正在加载视频...

视频加载失败

返回首页

Vik and Val Bercovici (WEKA) map where AI inference memory is headed. Every 100x cut in KV cache gets swallowed by ~10,000x more usage, so demand climbs. - NVLink beats the board: 128 lanes vs 32 PCIe - WEKA serves NAND-backed storage faster than DRAM over network - DeepSeek's...

semidoped's profile picture

Semi Doped

8,212 subscribers

60,918 次观看 • 2 个月前 •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

Memory is eating the world When consumers feel the heat, you know its a problem. While big-3 memory companies are enjoying sweet-sweet profit margins…the rest of the electronics market is cracking under pricing pressure. - GoPro is more like GoingUnder - Apple products 10-30% pricier - Even music synths are more expensive Is it price gouging at this point? Is this revenge for all the memory cyclicality? When will this end? Chapters: 0:00 Memory crisis hits 1:20 AI impacting consumers 3:00 AI causing inflation 6:48 Consumer demand drop? 8:49 AI demand inelastic 10:56 Long-term memory outlook 11:02 GoPro's memory woes 12:24 Apple's pricing power 18:20 Apple seeks CXMT DRAM 21:38 Shrinkflation for phones 23:16 Korea's memory investment 26:10 Micron's killing profits 33:00 Why AI needs so much DRAM 40:27 Future of AI training 44:34 Cost-optimizing inference Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar $MU
50:55

Memory is eating the world When consumers feel the heat, you know its a problem. While big-3 memory companies are enjoying sweet-sweet profit margins…the rest of the electronics market is cracking under pricing pressure. - GoPro is more like GoingUnder - Apple products 10-30% pricier - Even music synths are more expensive Is it price gouging at this point? Is this revenge for all the memory cyclicality? When will this end? Chapters: 0:00 Memory crisis hits 1:20 AI impacting consumers 3:00 AI causing inflation 6:48 Consumer demand drop? 8:49 AI demand inelastic 10:56 Long-term memory outlook 11:02 GoPro's memory woes 12:24 Apple's pricing power 18:20 Apple seeks CXMT DRAM 21:38 Shrinkflation for phones 23:16 Korea's memory investment 26:10 Micron's killing profits 33:00 Why AI needs so much DRAM 40:27 Future of AI training 44:34 Cost-optimizing inference Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar $MU

Semi Doped

34,810 次观看 • 2 个月前

New episode: Austin and Vik break down Qualcomm's Investor Day - HBC, Alphawave, Modular, more • 2/3 of revenue from data center, auto, and IoT by 2029, not handsets • HBC stacks LPDDR on the logic die, not beside it: 2,000 lanes become 100,000 • Disaggregation lets latecomers slot a decode rack beside Nvidia • Alphawave is like QCOM's Mellanox; Modular (Chris Lattner) is its CUDA • C1000 CPU wins Meta, but ships ~2 yrs into a CPU shortage Chapters: 0:00 Communications? That's just the start 4:08 Inside $QCOM investor day 9:16 Can Qualcomm build a data center business? 13:09 Disaggregated inference opens the door 17:57 High Bandwidth Compute: memory on the XPU 30:29 "No advanced packaging" just moves the problem 36:20 The roadmap, Alphawave, and Modular 46:00 The C1000 CPU and the agentic shortage 50:40 Cars as token generators, the $1T robotics bet 57:32 The memory market: MOAR Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar
59:43

New episode: Austin and Vik break down Qualcomm's Investor Day - HBC, Alphawave, Modular, more • 2/3 of revenue from data center, auto, and IoT by 2029, not handsets • HBC stacks LPDDR on the logic die, not beside it: 2,000 lanes become 100,000 • Disaggregation lets latecomers slot a decode rack beside Nvidia • Alphawave is like QCOM's Mellanox; Modular (Chris Lattner) is its CUDA • C1000 CPU wins Meta, but ships ~2 yrs into a CPU shortage Chapters: 0:00 Communications? That's just the start 4:08 Inside $QCOM investor day 9:16 Can Qualcomm build a data center business? 13:09 Disaggregated inference opens the door 17:57 High Bandwidth Compute: memory on the XPU 30:29 "No advanced packaging" just moves the problem 36:20 The roadmap, Alphawave, and Modular 46:00 The C1000 CPU and the agentic shortage 50:40 Cars as token generators, the $1T robotics bet 57:32 The memory market: MOAR Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar

Semi Doped

46,935 次观看 • 2 个月前

FEELING HOT HOT HOT! 🎙️ NEW EPISODE: OpenAI's Jalapeño! Vik and Austin react to OpenAI's Jalapeño inference chip, announced at Hot Chips by Richard Ho @cdleary. - "Dark silicon is cheaper than idle accelerators": one balanced chip beats a GPU + LPU pair? - Nine months from RTL to tapeout, a Blackwell-class chip designed with AI for EDA - NUMA-style local HBM slices per accelerator fix the "operands arrive late" problem - Broadcom ESUN scale-up: 128 chips per rack at 600 GB/s, 2,048 chips across 16 racks at 200G (still scale up!) - The "regret factor": opportunity cost of missing a future model beats the cost of generality Plus the extra spicy questions: should OpenAI sell it, how much of the RTL was AI-written, and where is Anthropic's chip? Chapters: 0:00 Hot Chips Reaction 2:32 Designing for User Experience 11:16 A Generalized Inference Chip 14:18 The Foundry-IDM Analogy 18:42 The 'Regret Factor' 21:02 The 9-Month Design Cycle 23:45 Challenging the Two-Chip Solution 35:08 Solving HBM Underutilization 36:46 The NUMA Architecture Solution 39:28 System-Level ESUN Networking 42:06 Dark Silicon vs. Idle Accelerators 49:08 A Wake-Up Call for the Industry 52:59 Where's Anthropic's Chip? Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar
54:32

FEELING HOT HOT HOT! 🎙️ NEW EPISODE: OpenAI's Jalapeño! Vik and Austin react to OpenAI's Jalapeño inference chip, announced at Hot Chips by Richard Ho @cdleary. - "Dark silicon is cheaper than idle accelerators": one balanced chip beats a GPU + LPU pair? - Nine months from RTL to tapeout, a Blackwell-class chip designed with AI for EDA - NUMA-style local HBM slices per accelerator fix the "operands arrive late" problem - Broadcom ESUN scale-up: 128 chips per rack at 600 GB/s, 2,048 chips across 16 racks at 200G (still scale up!) - The "regret factor": opportunity cost of missing a future model beats the cost of generality Plus the extra spicy questions: should OpenAI sell it, how much of the RTL was AI-written, and where is Anthropic's chip? Chapters: 0:00 Hot Chips Reaction 2:32 Designing for User Experience 11:16 A Generalized Inference Chip 14:18 The Foundry-IDM Analogy 18:42 The 'Regret Factor' 21:02 The 9-Month Design Cycle 23:45 Challenging the Two-Chip Solution 35:08 Solving HBM Underutilization 36:46 The NUMA Architecture Solution 39:28 System-Level ESUN Networking 42:06 Dark Silicon vs. Idle Accelerators 49:08 A Wake-Up Call for the Industry 52:59 Where's Anthropic's Chip? Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar

Semi Doped

21,136 次观看 • 19 天前

🎙️ NEW EPISODE: GlobalFoundries Thomas Barber: CPO, Silicon Photonics, 300mm, SiGe, OCI, NRZ "The enemy to me right now is copper," says Tom Barber. Austin sits down with Tom to unpack why GF thinks it can lead the shift to optical interconnects twice over: once with its long-running silicon photonics platform, and again with the specialty Silicon Germanium process the industry needs to drive it. Key Takeaways - CPO's real win isn't speed, it's power: saving 20-25 pJ/bit frees part of a data center's fixed 50-100 MW budget to go toward compute instead of moving bits. - GlobalFoundries leads photonics revenue for an unglamorous reason: it moved to 300mm wafers early, which yields 2.25x more die per wafer than the 200mm lines rivals still run. - CPO can end up more reliable than the pluggables it's replacing, not less, because it deletes the physical plug connector, and dust at that connector is the leading cause of field failures. - The OCI MSA picks NRZ over faster PAM4, deliberately going wide and slow, because NRZ's native bit error rate is a million times lower, which simplifies the receiver and cuts power. - GlobalFoundries stacks two specialty processes into one edge: micro-mirror couplers on the photonic side, and Silicon Germanium transistors hitting 350-400 GHz on the electrical side driving them. - The real competitor for silicon photonics isn't another foundry, it's copper itself — and that market is big enough that GlobalFoundries and TSMC can both win without taking share from each other. Chapters: 0:00 GlobalFoundries in Photonics 1:21 GF's Photonics Strategy 3:02 GF's Market Leadership 5:06 300mm Wafer Advantage 6:55 Copper's Range Limits 9:18 Pluggable to CPO 13:48 CPO Reliability 19:22 OCI MSA Explained 26:20 GF's Scale Platform 29:04 Micro Mirror Technology 31:32 Photonics vs. Copper 38:25 Silicon Germanium Advantage Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar
39:33

🎙️ NEW EPISODE: GlobalFoundries Thomas Barber: CPO, Silicon Photonics, 300mm, SiGe, OCI, NRZ "The enemy to me right now is copper," says Tom Barber. Austin sits down with Tom to unpack why GF thinks it can lead the shift to optical interconnects twice over: once with its long-running silicon photonics platform, and again with the specialty Silicon Germanium process the industry needs to drive it. Key Takeaways - CPO's real win isn't speed, it's power: saving 20-25 pJ/bit frees part of a data center's fixed 50-100 MW budget to go toward compute instead of moving bits. - GlobalFoundries leads photonics revenue for an unglamorous reason: it moved to 300mm wafers early, which yields 2.25x more die per wafer than the 200mm lines rivals still run. - CPO can end up more reliable than the pluggables it's replacing, not less, because it deletes the physical plug connector, and dust at that connector is the leading cause of field failures. - The OCI MSA picks NRZ over faster PAM4, deliberately going wide and slow, because NRZ's native bit error rate is a million times lower, which simplifies the receiver and cuts power. - GlobalFoundries stacks two specialty processes into one edge: micro-mirror couplers on the photonic side, and Silicon Germanium transistors hitting 350-400 GHz on the electrical side driving them. - The real competitor for silicon photonics isn't another foundry, it's copper itself — and that market is big enough that GlobalFoundries and TSMC can both win without taking share from each other. Chapters: 0:00 GlobalFoundries in Photonics 1:21 GF's Photonics Strategy 3:02 GF's Market Leadership 5:06 300mm Wafer Advantage 6:55 Copper's Range Limits 9:18 Pluggable to CPO 13:48 CPO Reliability 19:22 OCI MSA Explained 26:20 GF's Scale Platform 29:04 Micro Mirror Technology 31:32 Photonics vs. Copper 38:25 Silicon Germanium Advantage Get more of Austin and Vik daily, free! Sign up: Connect with Vik and Austin: Vik's Paid Substack: Austin's Paid Substack: Austin Lyons Vikram Sekar

Semi Doped

28,309 次观看 • 1 个月前

The Cost of Intelligence is Heading to Zero | Hyperspace P2P Distributed Cache We present to you our breakthrough cross-domain work across AI, distributed systems, cryptography, game theory to solve the primary structural inefficiency at the heart of AI infrastructure: most inference is redundant. Google has reported that only 15% of daily searches are truly novel. The rest are repeats or close variants. LLM inference inherits this same power-law distribution. Enterprise chatbots see 70-80% of queries fall into a handful of intent categories. System prompts are identical across 100% of requests within an application. The KV attention state for "You are a helpful assistant" has been computed billions of times, on millions of GPUs, identically. And yet every AI lab, every startup, every self-hosted deployment - computes and caches these results independently. There is no shared layer. No global memory. Every provider pays the full compute cost for every query, even when the answer already exists somewhere in the network. This is the problem Hyperspace solves where distributed cache operates at three levels, each catching a different class of redundancy: 1. Response cache Same prompt, same model, same parameters - instant cached response from any node in the network. SHA-256 hash lookup via DHT, with cryptographic cache proofs linking every response to its original inference execution. No trust required. Fetchers re-announce as providers, so popular responses replicate naturally across more nodes. 2. KV prefix cache Same system prompt tokens - skip the most expensive part of inference entirely. Prefill (computing Key-Value attention states) is deterministic: same model plus same tokens always produces identical KV state. The network caches these states using erasure coding and distributes them via the routing network. New questions that share a common prefix resume generation from cached state instead of recomputing from scratch. 3. Routing to cached nodes Instead of transferring KV state across the network for every request, Hyperspace routes the request to the node that already has the state loaded in VRAM. The request goes to the cache, not the cache to the request. Together, these three layers mean that 70-90% of inference requests at network scale never require full GPU computation. This work doesn't exist in isolation. It builds on research from across the industry: SGLang's RadixAttention demonstrated that automatic prefix sharing can yield up to 5x speedup on structured LLM workloads. Moonshot AI's Mooncake built an entire KV-cache-centric disaggregated architecture for production serving at Kimi. Anthropic, OpenAI, and Google all launched prompt caching products in 2024 - priced at 50-90% discounts - because system prompt reuse is so pervasive that it changes the economics of inference. What all of these systems share is a common limitation: they operate within a single organization's infrastructure. SGLang caches prefixes within one server. Mooncake disaggregates KV cache within one datacenter. Anthropic's prompt caching works within one API provider's fleet. None of them can share cached state across organizational boundaries. Hyperspace removes this boundary. The cache is global. A response computed by a node in Tokyo is immediately available to a node in Berlin. A KV prefix state generated for Qwen-32B on one machine is verifiable and reusable by any other machine running the same model. The routing network provides the delivery guarantees, the erasure coding provides the redundancy, and the cache proofs provide the trust. What this means for the cost of intelligence Big AI labs scale linearly: twice the users means twice the GPU spend. Every query is a cost center. Their internal caching helps, but it's siloed - Lab A's cache can't serve Lab B's users, and neither can serve a self-hosted Llama deployment. Hyperspace scales sub-linearly. Every new node that joins the network adds to the global cache. Every inference result enriches the cache for all future requests. The cache hit rate rises with network size because query distributions follow a power law - the most common questions are asked exponentially more often than rare ones. The implication is simple: as the network grows, the effective cost per inference drops. Not linearly. Logarithmically. At 10 million nodes, we estimate 75-90% of all inference requests can be served from cache, eliminating 400,000+ MWh of energy consumption per year and avoiding over 200,000 tons of CO2 emissions. The first person to ask a question pays the compute cost. Everyone after them gets the answer for free, with cryptographic proof that it's authentic. Training is competitive. Inference is shared Open-weight models are converging on quality with closed models. Labs will continue to differentiate on training - data curation, architecture innovation, RLHF tuning. That's where the real intellectual property lives. But inference is a commodity. Two copies of Qwen-32B running the same prompt produce the same KV state and the same response, byte for byte, regardless of whose GPU runs the matrix multiplication. There is no moat in multiplying matrices. The moat is in training the weights. A global distributed cache makes this separation explicit. It doesn't matter who trained the model. Once the weights are open, the inference cost approaches zero at scale - because the network remembers every answer and can prove it's correct. No lab, no matter how well-funded, can match this. They cannot share caches across competitors. They scale linearly. The network scales logarithmically. The marginal cost of intelligence approaches zero. That's the endgame.
0:23

The Cost of Intelligence is Heading to Zero | Hyperspace P2P Distributed Cache We present to you our breakthrough cross-domain work across AI, distributed systems, cryptography, game theory to solve the primary structural inefficiency at the heart of AI infrastructure: most inference is redundant. Google has reported that only 15% of daily searches are truly novel. The rest are repeats or close variants. LLM inference inherits this same power-law distribution. Enterprise chatbots see 70-80% of queries fall into a handful of intent categories. System prompts are identical across 100% of requests within an application. The KV attention state for "You are a helpful assistant" has been computed billions of times, on millions of GPUs, identically. And yet every AI lab, every startup, every self-hosted deployment - computes and caches these results independently. There is no shared layer. No global memory. Every provider pays the full compute cost for every query, even when the answer already exists somewhere in the network. This is the problem Hyperspace solves where distributed cache operates at three levels, each catching a different class of redundancy: 1. Response cache Same prompt, same model, same parameters - instant cached response from any node in the network. SHA-256 hash lookup via DHT, with cryptographic cache proofs linking every response to its original inference execution. No trust required. Fetchers re-announce as providers, so popular responses replicate naturally across more nodes. 2. KV prefix cache Same system prompt tokens - skip the most expensive part of inference entirely. Prefill (computing Key-Value attention states) is deterministic: same model plus same tokens always produces identical KV state. The network caches these states using erasure coding and distributes them via the routing network. New questions that share a common prefix resume generation from cached state instead of recomputing from scratch. 3. Routing to cached nodes Instead of transferring KV state across the network for every request, Hyperspace routes the request to the node that already has the state loaded in VRAM. The request goes to the cache, not the cache to the request. Together, these three layers mean that 70-90% of inference requests at network scale never require full GPU computation. This work doesn't exist in isolation. It builds on research from across the industry: SGLang's RadixAttention demonstrated that automatic prefix sharing can yield up to 5x speedup on structured LLM workloads. Moonshot AI's Mooncake built an entire KV-cache-centric disaggregated architecture for production serving at Kimi. Anthropic, OpenAI, and Google all launched prompt caching products in 2024 - priced at 50-90% discounts - because system prompt reuse is so pervasive that it changes the economics of inference. What all of these systems share is a common limitation: they operate within a single organization's infrastructure. SGLang caches prefixes within one server. Mooncake disaggregates KV cache within one datacenter. Anthropic's prompt caching works within one API provider's fleet. None of them can share cached state across organizational boundaries. Hyperspace removes this boundary. The cache is global. A response computed by a node in Tokyo is immediately available to a node in Berlin. A KV prefix state generated for Qwen-32B on one machine is verifiable and reusable by any other machine running the same model. The routing network provides the delivery guarantees, the erasure coding provides the redundancy, and the cache proofs provide the trust. What this means for the cost of intelligence Big AI labs scale linearly: twice the users means twice the GPU spend. Every query is a cost center. Their internal caching helps, but it's siloed - Lab A's cache can't serve Lab B's users, and neither can serve a self-hosted Llama deployment. Hyperspace scales sub-linearly. Every new node that joins the network adds to the global cache. Every inference result enriches the cache for all future requests. The cache hit rate rises with network size because query distributions follow a power law - the most common questions are asked exponentially more often than rare ones. The implication is simple: as the network grows, the effective cost per inference drops. Not linearly. Logarithmically. At 10 million nodes, we estimate 75-90% of all inference requests can be served from cache, eliminating 400,000+ MWh of energy consumption per year and avoiding over 200,000 tons of CO2 emissions. The first person to ask a question pays the compute cost. Everyone after them gets the answer for free, with cryptographic proof that it's authentic. Training is competitive. Inference is shared Open-weight models are converging on quality with closed models. Labs will continue to differentiate on training - data curation, architecture innovation, RLHF tuning. That's where the real intellectual property lives. But inference is a commodity. Two copies of Qwen-32B running the same prompt produce the same KV state and the same response, byte for byte, regardless of whose GPU runs the matrix multiplication. There is no moat in multiplying matrices. The moat is in training the weights. A global distributed cache makes this separation explicit. It doesn't matter who trained the model. Once the weights are open, the inference cost approaches zero at scale - because the network remembers every answer and can prove it's correct. No lab, no matter how well-funded, can match this. They cannot share caches across competitors. They scale linearly. The network scales logarithmically. The marginal cost of intelligence approaches zero. That's the endgame.

Varun

37,555 次观看 • 5 个月前

$NVDA $MU $SNDK $LITE PAPER OVERVIEW AND CORE CLAIMS The paper “KV Cache Transform Coding for Compact Storage in LLM Inference” introduces kvtc, a transform-coding pipeline that compresses transformer key-value (KV) caches primarily for storage and transfer in LLM serving, rather than for accelerating the per-token attention kernel during active decoding. The method combines 3 stages: (1) feature decorrelation via a PCA basis computed from a calibration dataset and reused across requests; (2) adaptive, variable-precision quantization with bit allocation solved via dynamic programming (DP), including groupwise scaling/shift overhead; and (3) lossless entropy coding (DEFLATE via nvCOMP in the reference implementation) to exploit residual redundancy after quantization. The central empirical claim is that KV tensors contain large, exploitable redundancy across heads and layers, enabling approximately 20× compression versus a 16-bit baseline with negligible degradation across a broad set of accuracy and long-context benchmarks, with materially higher compression (≥40×) available at modest quality cost in some regimes. The system claim is that such compression materially improves the economics of multi-turn, prefix-reuse serving by extending effective KV cache capacity in GPU HBM and host tiers (DRAM/NVMe) and by reducing inter-node and GPU↔host bandwidth demands, thereby improving cache hit rates and reducing time-to-first-token (TTFT) relative to recomputation when caches would otherwise be evicted. KV CACHE AS THE DOMINANT STATE VARIABLE IN INFERENCE ECONOMICS KV cache growth is linear in context length and is multiplicative in layers and attention heads, making it an increasingly dominant constraint as (a) context lengths expand, (b) models add layers and maintain large hidden dimensions, and (c) production workloads shift toward iterative and tool-augmented interactions that repeatedly reuse long prefixes. The paper uses the canonical 16-bit KV cache size formula (4·l·h·d_head·t) bytes and reports 16-bit KV cache sizes per 1K tokens of context that are already operationally large: 128MiB for Llama 3.1 8B, 160MiB for Mistral NeMo 12B, and 320MiB for Llama 3.3 70B Instruct. In binary units, these figures imply per-token KV footprints of 128KiB/token (Llama 3.1 8B), 160KiB/token (Mistral NeMo 12B), and 320KiB/token (Llama 3.3 70B Instruct) at 16-bit. For a 10K-token prompt (10×1K in the paper’s binary convention), the 16-bit KV cache sizes scale to approximately 1.25GiB (Llama 3.1 8B), 1.56GiB (Mistral NeMo 12B), and 3.13GiB (Llama 3.3 70B Instruct). These magnitudes explain why stale caches create a throughput–latency dilemma: retaining them in HBM maximizes responsiveness on future turns but crowds out concurrent sessions; evicting them forces quadratic-cost prefill recomputation and increases TTFT; offloading them to host or storage introduces large transfer overhead and consumes DRAM/NVMe capacity. A key operational nuance emphasized is that modern serving stacks increasingly treat KV caches as a database, leveraging block paging and shared-prefix reuse. In the common disaggregated serving design (separate prefill and decode nodes), KV cache transfer becomes a dominant category of cross-node traffic. Under that design, any reduction in KV cache size directly increases effective fabric capacity and reduces tail latency attributable to congestion, while also enabling longer cache lifetimes in “hot” (HBM) and “warm” (CPU DRAM) tiers that raise cache hit rates and reduce recomputation frequency. The paper’s quantitative example illustrates the economic stakes: a 1,000-line code file tokenized at ~10 tokens/line yields ~10K tokens; for Llama 3.3 70B, an 8-bit KV cache for that context is ~1.6GiB. Reuse across subsequent turns or parallel chats around the same file is valuable, but HBM scarcity makes retaining many such caches infeasible without compression. TECHNICAL MECHANISM: WHY KV CACHES ARE COMPRESSIBLE AND HOW KVTC EXPLOITS IT The technical rationale begins with an empirical observation: keys (and, to a lesser extent, values) across different attention heads can be aligned into a shared latent space using orthogonal transformations (Procrustes alignment). This supports the hypothesis that head-specific projections introduce rotations of a common subspace rather than completely distinct information, implying that concatenating across heads and layers should reveal low-rank structure suitable for linear decorrelation and dimensionality reduction. The method operationalizes this using a PCA/SVD basis learned from calibration data rather than recomputing a decomposition per prompt. This design choice targets production viability: per-prompt SVD is computationally expensive and scales poorly with long prompts and frequent cache updates. kvtc is explicitly structured as an offline-calibrated, online-applied codec: Calibration (performed 1 time per model and compression setting for DP allocation) A calibration dataset is forwarded through the model to collect KV caches. Token positions are pooled, and a subset of positions is sampled. Keys and values are processed separately. Several implementation choices are highlighted as decisive for stability: Rotary positional embeddings are effectively removed prior to compression (“undo positional rotations”), because positional rotations degrade the apparent low-rank structure of keys. “Attention sink” tokens (the earliest tokens in the sequence) and a sliding window of most recent tokens are excluded from compression because they disproportionately affect attention patterns and are empirically more sensitive to reconstruction error. Cross-layer concatenation is used: keys (or values) from multiple layers and heads at the same token position are concatenated along the feature axis to form a higher-dimensional feature vector. PCA is computed over these concatenated vectors, improving robustness relative to per-layer or per-head PCA. The PCA basis is computed via SVD of centered calibration data, using randomized SVD for scalability with a target rank cutoff. The paper reports calibration regimes of 160K tokens for several models with a 10K PCA dimension cutoff (8K for Qwen variants with fewer KV heads), selected to fit within a single 80GB H100 memory envelope and complete within minutes. A critical economic detail is that the same PCA basis can be reused across multiple compression ratios; only the DP-derived precision assignment changes per compression target. Compression (applied between inference phases) Compression operates on stored KV cache tensors, not on weights, and does not modify attention computation. The KV cache is projected into the PCA basis, quantized, packed, and then entropy-coded. Compression is positioned as a background or between-phase operation (after decoding, or between prefill and decode), executed on GPU or CPU depending on where the cache currently resides. The design intent is that compression should not sit on the critical per-token decoding path; it is a storage and transport optimization. Decompression (performed prior to reuse) Decompression reverses the entropy coding and quantization and applies the inverse PCA projection. A practical latency optimization is proposed: inverse projection can be performed layer-by-layer using submatrices of the PCA basis, allowing generation to begin before the full cache is reconstructed, reducing TTFT. Quantization and bit allocation are the core differentiators versus simpler PCA truncation. PCA provides ordered components by variance; kvtc uses DP to allocate a global bit budget across PCA coordinates (and across groups of coordinates) to minimize reconstruction error in the decorrelated domain. Groups of subsequent PCA coordinates share 16-bit shift and scale factors (a microscaling-inspired design), and the DP algorithm jointly selects group size and precision type under a bit budget, including the overhead of per-group metadata. DP commonly assigns 0 bits to many trailing PCA components, which both increases compression and provides a mechanism to trim the PCA basis to the subset of components that actually carry payload, reducing compute and storage overhead of the projection matrices in deployment. Lossless entropy coding then exploits the structure induced by quantization. DEFLATE is used in the reference implementation, and the paper emphasizes that the incremental gain from the lossless stage is content-dependent but meaningful, with an average uplift of ~1.23× on top of quantization in the reported regime. An ablation in the appendices indicates that GPU-friendly variants (GDeflate) can achieve nearly identical compression ratios (≤0.1 difference in measured cases), implying that throughput-optimized lossless codecs can likely be substituted without sacrificing meaningful compression. EMPIRICAL RESULTS: ACCURACY, COMPRESSION, AND LATENCY General-purpose 8B–12B dense models The paper evaluates Llama 3.1 8B, MN-Minitron 8B, and Mistral NeMo 12B across math/knowledge (GSM8K, MMLU) and long-context tasks (Qasper, Lost in the Middle, RULER Variable Tracking) under a simulated multi-turn regime where compression/decompression is applied periodically, with a sliding window of recent tokens excluded. A consistent pattern appears: kvtc maintains near-vanilla performance through 16× compression settings, and remains competitive at 32×, with degradation becoming task- and model-dependent at 64×, particularly on long-context retrieval metrics when compression is pushed aggressively. Selected quantitative anchor points from the paper’s standard-error table (all values are reported with the paper’s evaluation setup and token-window exclusions): Llama 3.1 8B Vanilla: GSM8K 56.8, MMLU 60.5, Qasper 40.4, LITM 99.4, RULER-VT 99.8 kvtc16×: GSM8K 56.9, MMLU 60.1, Qasper 40.7, LITM 99.3, RULER-VT 99.1 kvtc32×: GSM8K 57.8, MMLU 60.6, Qasper 39.4, LITM 99.1, RULER-VT 98.9 kvtc64×: GSM8K 57.2, MMLU 60.7, Qasper 37.8, LITM 90.2, RULER-VT 95.9 These results indicate that, for this model, long-context sensitivity emerges at 64× with meaningful drops in LITM and RULER-VT, while math/knowledge scores remain stable, implying a differential sensitivity consistent with key-vector precision being more critical for retrieval-style behavior. Mistral NeMo 12B Vanilla: GSM8K 61.9, MMLU 64.5, Qasper 38.4, LITM 99.5, RULER-VT 99.8 kvtc16×: GSM8K 62.0, MMLU 64.4, Qasper 37.6, LITM 99.8, RULER-VT 99.5 kvtc32×: GSM8K 62.2, MMLU 63.8, Qasper 37.5, LITM 99.6, RULER-VT 98.7 kvtc64×: GSM8K 61.9, MMLU 61.4, Qasper 38.0, LITM 95.3, RULER-VT 98.0 Here, degradation at 64× is visible but materially smaller than the Llama 3.1 8B LITM drop, suggesting model-architecture or training-data differences can change the tolerance envelope for aggressive KV cache distortion. MN-Minitron 8B Vanilla: GSM8K 59.1, MMLU 64.3, Qasper 38.2, LITM 99.8, RULER-VT 99.4 kvtc16×: GSM8K 60.3, MMLU 64.1, Qasper 38.6, LITM 99.3, RULER-VT 98.8 kvtc32×: GSM8K 59.1, MMLU 63.7, Qasper 37.7, LITM 86.9, RULER-VT 96.0 kvtc64×: GSM8K 57.8, MMLU 62.1, Qasper 38.1, LITM 59.5, RULER-VT 93.4 This model shows markedly higher sensitivity on LITM at 32× and 64×, despite stable short-context metrics, reinforcing that “compression safety” is not monotonic in parameter count and that pruning/distillation choices can alter KV cache redundancy or robustness. Comparisons to baselines The paper compares kvtc to quantization baselines (KIVI, GEAR, FP8) and eviction baselines (H2O, TOVA), plus an SVD-based prefill-optimization method (xKV). Across the reported tasks: Low-bit quantization methods at modest compression (2-bit KV schemes) show earlier degradation in long-context behavior than kvtc at substantially higher compression settings. Eviction methods perform poorly as generic compressors for long-context tasks, consistent with their objective function (selective pruning) being misaligned with “lossless-ish storage for reuse.” xKV shows competitive results on some tasks but a consistent underperformance on Qasper relative to kvtc and vanilla in the provided tables, consistent with method-specific distortions introduced by its decomposition regime. Reasoning models and high-variance tasks For DeepSeek-R1-distilled Qwen 2.5 reasoning models, the paper evaluates AIME 2024/2025 and LiveCodeBench coding. Results are averaged over 8 runs with large variance, but a key inference is that kvtc at ~9×–21× compression achieves broadly similar AIME scores within variance bands, while coding performance remains stable at ~9× and degrades more visibly at ~18×–21× on the 7B model. An important nuance is that smaller reasoning models already have smaller KV footprints (reported ~29KiB/token for Qwen R1 1.5B versus 131KiB/token for Llama 3.1 8B), so the economic value of aggressive KV cache compression is proportionally higher for large models and long contexts than for small models with short contexts, unless the serving system’s bottleneck is dominated by cache transfer rather than HBM capacity. Multi-GPU inference and pipeline parallel For Llama 3.3 70B Instruct run pipeline-parallel across 4 GPUs (20 layers per GPU), the paper compresses KV cache chunks independently per GPU. On MATH-500, the reported accuracy declines from 75.6 (vanilla) to 74.4 at 10× and 72.6 at 20×, with standard errors near ~1.9. NIAH and LITM remain at 100.0 for all tested ratios in that table. The paper notes that joint compression across chunks could improve accuracy for some offload scenarios but is not required for feasibility, highlighting an engineering trade-off between deployment simplicity in distributed settings and optimal global compression. Latency and TTFT economics A critical system result is the measured compression/decompression latency on an H100 for a non-fused implementation. For Mistral NeMo 12B in bfloat16: BS=8, CTX=8K: compression 379ms, decompression 267ms; vanilla recompute TTFT 3098ms; kvtc decompression TTFT 380ms BS=2, CTX=16K: compression 194ms, decompression 143ms; vanilla recompute TTFT 1780ms; kvtc decompression TTFT 208ms These measurements imply that, when a cache would otherwise be recomputed, decompressing a stored compressed cache can reduce TTFT by ~8×–9× in these scenarios, even without kernel fusion. The decomposition of runtime shows PCA projection and entropy coding as the largest contributors, implying that GPU-optimized kernels and faster GPU-native lossless codecs could reduce overhead further. The fundamental economic conclusion is that, in multi-turn settings with long prefixes, compression-induced overhead is likely dominated by the avoided prefill compute and avoided transfer overhead for uncompressed caches. KEY DEPLOYMENT-SENSITIVE DESIGN CHOICES AND FAILURE MODES Several design choices appear to be “hard requirements” rather than optional optimizations: Sink tokens and sliding window exclusions The paper’s ablations show that compressing early “sink” tokens can catastrophically degrade accuracy at high compression ratios (example: Llama 3.1 8B at 64× collapses on multiple tasks when sink tokens are compressed). Similarly, compressing the most recent tokens hurts performance, motivating a sliding window (default 128 tokens) that remains uncompressed. This introduces a predictable engineering constraint: kvtc is not a uniform compression of the full cache; it is a policy-driven, token-position-dependent codec. Production integration therefore requires correct handling of token positions, attention sinks, and window management, and these policies must be aligned with attention-kernel behavior and model-specific sink dynamics. RoPE handling Removing positional rotations prior to compression is described as important for preserving low-rank structure. In deployment, this implies that the codec must be position-aware and must invert and reapply RoPE correctly. This is an additional source of complexity relative to pure per-token quantization and is sensitive to model variants and RoPE parameterizations. Calibration set representativeness The method’s quality hinges on the PCA basis generalizing from calibration data to production data. The paper demonstrates relative stability with 160K–200K calibration tokens and explores domain shifts (general web text vs math traces vs code). Results suggest that moderate domain mismatch is tolerated at 16×–64×, while extreme compression (e.g., 256× in ablations) becomes materially more sensitive to calibration choice. In production, this implies that operators targeting the “negligible degradation” regime should be able to calibrate with broadly representative corpora, while operators targeting ultra-high compression for specialized workloads should expect tighter coupling between calibration domain and achieved quality. PCA matrix storage overhead and operational footprint A non-trivial hidden cost is the need to store PCA projection matrices per model. The paper reports that, prior to DP trimming, PCA matrices stored at 16-bit can amount to a meaningful fraction of model parameter count (examples reported: ~2.4% for Llama 3.3 70B, ~8.7% for Llama 3.1 8B). This overhead is amortized across all cached sessions for a model but competes with HBM/DRAM budgets in multi-model serving. DP-driven trimming can reduce this overhead at higher compression ratios by removing zero-bit components, but the directionality is not guaranteed at low compression ratios if many components remain active. In distributed inference (pipeline parallel), per-chunk PCA can reduce matrix sizes, but may reduce cross-layer decorrelation benefits if fewer layers are concatenated. SYSTEM-LEVEL IMPLICATIONS FOR GENERATIVE AI INFRASTRUCTURE GPU AND HBM The principal infrastructure implication is that KV cache compression at storage time targets the dominant memory allocator stressor in stateful serving: the accumulation of idle or warm conversation state. For workloads with long reusable prefixes (code assistants, enterprise agents with large system prompts, repeated RAG scaffolds, document chat), the limiting resource frequently becomes HBM reserved for KV caches rather than compute. By compressing stale caches by ~20× (or more), the same HBM budget can retain a materially larger working set of cached prefixes, increasing cache hit rates and reducing recomputation. This effect is multiplicative with cache-aware routing and prefix sharing: more prefixes can remain resident (hot or warm) and can be routed to nodes that already hold them, improving both throughput and tail latency. However, kvtc as described does not reduce the active KV cache footprint during the actual attention computation for a currently decoding sequence, because the model operates on decompressed KV caches during decoding. Therefore, the method does not directly reduce HBM bandwidth consumed by attention kernels during steady-state decode, and does not directly address the “memory traffic per generated token” bottleneck that motivates online KV quantization and eviction strategies. The primary HBM benefit is increased effective capacity for caches between turns and reduced HBM pressure from storing many idle sessions, not reduced per-token decode bandwidth. Compression and decompression themselves consume GPU compute and memory bandwidth. The measured decompression TTFT of ~208ms–380ms in the provided benchmarks indicates that the overhead is real but can be materially smaller than recomputation of long prefixes. In an HBM-constrained serving environment, this overhead can be interpreted as a trade between (a) maintaining more caches warm and paying decompression on reuse versus (b) evicting caches and paying full prefill recomputation. The decision boundary will depend on distribution of inter-turn idle times, probability of reuse, and SLA sensitivity to TTFT. kvtc expands the feasible region where keeping caches is economically rational, especially for long prompts. CPU AND DRAM The method implies a stronger role for CPU DRAM as a warm KV cache tier. A ~20× compression ratio changes the practical scale of “warm state” that can be stored per server. Using the paper’s reported KV cache sizes, a 10K-token 16-bit KV cache for Llama 3.3 70B is ~3.13GiB; compressing by ~20× would reduce this to ~160MiB. At that size, storing hundreds to thousands of warm conversation states in DRAM becomes materially more feasible, increasing cache hit rates and reducing NVMe dependence. This can shift system design from “HBM-only hot caches with aggressive eviction” toward “HBM hot + DRAM warm with long retention,” which is structurally analogous to CPU page cache hierarchies in classical systems design. CPU compute implications depend on where compression is executed. The paper explicitly allows compression on CPU if the cache is already in storage, but the strongest bandwidth savings are achieved when compression happens before moving KV caches off the GPU. If an operator chooses GPU-side compression prior to PCIe/NVLink transfer, CPU compute overhead is modest (orchestrating and DP calibration offline). If an operator instead transfers uncompressed caches to CPU for compression, bandwidth savings are forfeited and CPU memory bandwidth becomes a bottleneck. Therefore, the most economically coherent deployment path is GPU-native compression/decompression with CPU DRAM used as the warm storage reservoir.
12:09

$NVDA $MU $SNDK $LITE PAPER OVERVIEW AND CORE CLAIMS The paper “KV Cache Transform Coding for Compact Storage in LLM Inference” introduces kvtc, a transform-coding pipeline that compresses transformer key-value (KV) caches primarily for storage and transfer in LLM serving, rather than for accelerating the per-token attention kernel during active decoding. The method combines 3 stages: (1) feature decorrelation via a PCA basis computed from a calibration dataset and reused across requests; (2) adaptive, variable-precision quantization with bit allocation solved via dynamic programming (DP), including groupwise scaling/shift overhead; and (3) lossless entropy coding (DEFLATE via nvCOMP in the reference implementation) to exploit residual redundancy after quantization. The central empirical claim is that KV tensors contain large, exploitable redundancy across heads and layers, enabling approximately 20× compression versus a 16-bit baseline with negligible degradation across a broad set of accuracy and long-context benchmarks, with materially higher compression (≥40×) available at modest quality cost in some regimes. The system claim is that such compression materially improves the economics of multi-turn, prefix-reuse serving by extending effective KV cache capacity in GPU HBM and host tiers (DRAM/NVMe) and by reducing inter-node and GPU↔host bandwidth demands, thereby improving cache hit rates and reducing time-to-first-token (TTFT) relative to recomputation when caches would otherwise be evicted. KV CACHE AS THE DOMINANT STATE VARIABLE IN INFERENCE ECONOMICS KV cache growth is linear in context length and is multiplicative in layers and attention heads, making it an increasingly dominant constraint as (a) context lengths expand, (b) models add layers and maintain large hidden dimensions, and (c) production workloads shift toward iterative and tool-augmented interactions that repeatedly reuse long prefixes. The paper uses the canonical 16-bit KV cache size formula (4·l·h·d_head·t) bytes and reports 16-bit KV cache sizes per 1K tokens of context that are already operationally large: 128MiB for Llama 3.1 8B, 160MiB for Mistral NeMo 12B, and 320MiB for Llama 3.3 70B Instruct. In binary units, these figures imply per-token KV footprints of 128KiB/token (Llama 3.1 8B), 160KiB/token (Mistral NeMo 12B), and 320KiB/token (Llama 3.3 70B Instruct) at 16-bit. For a 10K-token prompt (10×1K in the paper’s binary convention), the 16-bit KV cache sizes scale to approximately 1.25GiB (Llama 3.1 8B), 1.56GiB (Mistral NeMo 12B), and 3.13GiB (Llama 3.3 70B Instruct). These magnitudes explain why stale caches create a throughput–latency dilemma: retaining them in HBM maximizes responsiveness on future turns but crowds out concurrent sessions; evicting them forces quadratic-cost prefill recomputation and increases TTFT; offloading them to host or storage introduces large transfer overhead and consumes DRAM/NVMe capacity. A key operational nuance emphasized is that modern serving stacks increasingly treat KV caches as a database, leveraging block paging and shared-prefix reuse. In the common disaggregated serving design (separate prefill and decode nodes), KV cache transfer becomes a dominant category of cross-node traffic. Under that design, any reduction in KV cache size directly increases effective fabric capacity and reduces tail latency attributable to congestion, while also enabling longer cache lifetimes in “hot” (HBM) and “warm” (CPU DRAM) tiers that raise cache hit rates and reduce recomputation frequency. The paper’s quantitative example illustrates the economic stakes: a 1,000-line code file tokenized at ~10 tokens/line yields ~10K tokens; for Llama 3.3 70B, an 8-bit KV cache for that context is ~1.6GiB. Reuse across subsequent turns or parallel chats around the same file is valuable, but HBM scarcity makes retaining many such caches infeasible without compression. TECHNICAL MECHANISM: WHY KV CACHES ARE COMPRESSIBLE AND HOW KVTC EXPLOITS IT The technical rationale begins with an empirical observation: keys (and, to a lesser extent, values) across different attention heads can be aligned into a shared latent space using orthogonal transformations (Procrustes alignment). This supports the hypothesis that head-specific projections introduce rotations of a common subspace rather than completely distinct information, implying that concatenating across heads and layers should reveal low-rank structure suitable for linear decorrelation and dimensionality reduction. The method operationalizes this using a PCA/SVD basis learned from calibration data rather than recomputing a decomposition per prompt. This design choice targets production viability: per-prompt SVD is computationally expensive and scales poorly with long prompts and frequent cache updates. kvtc is explicitly structured as an offline-calibrated, online-applied codec: Calibration (performed 1 time per model and compression setting for DP allocation) A calibration dataset is forwarded through the model to collect KV caches. Token positions are pooled, and a subset of positions is sampled. Keys and values are processed separately. Several implementation choices are highlighted as decisive for stability: Rotary positional embeddings are effectively removed prior to compression (“undo positional rotations”), because positional rotations degrade the apparent low-rank structure of keys. “Attention sink” tokens (the earliest tokens in the sequence) and a sliding window of most recent tokens are excluded from compression because they disproportionately affect attention patterns and are empirically more sensitive to reconstruction error. Cross-layer concatenation is used: keys (or values) from multiple layers and heads at the same token position are concatenated along the feature axis to form a higher-dimensional feature vector. PCA is computed over these concatenated vectors, improving robustness relative to per-layer or per-head PCA. The PCA basis is computed via SVD of centered calibration data, using randomized SVD for scalability with a target rank cutoff. The paper reports calibration regimes of 160K tokens for several models with a 10K PCA dimension cutoff (8K for Qwen variants with fewer KV heads), selected to fit within a single 80GB H100 memory envelope and complete within minutes. A critical economic detail is that the same PCA basis can be reused across multiple compression ratios; only the DP-derived precision assignment changes per compression target. Compression (applied between inference phases) Compression operates on stored KV cache tensors, not on weights, and does not modify attention computation. The KV cache is projected into the PCA basis, quantized, packed, and then entropy-coded. Compression is positioned as a background or between-phase operation (after decoding, or between prefill and decode), executed on GPU or CPU depending on where the cache currently resides. The design intent is that compression should not sit on the critical per-token decoding path; it is a storage and transport optimization. Decompression (performed prior to reuse) Decompression reverses the entropy coding and quantization and applies the inverse PCA projection. A practical latency optimization is proposed: inverse projection can be performed layer-by-layer using submatrices of the PCA basis, allowing generation to begin before the full cache is reconstructed, reducing TTFT. Quantization and bit allocation are the core differentiators versus simpler PCA truncation. PCA provides ordered components by variance; kvtc uses DP to allocate a global bit budget across PCA coordinates (and across groups of coordinates) to minimize reconstruction error in the decorrelated domain. Groups of subsequent PCA coordinates share 16-bit shift and scale factors (a microscaling-inspired design), and the DP algorithm jointly selects group size and precision type under a bit budget, including the overhead of per-group metadata. DP commonly assigns 0 bits to many trailing PCA components, which both increases compression and provides a mechanism to trim the PCA basis to the subset of components that actually carry payload, reducing compute and storage overhead of the projection matrices in deployment. Lossless entropy coding then exploits the structure induced by quantization. DEFLATE is used in the reference implementation, and the paper emphasizes that the incremental gain from the lossless stage is content-dependent but meaningful, with an average uplift of ~1.23× on top of quantization in the reported regime. An ablation in the appendices indicates that GPU-friendly variants (GDeflate) can achieve nearly identical compression ratios (≤0.1 difference in measured cases), implying that throughput-optimized lossless codecs can likely be substituted without sacrificing meaningful compression. EMPIRICAL RESULTS: ACCURACY, COMPRESSION, AND LATENCY General-purpose 8B–12B dense models The paper evaluates Llama 3.1 8B, MN-Minitron 8B, and Mistral NeMo 12B across math/knowledge (GSM8K, MMLU) and long-context tasks (Qasper, Lost in the Middle, RULER Variable Tracking) under a simulated multi-turn regime where compression/decompression is applied periodically, with a sliding window of recent tokens excluded. A consistent pattern appears: kvtc maintains near-vanilla performance through 16× compression settings, and remains competitive at 32×, with degradation becoming task- and model-dependent at 64×, particularly on long-context retrieval metrics when compression is pushed aggressively. Selected quantitative anchor points from the paper’s standard-error table (all values are reported with the paper’s evaluation setup and token-window exclusions): Llama 3.1 8B Vanilla: GSM8K 56.8, MMLU 60.5, Qasper 40.4, LITM 99.4, RULER-VT 99.8 kvtc16×: GSM8K 56.9, MMLU 60.1, Qasper 40.7, LITM 99.3, RULER-VT 99.1 kvtc32×: GSM8K 57.8, MMLU 60.6, Qasper 39.4, LITM 99.1, RULER-VT 98.9 kvtc64×: GSM8K 57.2, MMLU 60.7, Qasper 37.8, LITM 90.2, RULER-VT 95.9 These results indicate that, for this model, long-context sensitivity emerges at 64× with meaningful drops in LITM and RULER-VT, while math/knowledge scores remain stable, implying a differential sensitivity consistent with key-vector precision being more critical for retrieval-style behavior. Mistral NeMo 12B Vanilla: GSM8K 61.9, MMLU 64.5, Qasper 38.4, LITM 99.5, RULER-VT 99.8 kvtc16×: GSM8K 62.0, MMLU 64.4, Qasper 37.6, LITM 99.8, RULER-VT 99.5 kvtc32×: GSM8K 62.2, MMLU 63.8, Qasper 37.5, LITM 99.6, RULER-VT 98.7 kvtc64×: GSM8K 61.9, MMLU 61.4, Qasper 38.0, LITM 95.3, RULER-VT 98.0 Here, degradation at 64× is visible but materially smaller than the Llama 3.1 8B LITM drop, suggesting model-architecture or training-data differences can change the tolerance envelope for aggressive KV cache distortion. MN-Minitron 8B Vanilla: GSM8K 59.1, MMLU 64.3, Qasper 38.2, LITM 99.8, RULER-VT 99.4 kvtc16×: GSM8K 60.3, MMLU 64.1, Qasper 38.6, LITM 99.3, RULER-VT 98.8 kvtc32×: GSM8K 59.1, MMLU 63.7, Qasper 37.7, LITM 86.9, RULER-VT 96.0 kvtc64×: GSM8K 57.8, MMLU 62.1, Qasper 38.1, LITM 59.5, RULER-VT 93.4 This model shows markedly higher sensitivity on LITM at 32× and 64×, despite stable short-context metrics, reinforcing that “compression safety” is not monotonic in parameter count and that pruning/distillation choices can alter KV cache redundancy or robustness. Comparisons to baselines The paper compares kvtc to quantization baselines (KIVI, GEAR, FP8) and eviction baselines (H2O, TOVA), plus an SVD-based prefill-optimization method (xKV). Across the reported tasks: Low-bit quantization methods at modest compression (2-bit KV schemes) show earlier degradation in long-context behavior than kvtc at substantially higher compression settings. Eviction methods perform poorly as generic compressors for long-context tasks, consistent with their objective function (selective pruning) being misaligned with “lossless-ish storage for reuse.” xKV shows competitive results on some tasks but a consistent underperformance on Qasper relative to kvtc and vanilla in the provided tables, consistent with method-specific distortions introduced by its decomposition regime. Reasoning models and high-variance tasks For DeepSeek-R1-distilled Qwen 2.5 reasoning models, the paper evaluates AIME 2024/2025 and LiveCodeBench coding. Results are averaged over 8 runs with large variance, but a key inference is that kvtc at ~9×–21× compression achieves broadly similar AIME scores within variance bands, while coding performance remains stable at ~9× and degrades more visibly at ~18×–21× on the 7B model. An important nuance is that smaller reasoning models already have smaller KV footprints (reported ~29KiB/token for Qwen R1 1.5B versus 131KiB/token for Llama 3.1 8B), so the economic value of aggressive KV cache compression is proportionally higher for large models and long contexts than for small models with short contexts, unless the serving system’s bottleneck is dominated by cache transfer rather than HBM capacity. Multi-GPU inference and pipeline parallel For Llama 3.3 70B Instruct run pipeline-parallel across 4 GPUs (20 layers per GPU), the paper compresses KV cache chunks independently per GPU. On MATH-500, the reported accuracy declines from 75.6 (vanilla) to 74.4 at 10× and 72.6 at 20×, with standard errors near ~1.9. NIAH and LITM remain at 100.0 for all tested ratios in that table. The paper notes that joint compression across chunks could improve accuracy for some offload scenarios but is not required for feasibility, highlighting an engineering trade-off between deployment simplicity in distributed settings and optimal global compression. Latency and TTFT economics A critical system result is the measured compression/decompression latency on an H100 for a non-fused implementation. For Mistral NeMo 12B in bfloat16: BS=8, CTX=8K: compression 379ms, decompression 267ms; vanilla recompute TTFT 3098ms; kvtc decompression TTFT 380ms BS=2, CTX=16K: compression 194ms, decompression 143ms; vanilla recompute TTFT 1780ms; kvtc decompression TTFT 208ms These measurements imply that, when a cache would otherwise be recomputed, decompressing a stored compressed cache can reduce TTFT by ~8×–9× in these scenarios, even without kernel fusion. The decomposition of runtime shows PCA projection and entropy coding as the largest contributors, implying that GPU-optimized kernels and faster GPU-native lossless codecs could reduce overhead further. The fundamental economic conclusion is that, in multi-turn settings with long prefixes, compression-induced overhead is likely dominated by the avoided prefill compute and avoided transfer overhead for uncompressed caches. KEY DEPLOYMENT-SENSITIVE DESIGN CHOICES AND FAILURE MODES Several design choices appear to be “hard requirements” rather than optional optimizations: Sink tokens and sliding window exclusions The paper’s ablations show that compressing early “sink” tokens can catastrophically degrade accuracy at high compression ratios (example: Llama 3.1 8B at 64× collapses on multiple tasks when sink tokens are compressed). Similarly, compressing the most recent tokens hurts performance, motivating a sliding window (default 128 tokens) that remains uncompressed. This introduces a predictable engineering constraint: kvtc is not a uniform compression of the full cache; it is a policy-driven, token-position-dependent codec. Production integration therefore requires correct handling of token positions, attention sinks, and window management, and these policies must be aligned with attention-kernel behavior and model-specific sink dynamics. RoPE handling Removing positional rotations prior to compression is described as important for preserving low-rank structure. In deployment, this implies that the codec must be position-aware and must invert and reapply RoPE correctly. This is an additional source of complexity relative to pure per-token quantization and is sensitive to model variants and RoPE parameterizations. Calibration set representativeness The method’s quality hinges on the PCA basis generalizing from calibration data to production data. The paper demonstrates relative stability with 160K–200K calibration tokens and explores domain shifts (general web text vs math traces vs code). Results suggest that moderate domain mismatch is tolerated at 16×–64×, while extreme compression (e.g., 256× in ablations) becomes materially more sensitive to calibration choice. In production, this implies that operators targeting the “negligible degradation” regime should be able to calibrate with broadly representative corpora, while operators targeting ultra-high compression for specialized workloads should expect tighter coupling between calibration domain and achieved quality. PCA matrix storage overhead and operational footprint A non-trivial hidden cost is the need to store PCA projection matrices per model. The paper reports that, prior to DP trimming, PCA matrices stored at 16-bit can amount to a meaningful fraction of model parameter count (examples reported: ~2.4% for Llama 3.3 70B, ~8.7% for Llama 3.1 8B). This overhead is amortized across all cached sessions for a model but competes with HBM/DRAM budgets in multi-model serving. DP-driven trimming can reduce this overhead at higher compression ratios by removing zero-bit components, but the directionality is not guaranteed at low compression ratios if many components remain active. In distributed inference (pipeline parallel), per-chunk PCA can reduce matrix sizes, but may reduce cross-layer decorrelation benefits if fewer layers are concatenated. SYSTEM-LEVEL IMPLICATIONS FOR GENERATIVE AI INFRASTRUCTURE GPU AND HBM The principal infrastructure implication is that KV cache compression at storage time targets the dominant memory allocator stressor in stateful serving: the accumulation of idle or warm conversation state. For workloads with long reusable prefixes (code assistants, enterprise agents with large system prompts, repeated RAG scaffolds, document chat), the limiting resource frequently becomes HBM reserved for KV caches rather than compute. By compressing stale caches by ~20× (or more), the same HBM budget can retain a materially larger working set of cached prefixes, increasing cache hit rates and reducing recomputation. This effect is multiplicative with cache-aware routing and prefix sharing: more prefixes can remain resident (hot or warm) and can be routed to nodes that already hold them, improving both throughput and tail latency. However, kvtc as described does not reduce the active KV cache footprint during the actual attention computation for a currently decoding sequence, because the model operates on decompressed KV caches during decoding. Therefore, the method does not directly reduce HBM bandwidth consumed by attention kernels during steady-state decode, and does not directly address the “memory traffic per generated token” bottleneck that motivates online KV quantization and eviction strategies. The primary HBM benefit is increased effective capacity for caches between turns and reduced HBM pressure from storing many idle sessions, not reduced per-token decode bandwidth. Compression and decompression themselves consume GPU compute and memory bandwidth. The measured decompression TTFT of ~208ms–380ms in the provided benchmarks indicates that the overhead is real but can be materially smaller than recomputation of long prefixes. In an HBM-constrained serving environment, this overhead can be interpreted as a trade between (a) maintaining more caches warm and paying decompression on reuse versus (b) evicting caches and paying full prefill recomputation. The decision boundary will depend on distribution of inter-turn idle times, probability of reuse, and SLA sensitivity to TTFT. kvtc expands the feasible region where keeping caches is economically rational, especially for long prompts. CPU AND DRAM The method implies a stronger role for CPU DRAM as a warm KV cache tier. A ~20× compression ratio changes the practical scale of “warm state” that can be stored per server. Using the paper’s reported KV cache sizes, a 10K-token 16-bit KV cache for Llama 3.3 70B is ~3.13GiB; compressing by ~20× would reduce this to ~160MiB. At that size, storing hundreds to thousands of warm conversation states in DRAM becomes materially more feasible, increasing cache hit rates and reducing NVMe dependence. This can shift system design from “HBM-only hot caches with aggressive eviction” toward “HBM hot + DRAM warm with long retention,” which is structurally analogous to CPU page cache hierarchies in classical systems design. CPU compute implications depend on where compression is executed. The paper explicitly allows compression on CPU if the cache is already in storage, but the strongest bandwidth savings are achieved when compression happens before moving KV caches off the GPU. If an operator chooses GPU-side compression prior to PCIe/NVLink transfer, CPU compute overhead is modest (orchestrating and DP calibration offline). If an operator instead transfers uncompressed caches to CPU for compression, bandwidth savings are forfeited and CPU memory bandwidth becomes a bottleneck. Therefore, the most economically coherent deployment path is GPU-native compression/decompression with CPU DRAM used as the warm storage reservoir.

TheValueist

16,549 次观看 • 7 个月前

$MU $SNDK $LITE $VRT NVIDIA and Groq: 2nd and 3rd Order Strategic Infrastructure Effects and Market Implications Public reporting indicates NVIDIA has agreed to acquire Groq for approximately $20,000,000,000 in cash, while excluding Groq’s nascent cloud business from the transaction perimeter. The reported carve-out materially constrains the immediate, direct linkage from the acquisition to incremental, NVIDIA-controlled data center capacity build-out because GroqCloud appears to be the principal channel through which Groq hardware is currently monetized at scale as a service. The infrastructure-market implications therefore depend primarily on post-close product strategy: whether NVIDIA (1) commercializes Groq silicon as a distinct inference product line and drives broad deployment through OEM/ODM channels and partners, (2) uses the acquisition mainly to absorb IP and talent while de-emphasizing standalone Groq hardware volumes, or (3) uses Groq technology to reshape NVIDIA’s own inference systems and networking roadmaps. The dominant transmission mechanism into memory, networking, and facility infrastructure markets is the degree to which NVIDIA shifts incremental inference deployments away from GPU architectures that are tightly coupled to external high-bandwidth memory (HBM) and toward Groq’s current architecture, which emphasizes large on-chip SRAM, deterministic compiler-scheduled execution, and direct chip-to-chip connectivity. Independent and company-published materials describe Groq’s current-generation approach as having no external memory, keeping weights and KV cache on-chip during processing, and requiring model sharding across multiple chips due to limited on-chip SRAM per device. That architectural choice is directionally HBM-negative on a per-accelerator basis and ambiguous for DRAM, NAND, networking, power, and cooling on a per-token basis because the design can reduce memory wall losses and tail-latency overhead while potentially increasing the number of chips and interconnect endpoints required to serve large models and long-context workloads. HBM implications are the most mechanically straightforward but should be framed as second-derivative rather than absolute. If Groq-class inference silicon meaningfully displaces NVIDIA GPU-based inference deployments, incremental HBM bit demand tied to inference growth could be reduced relative to a GPU-only baseline because Groq’s current approach does not appear to attach HBM stacks to each accelerator. However, current market structure suggests HBM remains supply-constrained and is being pulled by multiple vectors including continued GPU training scale and high-capacity inference configurations, with leading suppliers signaling tight conditions extending beyond 2026. In that environment, reduced inference-driven HBM intensity could primarily reallocate scarce HBM supply toward higher-end training and premium inference GPUs rather than creating an outright volume collapse, preserving high utilization of HBM capacity while potentially affecting the slope of pricing power and capacity expansion urgency over a multi-year horizon. The key downside scenario for the HBM complex would be a durable architectural bifurcation where “good-enough” inference shifts disproportionately to HBM-less ASICs across a broad swath of deployments (latency-sensitive, batch-1, cost-per-token optimized), while training remains GPU-HBM dominated; such a split would reduce the portion of future inference compute that naturally monetizes through HBM content and could compress the incremental HBM-per-AI-dollar ratio. The key upside/neutral scenario for HBM is that the supply chain remains fully allocated regardless, with NVIDIA using any “freed” HBM to ship more high-end GPUs into training and long-context inference, especially as roadmaps increase HBM per GPU, sustaining robust aggregate bit demand even if inference becomes more heterogeneous. Conventional DRAM implications split into 2 channels: (1) DRAM wafer capacity diversion into HBM and (2) DDR content per server in AI clusters. Supplier commentary indicates that AI-driven memory demand is supporting elevated DRAM markets more broadly, and HBM production is resource-intensive versus conventional DRAM, tightening supply for DDR products in parallel. A meaningful NVIDIA pivot to an inference architecture that reduces HBM dependence could, at the margin, ease the most acute HBM-driven bottlenecks and allow memory manufacturers more flexibility in balancing DRAM mix, which could be modestly DDR-positive on the supply side (less crowding-out) even if it is DDR-neutral or slightly negative on the demand side (if per-node CPU/DDR requirements decline due to more efficient accelerator utilization). The dominant practical outcome is likely that DDR demand remains supported by broad AI server proliferation and increasing memory footprints at the system level (CPUs, networking stacks, caching layers, retrieval-augmented pipelines), while HBM remains the premium profit pool; therefore, any HBM displacement that increases total server volumes could indirectly keep DDR demand resilient even if DDR per accelerator is not rising materially. NAND flash implications are comparatively indirect and volume-driven rather than architecture-driven. Inference clusters require SSD capacity for model storage, container images, logging, and increasingly for fast local retrieval indices and embedding stores, but the storage footprint per unit of compute is typically smaller than in training pipelines that stage large datasets and checkpoints. If NVIDIA uses Groq to lower inference cost and latency enough to expand the total number of inference deployment locations (regional colocation, enterprise on-prem, sovereign footprints), aggregate SSD attach could rise through geographic fragmentation and replication of model artifacts across more sites, even if per-site storage is modest. The NAND effect is therefore likely to be demand-broadening and mix-positive (datacenter SSDs) but not a primary swing factor versus the macro AI capex cycle and consumer/device cycles. Hard disk drive (HDD) markets should see negligible direct sensitivity because nearline HDD demand is driven by bulk storage and cloud archiving economics, while inference acceleration choices primarily reshape compute and network layers; any HDD benefit would be a tertiary function of overall data center square footage expansion rather than a direct consequence of Groq silicon displacing GPUs. Optical networking implications require separating (1) intra-cluster back-end fabrics that connect accelerators and (2) front-end / data center interconnect (DCI) that connects sites and regions. Groq’s own positioning and third-party reporting suggest scaling beyond a single node or rack relies on high-bandwidth fabrics and, in some described configurations, optical interconnect scaling across hundreds of chips. If NVIDIA commercializes Groq at scale, 2 offsetting forces emerge: lower cost-per-token and improved latency could expand inference throughput and drive more east-west traffic, increasing demand for high-speed switching and optics; conversely, if Groq delivers materially higher utilization and tokens per unit of network bandwidth for certain workloads, the network required per served token could decline. Public NVIDIA materials already indicate an aggressive photonics roadmap aimed at scaling AI factories, including co-packaged optics (CPO) switches and explicit collaboration with Coherent and Lumentum in the silicon photonics supply chain. That linkage is important because it suggests that, independent of Groq, NVIDIA is already pushing optics integration deeper into the switch package to reduce power and increase resiliency; Groq increases the strategic incentive to reduce network power and latency if inference becomes even more distributed and latency-sensitive. For Lumentum and Coherent specifically, the net implication is less about “more optics versus fewer optics” and more about a shift in optics form factor and value capture. Co-packaged optics can reduce reliance on pluggable transceivers in some switch architectures while increasing demand for integrated photonic engines, lasers, fiber attach, packaging processes, and component-level supply. NVIDIA’s own announcements explicitly position Coherent and Lumentum as collaborators in creating the integrated silicon/optics process and supply chain for photonics switches. If Groq accelerates the transition to very large-scale fabrics (more endpoints, higher port speeds, tighter power envelopes), that tends to pull forward CPO adoption and amplifies demand for the underlying photonics components even if the conventional pluggable module TAM is structurally pressured over time. If Groq instead pushes inference toward smaller, more localized pods (closer to users, more regional colocation), that can be optics-positive for DCI and metro connectivity because more sites must be interconnected at high bandwidth with low latency, favoring coherent optics and high-speed interconnect between facilities. The principal risk for optics suppliers is timing and margin structure: a faster move to NVIDIA-driven integrated photonics could concentrate bargaining power and compress margins for commoditized transceiver modules while favoring suppliers with differentiated lasers, integration capability, and qualification depth in NVIDIA’s CPO ecosystem. AEC and copper interconnect implications hinge on whether Groq deployment increases the density of short-reach links inside racks and rows. High-speed copper remains structurally advantaged at very short distances on cost, power, and serviceability, but reaches become constrained as lane speeds and aggregate bandwidth rise, creating a role for active electrical cables (AECs), retimers, and signal-conditioning silicon. Credo explicitly positions its AEC products as enabling reliable lossless 800G connectivity for AI clusters, and the company has highlighted participation at NVIDIA GTC with content focused on extending PCIe/CXL using AECs, indicating relevance to next-generation system topologies that require longer reach and higher signal integrity than passive copper can deliver. If NVIDIA turns Groq into a widely deployed inference card or chassis product, the likely near-term effect is AEC-positive because (1) more inference throughput tends to increase top-of-rack connectivity requirements, (2) distributing inference across more racks and sites increases short-reach links per unit of delivered service, and (3) PCIe-attached accelerator architectures tend to require robust signal conditioning as systems move to PCIe 6.x and beyond. Groq workshop materials explicitly reference GroqCard and GroqNode form factors, reinforcing that PCIe-attached deployment has been central to Groq’s current packaging strategy. The main countervailing risk is that Groq’s deterministic chip-to-chip fabric could be implemented primarily through backplanes and direct board-level connectivity that reduces the need for merchant AECs inside the box; in that case, incremental AEC demand would concentrate more in rack-to-switch and node-to-fabric links rather than within-chassis chip fabrics. Astera Labs implications are connectivity-architecture sensitive and, on balance, skew positive if NVIDIA increases heterogeneity and disaggregation in AI systems. NVIDIA has publicly positioned NVLink Fusion as a pathway for partners to build semi-custom AI infrastructure and has explicitly identified Astera Labs as a partner in that ecosystem, with Astera describing NVLink-related solutions expanding its connectivity platform across PCIe, CXL, and Ethernet plus fleet observability software. A Groq acquisition increases the probability that NVIDIA offers a broader menu of accelerators (training GPUs, inference-focused ASICs) and therefore increases the importance of scalable, high-reliability connectivity, retiming, switching, and telemetry across mixed topologies. If Groq silicon remains PCIe-attached in many deployments, PCIe 6.x retimers/switches and active cable modules become more central, aligning with Astera’s core portfolio. If NVIDIA instead integrates Groq concepts into scale-up fabrics (NVLink-like domains) or uses Groq to expand into inference “appliances” that must be rapidly deployed in colocation environments, the need for standard-compliant, serviceable connectivity with strong RAS/telemetry increases, again aligning with Astera’s positioning. Power equipment and cooling implications for Vertiv and adjacent suppliers should be viewed through the lens of rack power density, cooling modality (air vs liquid), and site deployment model (hyperscale campuses vs distributed colocation/enterprise). Groq claims its LPU and rack designs are “air-cooled by design” and require no complex cooling and power infrastructure, and third-party reporting has described Groq’s approach as relying on parallelism across many lower-power units rather than extreme per-chip performance. If NVIDIA scales Groq as a mainstream inference platform, the mix of data center cooling spend could shift modestly away from the highest-density liquid-cooled racks toward more air-cooled or hybrid deployments, particularly for inference pods placed in existing facilities that cannot easily retrofit for very high rack heat flux. That would be a mix headwind for suppliers most levered exclusively to high-end liquid cooling attachments per rack, but it is not necessarily a volume headwind for Vertiv given the company’s broad exposure to both power and cooling infrastructure and the likelihood that total AI deployment locations expand. Vertiv’s own industry commentary emphasizes that AI racks require higher power-density UPS, batteries, power distribution equipment, and switchgear capable of handling rapid load transients, and that hybrid cooling systems will evolve across deployment environments. Those statements align with a world where inference growth increases the count of powered racks and raises the operational complexity of power delivery even if per-rack density is lower than the most extreme training clusters. The most material infrastructure impact may occur outside the rack and upstream of the data hall: grid interconnects, substations, transformers, switchgear, generators, and utility-scale generation additions. Recent regulatory actions in the U.S. highlight that projected data center demand is already driving large planned increases in electricity generation capacity, underscoring that power availability is a binding constraint. In that context, an inference architecture that lowers joules per token could reduce the power required per unit of inference delivered, but it can also accelerate demand by lowering cost and improving latency, increasing the total volume of inference served (a classic rebound effect). The net outcome is likely continued, elevated demand for power infrastructure even if efficiency improves, with the key swing factor being whether AI capex remains on a multi-year growth trajectory or enters a digestion phase. Other data center infrastructure implications include server/ODM mix, facility design standardization, and networking architecture choices. If NVIDIA positions Groq-based inference as a broadly distributable “standard server + accelerator” solution rather than as an integrated, liquid-cooled rack like GB200 NVL72, spend could shift toward more conventional air-cooled server designs, higher unit volumes of mainstream racks, and faster deployment in colocation footprints, increasing demand for modular power rooms, busways, and rapidly deployable cooling solutions. If NVIDIA instead integrates Groq into its “AI factory” paradigm, the primary effect is likely acceleration of dense back-end fabric build-outs and a faster push toward photonics switching, increasing demand for fiber plant, connectors, and integrated optics supply chains while potentially compressing the lifecycle of transitional architectures based on pluggable optics and mid-reach copper. NVIDIA’s stated roadmap toward co-packaged optics and silicon photonics switches is already oriented toward scaling to very large GPU counts; adding a high-end inference ASIC increases the strategic importance of power-efficient, low-latency fabrics because inference economics become increasingly sensitive to network overhead as compute cost declines. Across the covered segments, the most defensible base case is limited near-term dislocation and a medium-term increase in uncertainty around memory intensity per unit of inference growth. HBM faces the clearest relative risk from an HBM-less inference platform, but supply tightness and GPU training roadmaps reduce the probability of an absolute demand shock over the next 12–24 months. Optical, AEC/copper, and power/cooling are more likely to remain volume-supported because they scale with endpoint count, deployment fragmentation, and total data center footprint, and those tend to rise when inference becomes cheaper and more widely deployed. The highest-conviction second-order effect is a shift in infrastructure mix: incrementally more distributed inference deployments (favoring colocation power/cooling standardization, DCI optics, and serviceable short-reach interconnect) and a gradual migration from pluggable optics toward integrated photonics in back-end fabrics (favoring suppliers positioned in the CPO ecosystem).
9:00

$MU $SNDK $LITE $VRT NVIDIA and Groq: 2nd and 3rd Order Strategic Infrastructure Effects and Market Implications Public reporting indicates NVIDIA has agreed to acquire Groq for approximately $20,000,000,000 in cash, while excluding Groq’s nascent cloud business from the transaction perimeter. The reported carve-out materially constrains the immediate, direct linkage from the acquisition to incremental, NVIDIA-controlled data center capacity build-out because GroqCloud appears to be the principal channel through which Groq hardware is currently monetized at scale as a service. The infrastructure-market implications therefore depend primarily on post-close product strategy: whether NVIDIA (1) commercializes Groq silicon as a distinct inference product line and drives broad deployment through OEM/ODM channels and partners, (2) uses the acquisition mainly to absorb IP and talent while de-emphasizing standalone Groq hardware volumes, or (3) uses Groq technology to reshape NVIDIA’s own inference systems and networking roadmaps. The dominant transmission mechanism into memory, networking, and facility infrastructure markets is the degree to which NVIDIA shifts incremental inference deployments away from GPU architectures that are tightly coupled to external high-bandwidth memory (HBM) and toward Groq’s current architecture, which emphasizes large on-chip SRAM, deterministic compiler-scheduled execution, and direct chip-to-chip connectivity. Independent and company-published materials describe Groq’s current-generation approach as having no external memory, keeping weights and KV cache on-chip during processing, and requiring model sharding across multiple chips due to limited on-chip SRAM per device. That architectural choice is directionally HBM-negative on a per-accelerator basis and ambiguous for DRAM, NAND, networking, power, and cooling on a per-token basis because the design can reduce memory wall losses and tail-latency overhead while potentially increasing the number of chips and interconnect endpoints required to serve large models and long-context workloads. HBM implications are the most mechanically straightforward but should be framed as second-derivative rather than absolute. If Groq-class inference silicon meaningfully displaces NVIDIA GPU-based inference deployments, incremental HBM bit demand tied to inference growth could be reduced relative to a GPU-only baseline because Groq’s current approach does not appear to attach HBM stacks to each accelerator. However, current market structure suggests HBM remains supply-constrained and is being pulled by multiple vectors including continued GPU training scale and high-capacity inference configurations, with leading suppliers signaling tight conditions extending beyond 2026. In that environment, reduced inference-driven HBM intensity could primarily reallocate scarce HBM supply toward higher-end training and premium inference GPUs rather than creating an outright volume collapse, preserving high utilization of HBM capacity while potentially affecting the slope of pricing power and capacity expansion urgency over a multi-year horizon. The key downside scenario for the HBM complex would be a durable architectural bifurcation where “good-enough” inference shifts disproportionately to HBM-less ASICs across a broad swath of deployments (latency-sensitive, batch-1, cost-per-token optimized), while training remains GPU-HBM dominated; such a split would reduce the portion of future inference compute that naturally monetizes through HBM content and could compress the incremental HBM-per-AI-dollar ratio. The key upside/neutral scenario for HBM is that the supply chain remains fully allocated regardless, with NVIDIA using any “freed” HBM to ship more high-end GPUs into training and long-context inference, especially as roadmaps increase HBM per GPU, sustaining robust aggregate bit demand even if inference becomes more heterogeneous. Conventional DRAM implications split into 2 channels: (1) DRAM wafer capacity diversion into HBM and (2) DDR content per server in AI clusters. Supplier commentary indicates that AI-driven memory demand is supporting elevated DRAM markets more broadly, and HBM production is resource-intensive versus conventional DRAM, tightening supply for DDR products in parallel. A meaningful NVIDIA pivot to an inference architecture that reduces HBM dependence could, at the margin, ease the most acute HBM-driven bottlenecks and allow memory manufacturers more flexibility in balancing DRAM mix, which could be modestly DDR-positive on the supply side (less crowding-out) even if it is DDR-neutral or slightly negative on the demand side (if per-node CPU/DDR requirements decline due to more efficient accelerator utilization). The dominant practical outcome is likely that DDR demand remains supported by broad AI server proliferation and increasing memory footprints at the system level (CPUs, networking stacks, caching layers, retrieval-augmented pipelines), while HBM remains the premium profit pool; therefore, any HBM displacement that increases total server volumes could indirectly keep DDR demand resilient even if DDR per accelerator is not rising materially. NAND flash implications are comparatively indirect and volume-driven rather than architecture-driven. Inference clusters require SSD capacity for model storage, container images, logging, and increasingly for fast local retrieval indices and embedding stores, but the storage footprint per unit of compute is typically smaller than in training pipelines that stage large datasets and checkpoints. If NVIDIA uses Groq to lower inference cost and latency enough to expand the total number of inference deployment locations (regional colocation, enterprise on-prem, sovereign footprints), aggregate SSD attach could rise through geographic fragmentation and replication of model artifacts across more sites, even if per-site storage is modest. The NAND effect is therefore likely to be demand-broadening and mix-positive (datacenter SSDs) but not a primary swing factor versus the macro AI capex cycle and consumer/device cycles. Hard disk drive (HDD) markets should see negligible direct sensitivity because nearline HDD demand is driven by bulk storage and cloud archiving economics, while inference acceleration choices primarily reshape compute and network layers; any HDD benefit would be a tertiary function of overall data center square footage expansion rather than a direct consequence of Groq silicon displacing GPUs. Optical networking implications require separating (1) intra-cluster back-end fabrics that connect accelerators and (2) front-end / data center interconnect (DCI) that connects sites and regions. Groq’s own positioning and third-party reporting suggest scaling beyond a single node or rack relies on high-bandwidth fabrics and, in some described configurations, optical interconnect scaling across hundreds of chips. If NVIDIA commercializes Groq at scale, 2 offsetting forces emerge: lower cost-per-token and improved latency could expand inference throughput and drive more east-west traffic, increasing demand for high-speed switching and optics; conversely, if Groq delivers materially higher utilization and tokens per unit of network bandwidth for certain workloads, the network required per served token could decline. Public NVIDIA materials already indicate an aggressive photonics roadmap aimed at scaling AI factories, including co-packaged optics (CPO) switches and explicit collaboration with Coherent and Lumentum in the silicon photonics supply chain. That linkage is important because it suggests that, independent of Groq, NVIDIA is already pushing optics integration deeper into the switch package to reduce power and increase resiliency; Groq increases the strategic incentive to reduce network power and latency if inference becomes even more distributed and latency-sensitive. For Lumentum and Coherent specifically, the net implication is less about “more optics versus fewer optics” and more about a shift in optics form factor and value capture. Co-packaged optics can reduce reliance on pluggable transceivers in some switch architectures while increasing demand for integrated photonic engines, lasers, fiber attach, packaging processes, and component-level supply. NVIDIA’s own announcements explicitly position Coherent and Lumentum as collaborators in creating the integrated silicon/optics process and supply chain for photonics switches. If Groq accelerates the transition to very large-scale fabrics (more endpoints, higher port speeds, tighter power envelopes), that tends to pull forward CPO adoption and amplifies demand for the underlying photonics components even if the conventional pluggable module TAM is structurally pressured over time. If Groq instead pushes inference toward smaller, more localized pods (closer to users, more regional colocation), that can be optics-positive for DCI and metro connectivity because more sites must be interconnected at high bandwidth with low latency, favoring coherent optics and high-speed interconnect between facilities. The principal risk for optics suppliers is timing and margin structure: a faster move to NVIDIA-driven integrated photonics could concentrate bargaining power and compress margins for commoditized transceiver modules while favoring suppliers with differentiated lasers, integration capability, and qualification depth in NVIDIA’s CPO ecosystem. AEC and copper interconnect implications hinge on whether Groq deployment increases the density of short-reach links inside racks and rows. High-speed copper remains structurally advantaged at very short distances on cost, power, and serviceability, but reaches become constrained as lane speeds and aggregate bandwidth rise, creating a role for active electrical cables (AECs), retimers, and signal-conditioning silicon. Credo explicitly positions its AEC products as enabling reliable lossless 800G connectivity for AI clusters, and the company has highlighted participation at NVIDIA GTC with content focused on extending PCIe/CXL using AECs, indicating relevance to next-generation system topologies that require longer reach and higher signal integrity than passive copper can deliver. If NVIDIA turns Groq into a widely deployed inference card or chassis product, the likely near-term effect is AEC-positive because (1) more inference throughput tends to increase top-of-rack connectivity requirements, (2) distributing inference across more racks and sites increases short-reach links per unit of delivered service, and (3) PCIe-attached accelerator architectures tend to require robust signal conditioning as systems move to PCIe 6.x and beyond. Groq workshop materials explicitly reference GroqCard and GroqNode form factors, reinforcing that PCIe-attached deployment has been central to Groq’s current packaging strategy. The main countervailing risk is that Groq’s deterministic chip-to-chip fabric could be implemented primarily through backplanes and direct board-level connectivity that reduces the need for merchant AECs inside the box; in that case, incremental AEC demand would concentrate more in rack-to-switch and node-to-fabric links rather than within-chassis chip fabrics. Astera Labs implications are connectivity-architecture sensitive and, on balance, skew positive if NVIDIA increases heterogeneity and disaggregation in AI systems. NVIDIA has publicly positioned NVLink Fusion as a pathway for partners to build semi-custom AI infrastructure and has explicitly identified Astera Labs as a partner in that ecosystem, with Astera describing NVLink-related solutions expanding its connectivity platform across PCIe, CXL, and Ethernet plus fleet observability software. A Groq acquisition increases the probability that NVIDIA offers a broader menu of accelerators (training GPUs, inference-focused ASICs) and therefore increases the importance of scalable, high-reliability connectivity, retiming, switching, and telemetry across mixed topologies. If Groq silicon remains PCIe-attached in many deployments, PCIe 6.x retimers/switches and active cable modules become more central, aligning with Astera’s core portfolio. If NVIDIA instead integrates Groq concepts into scale-up fabrics (NVLink-like domains) or uses Groq to expand into inference “appliances” that must be rapidly deployed in colocation environments, the need for standard-compliant, serviceable connectivity with strong RAS/telemetry increases, again aligning with Astera’s positioning. Power equipment and cooling implications for Vertiv and adjacent suppliers should be viewed through the lens of rack power density, cooling modality (air vs liquid), and site deployment model (hyperscale campuses vs distributed colocation/enterprise). Groq claims its LPU and rack designs are “air-cooled by design” and require no complex cooling and power infrastructure, and third-party reporting has described Groq’s approach as relying on parallelism across many lower-power units rather than extreme per-chip performance. If NVIDIA scales Groq as a mainstream inference platform, the mix of data center cooling spend could shift modestly away from the highest-density liquid-cooled racks toward more air-cooled or hybrid deployments, particularly for inference pods placed in existing facilities that cannot easily retrofit for very high rack heat flux. That would be a mix headwind for suppliers most levered exclusively to high-end liquid cooling attachments per rack, but it is not necessarily a volume headwind for Vertiv given the company’s broad exposure to both power and cooling infrastructure and the likelihood that total AI deployment locations expand. Vertiv’s own industry commentary emphasizes that AI racks require higher power-density UPS, batteries, power distribution equipment, and switchgear capable of handling rapid load transients, and that hybrid cooling systems will evolve across deployment environments. Those statements align with a world where inference growth increases the count of powered racks and raises the operational complexity of power delivery even if per-rack density is lower than the most extreme training clusters. The most material infrastructure impact may occur outside the rack and upstream of the data hall: grid interconnects, substations, transformers, switchgear, generators, and utility-scale generation additions. Recent regulatory actions in the U.S. highlight that projected data center demand is already driving large planned increases in electricity generation capacity, underscoring that power availability is a binding constraint. In that context, an inference architecture that lowers joules per token could reduce the power required per unit of inference delivered, but it can also accelerate demand by lowering cost and improving latency, increasing the total volume of inference served (a classic rebound effect). The net outcome is likely continued, elevated demand for power infrastructure even if efficiency improves, with the key swing factor being whether AI capex remains on a multi-year growth trajectory or enters a digestion phase. Other data center infrastructure implications include server/ODM mix, facility design standardization, and networking architecture choices. If NVIDIA positions Groq-based inference as a broadly distributable “standard server + accelerator” solution rather than as an integrated, liquid-cooled rack like GB200 NVL72, spend could shift toward more conventional air-cooled server designs, higher unit volumes of mainstream racks, and faster deployment in colocation footprints, increasing demand for modular power rooms, busways, and rapidly deployable cooling solutions. If NVIDIA instead integrates Groq into its “AI factory” paradigm, the primary effect is likely acceleration of dense back-end fabric build-outs and a faster push toward photonics switching, increasing demand for fiber plant, connectors, and integrated optics supply chains while potentially compressing the lifecycle of transitional architectures based on pluggable optics and mid-reach copper. NVIDIA’s stated roadmap toward co-packaged optics and silicon photonics switches is already oriented toward scaling to very large GPU counts; adding a high-end inference ASIC increases the strategic importance of power-efficient, low-latency fabrics because inference economics become increasingly sensitive to network overhead as compute cost declines. Across the covered segments, the most defensible base case is limited near-term dislocation and a medium-term increase in uncertainty around memory intensity per unit of inference growth. HBM faces the clearest relative risk from an HBM-less inference platform, but supply tightness and GPU training roadmaps reduce the probability of an absolute demand shock over the next 12–24 months. Optical, AEC/copper, and power/cooling are more likely to remain volume-supported because they scale with endpoint count, deployment fragmentation, and total data center footprint, and those tend to rise when inference becomes cheaper and more widely deployed. The highest-conviction second-order effect is a shift in infrastructure mix: incrementally more distributed inference deployments (favoring colocation power/cooling standardization, DCI optics, and serviceable short-reach interconnect) and a gradual migration from pluggable optics toward integrated photonics in back-end fabrics (favoring suppliers positioned in the CPO ecosystem).

TheValueist

76,250 次观看 • 8 个月前

I asked Dan Martell to walk me through every level of making money with AI. He gave me the most simple, practical advice I've ever heard on this subject. Level 1 - Making $0 - $100k Level 2 - Making $1m - $10m Level 3 - Building a $10m++ enterprise. 0:00 Only 5% of the World Has Ever Paid for AI 0:46 The Easiest Thing to Sell With AI Right Now 1:56 The Marcus and Sophie Framework 4:24 Theory of Constraints (Right Problem to Solve) 5:33 What Is the Number One Business Constraint 7:13 How to Leave Your Job and Go All In 8:27 Business Is Simple Find a Problem and Solve It 9:08 Stop Getting Ready to Get Ready 9:33 The Sarah Story One Text and $10K 9:53 Pull Up Your Phone and Message Your Contacts 11:05 Dan's Son Gets His First Client at $800/Month 12:41 Best Employee vs. Best Employer 13:59 What Other Services Can You Sell With AI 14:44 Sales Is Not Talking It's Asking 17:01 What to Do When You Hate Your Business 18:40 Pain and Pleasure Are the Only Two Motivators 19:13 They Haven't Made It a Must Yet 20:29 Make It a Must Not a Nice to Have 21:06 The Jen Story and the Gasping Moment 22:17 How to Find Your First 10 to 15 Clients 28:38 The Personal Brand Play 33:06 Vision Is What AI Cannot Do 34:55 Hard for Computers Easy for Humans 36:13 Level 2 Making Your First Million With AI 37:18 The Replacement Ladder Framework 37:39 Admin First Then Delivery Then Marketing 39:09 Why Marketing Is the Biggest AI Category 39:32 Why You Should Keep Sales for Yourself 40:00 Level 5 Leadership and AI Agents 41:41 What a Fully AI Systems Business Looks Like 43:13 The Gym Owner With Three Locations 46:16 Shutting Down the Company for Two Days 46:37 Teaching the Whole Team to Code in Claude 49:28 Wayne the 62 Year Old Who Made $12K a Month 52:38 I Only Share What Actually Works 53:21 Whisper Flow and Talking to Your AI 56:41 Claude Chat Claude Coworker and Claude Code 57:57 The Claude Browser Extension 58:49 Claude Code Is Not Just for Developers 1:00:06 How to Migrate Your AI Memory Across Tools 1:01:08 Level 3 $1M to $10M and the Brand Play 1:02:05 Nobody Buys AI They Buy Trust 1:03:25 Brand Is Association and Association Is Trust 1:05:12 A Million Followers Is $10M in Activated Revenue 1:07:03 How to Keep AI From Becoming Slop 1:07:42 Human in the Loop 1:08:16 The 10 80 10 Rule and Why AI Is Now the 80 1:10:01 The Team FIRED Themselves 1:11:45 Dan's Free AI Curriculum for Your Team
1:13:20

I asked Dan Martell to walk me through every level of making money with AI. He gave me the most simple, practical advice I've ever heard on this subject. Level 1 - Making $0 - $100k Level 2 - Making $1m - $10m Level 3 - Building a $10m++ enterprise. 0:00 Only 5% of the World Has Ever Paid for AI 0:46 The Easiest Thing to Sell With AI Right Now 1:56 The Marcus and Sophie Framework 4:24 Theory of Constraints (Right Problem to Solve) 5:33 What Is the Number One Business Constraint 7:13 How to Leave Your Job and Go All In 8:27 Business Is Simple Find a Problem and Solve It 9:08 Stop Getting Ready to Get Ready 9:33 The Sarah Story One Text and $10K 9:53 Pull Up Your Phone and Message Your Contacts 11:05 Dan's Son Gets His First Client at $800/Month 12:41 Best Employee vs. Best Employer 13:59 What Other Services Can You Sell With AI 14:44 Sales Is Not Talking It's Asking 17:01 What to Do When You Hate Your Business 18:40 Pain and Pleasure Are the Only Two Motivators 19:13 They Haven't Made It a Must Yet 20:29 Make It a Must Not a Nice to Have 21:06 The Jen Story and the Gasping Moment 22:17 How to Find Your First 10 to 15 Clients 28:38 The Personal Brand Play 33:06 Vision Is What AI Cannot Do 34:55 Hard for Computers Easy for Humans 36:13 Level 2 Making Your First Million With AI 37:18 The Replacement Ladder Framework 37:39 Admin First Then Delivery Then Marketing 39:09 Why Marketing Is the Biggest AI Category 39:32 Why You Should Keep Sales for Yourself 40:00 Level 5 Leadership and AI Agents 41:41 What a Fully AI Systems Business Looks Like 43:13 The Gym Owner With Three Locations 46:16 Shutting Down the Company for Two Days 46:37 Teaching the Whole Team to Code in Claude 49:28 Wayne the 62 Year Old Who Made $12K a Month 52:38 I Only Share What Actually Works 53:21 Whisper Flow and Talking to Your AI 56:41 Claude Chat Claude Coworker and Claude Code 57:57 The Claude Browser Extension 58:49 Claude Code Is Not Just for Developers 1:00:06 How to Migrate Your AI Memory Across Tools 1:01:08 Level 3 $1M to $10M and the Brand Play 1:02:05 Nobody Buys AI They Buy Trust 1:03:25 Brand Is Association and Association Is Trust 1:05:12 A Million Followers Is $10M in Activated Revenue 1:07:03 How to Keep AI From Becoming Slop 1:07:42 Human in the Loop 1:08:16 The 10 80 10 Rule and Why AI Is Now the 80 1:10:01 The Team FIRED Themselves 1:11:45 Dan's Free AI Curriculum for Your Team

Grant

168,385 次观看 • 2 个月前

Inside the Billionaire Backed Prediction Markets Hedge Fund. Run by a 24-Year-Old. Camilo Saravia (camilo), founder of BlueWalker Capital, a systematic prediction markets fund backed by Daniel Howard of Halo Capital. "I don't want more capital. I'm extremely long our equity." We cover: - Why insider trading in prediction markets is terrible for liquidity and GOOD for society - Prediction markets as cash-backed truth in a world of AI slop and disinformation - Why he turned down the allocator question entirely, and the "Goldilocks zone" that makes a fund this size work - His research team's actual mission statement: "collapse the entropy of the internet into signal" - Trading Spotify streams and measuring how fast Mamdani viralizes vs Cuomo on TikTok - Why beating earnings has almost no correlation with the stock going up, and why only testing reveals that - Mention markets as literal next-word prediction, and how makers got sniped out - Hiring missionaries with a mercenary work style, and why every hire takes a pay cut vs Citadel, Jane Street, Wintermute - Daniel Howard's mandate: faster, more risk, more aggression. "They haven't backed me to print 7% APY" - The abundance mindset, from a kid with immigrant parents sitting across from generational wealth - A venture mindset applied to public equities: pulling the thread from free cash flow down to Glassdoor culture - Why you never need to be binarily right: buy at 20, sell at 40, never wait for resolution Highlights: (00:00) Intro (00:56) Taker vs maker, reflexive vs proactive: the strategy map (03:08) What makes an event contract different from an equity (04:53) Insider trading in prediction markets: bug or feature (08:18) Cash-backed truth in a world of AI slop (12:43) Where edge actually comes from (14:19) The dataset: billions of records a day (17:44) Building a money management business from an empty office (21:34) Why asset management competes with software as a business model (23:51) How to underwrite elite talent (25:50) Missionaries vs mercenaries, and why the tension is the point (30:54) Recruiting against Citadel money (and losing on salary every time) (37:32) "We can't compromise speed": the Daniel Howard mandate (44:45) The abundance mindset (48:28) "Why should I invest?" / "I don't want more capital" (50:12) Collapsing the entropy of the internet (51:04) Spotify streams, TikTok virality, and mention markets (53:17) The one data provider he'd long if he could (55:49) There's a business behind everything (01:00:07) How to find the real drivers in any market (01:04:04) A venture mindset applied to public equities (01:10:06) Prediction markets 101: where to actually start (01:12:05) Why you don't need to be binarily right (01:13:19) Final question: building personal edge against the models
1:16:45

Inside the Billionaire Backed Prediction Markets Hedge Fund. Run by a 24-Year-Old. Camilo Saravia (camilo), founder of BlueWalker Capital, a systematic prediction markets fund backed by Daniel Howard of Halo Capital. "I don't want more capital. I'm extremely long our equity." We cover: - Why insider trading in prediction markets is terrible for liquidity and GOOD for society - Prediction markets as cash-backed truth in a world of AI slop and disinformation - Why he turned down the allocator question entirely, and the "Goldilocks zone" that makes a fund this size work - His research team's actual mission statement: "collapse the entropy of the internet into signal" - Trading Spotify streams and measuring how fast Mamdani viralizes vs Cuomo on TikTok - Why beating earnings has almost no correlation with the stock going up, and why only testing reveals that - Mention markets as literal next-word prediction, and how makers got sniped out - Hiring missionaries with a mercenary work style, and why every hire takes a pay cut vs Citadel, Jane Street, Wintermute - Daniel Howard's mandate: faster, more risk, more aggression. "They haven't backed me to print 7% APY" - The abundance mindset, from a kid with immigrant parents sitting across from generational wealth - A venture mindset applied to public equities: pulling the thread from free cash flow down to Glassdoor culture - Why you never need to be binarily right: buy at 20, sell at 40, never wait for resolution Highlights: (00:00) Intro (00:56) Taker vs maker, reflexive vs proactive: the strategy map (03:08) What makes an event contract different from an equity (04:53) Insider trading in prediction markets: bug or feature (08:18) Cash-backed truth in a world of AI slop (12:43) Where edge actually comes from (14:19) The dataset: billions of records a day (17:44) Building a money management business from an empty office (21:34) Why asset management competes with software as a business model (23:51) How to underwrite elite talent (25:50) Missionaries vs mercenaries, and why the tension is the point (30:54) Recruiting against Citadel money (and losing on salary every time) (37:32) "We can't compromise speed": the Daniel Howard mandate (44:45) The abundance mindset (48:28) "Why should I invest?" / "I don't want more capital" (50:12) Collapsing the entropy of the internet (51:04) Spotify streams, TikTok virality, and mention markets (53:17) The one data provider he'd long if he could (55:49) There's a business behind everything (01:00:07) How to find the real drivers in any market (01:04:04) A venture mindset applied to public equities (01:10:06) Prediction markets 101: where to actually start (01:12:05) Why you don't need to be binarily right (01:13:19) Final question: building personal edge against the models

Ethan Kho

127,051 次观看 • 1 个月前

BITCOIN AND ALTCOINS: BIGGEST CYCLE LOADING!!!! 🚨🚨🚨 00:00 Intro - BTC still around 200 WMA 00:10 BTC weekly chart - Still in the bear trend and buy zone 00:34 Norway sovereig wealth fund bought more MSTR 01:04 Saylor BTC video 01:58 Saylor BTC video 2 02:32 Stock market still in bull trend 03:05 SP500 video - victory cry 03:40 Time to think more bullish - BTC will shine 03:56 Ansem video - Start paying attention to crypto again 04:57 Bullmania AI - FOMO app and PUMP fun 06:34 Pump fun news - Callout rewards 07:45 SOL copying Robinhood - Stonkfun. Stocks pump memecoins 08:51 TREE protocol on top of LaunchOnSF - DeFi Summer of 2020 revival 13:41 Is it time to ape in? Dip your toes a bit 14:05 Welcome to the chat 14:19 Shout out to Bullmania AI - Sign up for the waitlist 15:26 Working on voice AI update - Ivan voice attempt 17:36 Gemini iincrease bitcoin holdings 17:45 Joe Rogan video - Shocked about Trump crypto earnings 18:58 New admnistration will bring problems for crypto 19:18 SpaceX and Tesla - Bullish since August 7th 20:20 Will Tesla be merged into SpaceX? 20:58 Tesla daily chart - potential price in 21:20 Advice to Elon - One ticker is better than two. Best for the pump and the mission 25:25 If merger happens - Premium will be paid 26:20 Metaplnet moved 4K BTC 26:29 OpenAI trouble in paradise, people leaving. 27:00 ChatGPT vs Claude 28:54 Trezor hack of customer data 32:00 What can happen after the hack 34:14 Tether passed the audit 34:39 Some quick news items 35:15 SPX prediction plebe noob mistake - A rant 36:53 Memory rally caused by fundamentals - video 39:57 Tom Lee video - Stocks going up on good news 41:56 Shout out to Bullmania - You need it before the bull market 44:22 Ansem vido - On new cycle. Get on chain 45:16 CashCat holding nicely 45:25 TOAD meme 45:42 WEB3 gaming recap - Remember Sandbox and the Metaverse fugazzi? 48:21 Ride the trend and max extract 48:43 Don't be too smart - You'll miss everything 49:11 Robinhod chain tVL close to 1B 49:28 Comparing AI bubble to Dot com bubble 50:11 Michael Burry predicting another crash 51:05 Q and A 51:27 Shout out to Bullmania 52:22 Bullmania AI - Michael Burry in full Cassandra mode 54:45 The curse of Cassandra 59:50 Outro - Bullmania user story
1:03:55

BITCOIN AND ALTCOINS: BIGGEST CYCLE LOADING!!!! 🚨🚨🚨 00:00 Intro - BTC still around 200 WMA 00:10 BTC weekly chart - Still in the bear trend and buy zone 00:34 Norway sovereig wealth fund bought more MSTR 01:04 Saylor BTC video 01:58 Saylor BTC video 2 02:32 Stock market still in bull trend 03:05 SP500 video - victory cry 03:40 Time to think more bullish - BTC will shine 03:56 Ansem video - Start paying attention to crypto again 04:57 Bullmania AI - FOMO app and PUMP fun 06:34 Pump fun news - Callout rewards 07:45 SOL copying Robinhood - Stonkfun. Stocks pump memecoins 08:51 TREE protocol on top of LaunchOnSF - DeFi Summer of 2020 revival 13:41 Is it time to ape in? Dip your toes a bit 14:05 Welcome to the chat 14:19 Shout out to Bullmania AI - Sign up for the waitlist 15:26 Working on voice AI update - Ivan voice attempt 17:36 Gemini iincrease bitcoin holdings 17:45 Joe Rogan video - Shocked about Trump crypto earnings 18:58 New admnistration will bring problems for crypto 19:18 SpaceX and Tesla - Bullish since August 7th 20:20 Will Tesla be merged into SpaceX? 20:58 Tesla daily chart - potential price in 21:20 Advice to Elon - One ticker is better than two. Best for the pump and the mission 25:25 If merger happens - Premium will be paid 26:20 Metaplnet moved 4K BTC 26:29 OpenAI trouble in paradise, people leaving. 27:00 ChatGPT vs Claude 28:54 Trezor hack of customer data 32:00 What can happen after the hack 34:14 Tether passed the audit 34:39 Some quick news items 35:15 SPX prediction plebe noob mistake - A rant 36:53 Memory rally caused by fundamentals - video 39:57 Tom Lee video - Stocks going up on good news 41:56 Shout out to Bullmania - You need it before the bull market 44:22 Ansem vido - On new cycle. Get on chain 45:16 CashCat holding nicely 45:25 TOAD meme 45:42 WEB3 gaming recap - Remember Sandbox and the Metaverse fugazzi? 48:21 Ride the trend and max extract 48:43 Don't be too smart - You'll miss everything 49:11 Robinhod chain tVL close to 1B 49:28 Comparing AI bubble to Dot com bubble 50:11 Michael Burry predicting another crash 51:05 Q and A 51:27 Shout out to Bullmania 52:22 Bullmania AI - Michael Burry in full Cassandra mode 54:45 The curse of Cassandra 59:50 Outro - Bullmania user story

Ivan on Tech 🍳📈💰 Head Trader @ Bullmania

19,469 次观看 • 1 个月前

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.
0:42

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.

Akshay 🚀

341,746 次观看 • 24 天前

Research suggests that up to 40% of cancer cases could be prevented through lifestyle changes. The evidence is now overwhelming: exercise is not just supportive—it’s a therapeutic intervention that recalibrates tumor biology, enhances treatment tolerance, and improves survival outcomes. Today’s interview features Dr. Kerry Courneya. With over 600 peer-reviewed studies, he is one of the most influential figures in exercise oncology. Even if you aren't someone who has personally experienced cancer in one form or another, you need to watch this episode. Episode 99 is Available now on X, YouTube, Spotify, and Apple Podcasts. Chapters: 0:00 - Introduction 1:47 - Why exercise should be effortful 2:33 - How to meaningfully reduce risk of cancer 6:22 - What type of exercise is best? 7:59 - How exercise reduces risk—even for smokers and the obese 10:48 - Weekend-only exercise 13:49 - 150 vs. 300 minutes per week (more is better—up to a point) 16:03 - Why pre-diagnosis exercise matters 19:09 - Why resilience to cancer treatment starts with exercise 21:01 - Why low muscle mass drives cancer death 23:58 - Why BMI fails to measure true obesity 27:51 - Why daily activity isn't enough (structured exercise matters) 29:34 - Breaking up sedentary time—do ‘exercise snacks’ help? 31:50 - Supplements vs. exercise 32:32 - Where exercise fits with chemo and immunotherapy 35:30 - Why rest is not the best medicine 41:20 - Aerobic vs. resistance 42:13 - How weight training improves 'chemo completion' 44:41 - Why exercise creates vulnerability in cancer cells (limitations do apply) 47:09 - Why exercise might be crucial for tumor elimination 53:03 - Why cardio may be better at clearing tumor cells 56:18 - When cancer spreads quickly—and when it doesn't 57:43 - Why liquid biopsies may prevent over-treatment 1:02:56 - Exercise-sensitive vs. exercise-resistant cancers 1:06:06 - Prostate cancer therapy—why strength training matters 1:08:10 - When exercise is the only therapy—does it work? 1:09:26 - Why HIIT reduces PSA in prostate cancer 1:11:40 - Avoiding overtreatment—can exercise buy you time? 1:12:00 - Why high-intensity exercise boosts anti-cancer biology 1:13:11 - Turning a diagnosis into a wake-up call 1:16:11 - Why oncologists are rethinking exercise 1:18:50 - Why exercise eases anxiety about cancer—proven psychological benefits 1:25:00 - Before, during, and after treatment 1:27:02 - Why exercise is unique among cancer therapies 1:28:16 - Why cancer patients stop exercising—the risky mistake almost everyone makes 1:30:41 - How to get sedentary cancer patients exercising (realistically) 1:33:15 - The $1 million case for including exercise 1:34:56 - Why recurrence trials haven't convinced doctors—yet 1:37:36 - The bottom-line message 1:37:55 - The myth of a cancer panacea (exercise included) 1:44:07 - What's the best $50 investment for staying active? 1:44:40 - Only 15 minutes per day—what’s the best anti-cancer exercise?
1:48:23

Research suggests that up to 40% of cancer cases could be prevented through lifestyle changes. The evidence is now overwhelming: exercise is not just supportive—it’s a therapeutic intervention that recalibrates tumor biology, enhances treatment tolerance, and improves survival outcomes. Today’s interview features Dr. Kerry Courneya. With over 600 peer-reviewed studies, he is one of the most influential figures in exercise oncology. Even if you aren't someone who has personally experienced cancer in one form or another, you need to watch this episode. Episode 99 is Available now on X, YouTube, Spotify, and Apple Podcasts. Chapters: 0:00 - Introduction 1:47 - Why exercise should be effortful 2:33 - How to meaningfully reduce risk of cancer 6:22 - What type of exercise is best? 7:59 - How exercise reduces risk—even for smokers and the obese 10:48 - Weekend-only exercise 13:49 - 150 vs. 300 minutes per week (more is better—up to a point) 16:03 - Why pre-diagnosis exercise matters 19:09 - Why resilience to cancer treatment starts with exercise 21:01 - Why low muscle mass drives cancer death 23:58 - Why BMI fails to measure true obesity 27:51 - Why daily activity isn't enough (structured exercise matters) 29:34 - Breaking up sedentary time—do ‘exercise snacks’ help? 31:50 - Supplements vs. exercise 32:32 - Where exercise fits with chemo and immunotherapy 35:30 - Why rest is not the best medicine 41:20 - Aerobic vs. resistance 42:13 - How weight training improves 'chemo completion' 44:41 - Why exercise creates vulnerability in cancer cells (limitations do apply) 47:09 - Why exercise might be crucial for tumor elimination 53:03 - Why cardio may be better at clearing tumor cells 56:18 - When cancer spreads quickly—and when it doesn't 57:43 - Why liquid biopsies may prevent over-treatment 1:02:56 - Exercise-sensitive vs. exercise-resistant cancers 1:06:06 - Prostate cancer therapy—why strength training matters 1:08:10 - When exercise is the only therapy—does it work? 1:09:26 - Why HIIT reduces PSA in prostate cancer 1:11:40 - Avoiding overtreatment—can exercise buy you time? 1:12:00 - Why high-intensity exercise boosts anti-cancer biology 1:13:11 - Turning a diagnosis into a wake-up call 1:16:11 - Why oncologists are rethinking exercise 1:18:50 - Why exercise eases anxiety about cancer—proven psychological benefits 1:25:00 - Before, during, and after treatment 1:27:02 - Why exercise is unique among cancer therapies 1:28:16 - Why cancer patients stop exercising—the risky mistake almost everyone makes 1:30:41 - How to get sedentary cancer patients exercising (realistically) 1:33:15 - The $1 million case for including exercise 1:34:56 - Why recurrence trials haven't convinced doctors—yet 1:37:36 - The bottom-line message 1:37:55 - The myth of a cancer panacea (exercise included) 1:44:07 - What's the best $50 investment for staying active? 1:44:40 - Only 15 minutes per day—what’s the best anti-cancer exercise?

Dr. Rhonda Patrick

220,075 次观看 • 1 年前

$AMD $5 Trillion is Inevitable LT| Agentic AI🧵 Agentic AI is the new $5 Trillion TAM 🚨🚨🚨 This thead will do Comp with $INTC and how to quantify this massive Agentic AI demand spike, and forcing Jensen to rush a CPU design. Global Agentic AI Market size is estimated to be $3-$5Trillion TAM by 2030(McKinsey) Quantifying the demand from agentic AI for AMD involves assessing the broader market growth for agentic systems, their unique computational requirements (particularly for CPUs in orchestration and reasoning tasks), and AMD's positioning very well through products like EPYC processors and partnerships. AMD EPYC Venice is the most superior choice in 2026-2027 for most Agentic AI workloads Agentic AI refers to autonomous AI agents that perform multi-step tasks, involving sequential logic, tool integration, and decision-making workloads that heavily rely on CPUs for handling orchestration, memory management, and context switching, rather than just GPU-parallelized training or batch inference. Agentic AI is often cited as 40-100x more "hungry" than traditional AI due to its continuous, 24/7 operation and complex workflows. This stems from factors like chain-of-thought reasoning (multiple LLM calls per query), API/tool interactions, memory management, and orchestration loops, which can generate 10-100x more tokens and require real-time responsiveness. For example, a single agentic query might trigger 5-20 model inferences, making it 10-20x more compute-intensive than simple chatbots, and the always-on nature compounds this to 40-100x overall. Nvidia's CEO has highlighted this as driving "easily 100x more computation" for inference in agentic/reasoning setups. AMD's EPYC Venice (6th Gen EPYC, codenamed "Venice") and Intel's Xeon 7 Diamond Rapids represent the pinnacle of server CPU technology in 2026, both targeting high-performance data center workloads like AI inference, agentic AI orchestration, cloud computing, and HPC. Venice builds on AMD's Zen 6 architecture, emphasizing core density and efficiency, while Diamond Rapids leverages Intel's Panther Cove P-cores for balanced performance. Both chips adopt similar advancements like 16-channel DDR5 memory and PCIe Gen 6, but differ in core counts, process nodes, and overall design philosophy. Intel has faced acute supply constraints across its Xeon lineup, including legacy nodes (Intel 7/3) and the ramping 18A process for next-gen parts. Intel shortage is expected with lead times up to 6 months or longer. 1. AMD EPYC Venice vs Intel Xeon 7 Diamond Rapids Architecture AMD: Zen 6 chiplet design with 8 CCDs and dual IODs Intel: Panther Cove P-cores; multi-die architecture with 4 compute tiles Core/Thread Count AMD: Up to 256 cores / 512 threads (Zen 6c variant) Intel: Up to 192 cores / 192 threads Process Node AMD: TSMC N2 (2nm) Intel: Intel 18A (1.8nm-class); in-house fab Memory Support AMD: 16-channel DDR5; up to 1.6 TB/s bandwidth. Intel: 16-channel DDR5 ; up to 1.6 TB/s bandwidth I/O and Connectivity AMD: PCIe Gen 6 (up to 128 lanes); twice the CPU-to-GPU bandwidth Intel: PCIe Gen 6 (up to 128 lanes); LGA 9324 socket Power (TDP) AMD: Starting 400-500W, potentially lower due to efficiency gains from TSMC 2nm Intel: Starting 400-500W, as it targets competitive efficiency Performance Projections AMD: Up to 70% uplift vs. 5th Gen Turin (1.7x in multi-threaded/AI tasks) Intel: ~40% faster than Granite Rapids (Xeon 6, 128-core). Lags AMD in per-core perf and 40-50% behind Venice core-for-core comp Target Workloads AMD: AI inference/orchestration, HPC, cloud virtualization. Partnerships Intel: Hyperscale AI, general enterprise. Custom silicon Pricing: AMD: estimated $10k-$20k for top SKUs Intel: estimated $8-$18k Availability: AMD: Significant Ramp H2 2026 due to higher allocation from TSMC Intel: H1-H2 2026 delayed, but trying to catch up Overall: ~Venice's 256 cores provide a 33% edge over Diamond Rapids' 192, making it superior for massively parallel tasks like AI training/inference or virtualization ~TSMC's N2 vs. Intel 18A debates rage on which is "better," but AMD's mature chiplet approach yields better density ( 32 cores/CCD vs. Intel's 48/tile). Venice's redesign reduces latency, aiding agentic AI where CPUs handle orchestration ~ Early projections show Venice widening AMD's lead matching or exceeding Diamond Rapids' perf with fewer watts in multi-threaded benchmarks. Intel's no-SMT design (to prioritize AI) handicaps it vs. AMD's 512 threads, though Clearwater Forest (E-core) could compete in density-focused niches. ~Power & Cooling: Both push above 400-500W, demanding liquid cooling. ~AMD been taking market share now above 40%. AMD EPYC Venice emerges as the superior choice in 2026 for most server workloads. Its higher core/thread count (256/512 vs. 192/192), stronger per-core performance, and architecture optimized for AI-driven tasks (agentic orchestration with GPU integration) provide decisive advantages in throughput, scalability, and efficiency. Projections indicate Venice delivering 1.7x the performance of prior gens while widening the gap over Intel ( 40-70% leads in multi-threaded benchmarks). AMD's fabless model with TSMC ensures reliable scaling, and its ecosystem ( open ROCm) appeals to AI adopters. Intel's Diamond Rapids is competitive in single-threaded enterprise apps and custom hyperscale ( NVLink), with potential fab advantages for supply/security. However, without SMT and lower density, it falls short in core-for-core battles—exposing Intel to another generation of AMD dominance unless 18A yields surprise efficiency gains. For data centers prioritizing raw compute ( AI, HPC), Venice wins; for Intel-centric ecosystems or specialized I/O, Diamond Rapids holds ground. Real benchmarks post-launch will confirm, but logic points to AMD pulling ahead. 2. Market size , Potential Revenue and Supply Global Agentic AI market size is projected to be $3-$5 Trillion by 2030 according to McKinsey, where consensus points to 40-50% CAGR driven by small to large enterprise demand. I also wrote a full thread on how and why Agentic AI is so explosive that AMD will blow all anlaysts estimate for subscribers. Link below if you are interested. AMD's data center segment hit a record $5.4B in Q4 2025 (up 39% YoY), with EPYC shipments ramping due to agentic demand. With 2GW of deployment in H2 2026, AMD AI data center revenue has $40-$50B+ at the lowest or most conservative projection; or Total Revenue in the $77-$94B For FY2026. However, Agentic AI massive demand spike could send EPYC revenue 3x to 4x in the next few years, potentially surpassing MI series GPU demand as enterprises prioritize CPU-dense Rack setups. This is pushing $NVDA Jensen to rush a CPU design and acquired Groq, a new CPU player due to this massive TAM. Noted that this is just popping just in weeks, highlighting we are just so early in this AI Supercycle and the pace of adoption is insane, and clearly productivity will skyrocket. Why? Because Agentic AI is 24/7 Smart AI agent working for you or your businesses is a mad compelling, and it is estimated to be 40-100x more Inference Hugnry! Many experts already said it is impossible to project this kind of Inference Demand. AI CapEx is expected to ramp up even more in 2027-2028-2029 and 2030 as Global Agentic AI is going to scale to $3-$5 Trillion TAM by 2030. The nature of Agentic is driving higher CPU/GPU ratio, with CPUs handling 50-90% of Agentic workflows. For example, The current Helios Rack: 18 compute trays per rack with 72 GPUs + 18 CPUs. The beauty of this $META and $AMD long term partnership is, that it is absolutely flexible to adjust racks to higher CPU rato or equal to service different needs. Helios rack can be easily swap to 2 GPUs 2CPUs or even CPUs only trays for dedicated orchestration/head nodes. You see, the beauty of this open rack-scale is flexibility and evolvability. If Agentic AI demand pushes much higher, AMD should be able to adjust variant trays without abandoning Heilos Rack. We can't talk just about massive Agentic AI demand without talking about the Supply side or TSMC. TSMC, AMD's primary foundry for advanced nodes ( Zen 6/Venice on N2/2nm), is addressing AI-driven shortages through massive expansions. TSMC accelerates fab construction with up to 10 facilities targeted for 2026. TSMC is accelerating its domestic manufacturing expansion, with industry sources indicating that as many as ten fabs could be under construction or preparing to begin operations across Taiwan’s major science parks. TSMC Capex: $52-56B in 2026 (up 37% YoY), with $45B already approved for new/upgraded capacities. 70-80% for advanced processes (2nm/A16), 10-20% for packaging (CoWoS quadrupling to 120-140K wafers/month by late 2026). In addition, Taiwanese companies (led by TSMC) commit to at least $250B in direct investments in US-based advanced semiconductor, AI, and energy production/innovation capacity.Taiwan provides $250B in government credit guarantees to facilitate additional investments and build a full US semiconductor ecosystem (including industrial parks). TSMC completed a second land purchase in Arizona (January 2026) for gigafab scaling, with an additional $100B+ (potentially four more modules) to further expand and qualify for tariff exemptions. AMD with secured 12GW from OpenAI and $META and massive Agentic AI will mean higher priority acess to 20-30% more wafers on TSMC advanced nodes, as TSMC has multi-year agreements with AMD for AI chips. Dr. C. C. Wei, CEO of TSMC quote: "I spend a lot of time in the last three or four months talking to my customer and then customers. Customer. I want to make sure that my customers demand are real. I talk to those cloud service providers, all of them. Their answer is. I'm quite satisfied with their answer. Actually they show me the evidence that the AI really help their business. So they grow their business successfully and he or she in their financial return. So I also double check their financial status. They are very rich." Amid shortages, the US buildout ensures AMD can ramp production of Instinct GPUs and EPYC CPUs without the constraints hitting competitors like Intel. By diversifying away from Taiwan (85% of advanced nodes today), the agreement mitigates supply disruptions, ensuring stable flows for AMD's chips. Scaling production and securing supply will matter for AMD the most in the next 5-10 years growth. The growth could be 80-100% YoY or higher; or it could be in the 60%. The aggressive TSMC supply ramp is reassuring the higher growth point. Conclusion: AMD stands at a pivotal inflection point in 2026, where the explosive rise of agentic AI demanding 40-100x more inference compute through its 24/7, multi-step orchestration positions the company to potentially triple its EPYC CPU revenue to $45-60B+ by 2028 while scaling Instinct GPUs to tens of billions annually by 2027. Agentic AI demand could push AI CapEx closer to $1 Trillion in 2027, far higher than most estimates. Dr. Lisa Su, AMD's visionary CEO, is masterfully securing supply to harness this massive demand by prioritizing operational execution and deep TSMC collaboration, ensuring readiness for the second-half 2026 AI ramp. Dr. Su has explicitly called out surging EPYC demand for agentic tasks where CPUs power head nodes and traditional workloads alongside GPUs while guiding for data center dominance through proactive capacity planning and partnerships like Nutanix ($150M investment for open agentic platforms) or providing tens of millions CPUs for OpenAI, $META, $ORCL, $AMZN, $MSFT, $GOOGL and others. Her strategy includes multi-year TSMC agreements for advanced nodes (N2 for Venice CPUs and future Instincts), diversifying beyond Taiwan to mitigate risks, and unveiling innovations like the MI455X GPU at CES 2026, which she touted as enabling "the next trillion-dollar market opportunity" in physical AI. Dr. Su's forward-looking vision predicting AI reaching 5 billion users emphasizes "AI everywhere," backed by hardware like Ryzen AI chips, all while declaring demand "going through the roof" and committing to scale without bottlenecks. TSMC's aggressive ramp-up, fueled by $52-56B in 2026 capex (up 37% YoY) and 10+ new fabs across Taiwan, the US (Arizona cluster expanding to 6+ modules with $165B+ investment), Japan, and Europe, provides profound reassurance for AMD's supply stability. The January 2026 US-Taiwan agreement committing $250B in investments and credit guarantees for US reshoring accelerates this, granting tariff relief (15% rates with 1.5-2.5x exemptions) tied to capacity buildouts, enabling TSMC to potentially double output over the decade to meet AI wafer hunger. This translates to 20-30% higher wafer allocations on key nodes, sidestepping Intel-like shortages and empowering Dr. Su's team to deliver on hyperscaler demands without disruption. Ultimately, this synergy cements AMD's leadership in the agentic era, promising sustained growth, $5T+ valuations at scale, and a resilient path forward as AI reshapes the world. This is NOT Financial Advice! Video source: AMD CES 2026
2:49

$AMD $5 Trillion is Inevitable LT| Agentic AI🧵 Agentic AI is the new $5 Trillion TAM 🚨🚨🚨 This thead will do Comp with $INTC and how to quantify this massive Agentic AI demand spike, and forcing Jensen to rush a CPU design. Global Agentic AI Market size is estimated to be $3-$5Trillion TAM by 2030(McKinsey) Quantifying the demand from agentic AI for AMD involves assessing the broader market growth for agentic systems, their unique computational requirements (particularly for CPUs in orchestration and reasoning tasks), and AMD's positioning very well through products like EPYC processors and partnerships. AMD EPYC Venice is the most superior choice in 2026-2027 for most Agentic AI workloads Agentic AI refers to autonomous AI agents that perform multi-step tasks, involving sequential logic, tool integration, and decision-making workloads that heavily rely on CPUs for handling orchestration, memory management, and context switching, rather than just GPU-parallelized training or batch inference. Agentic AI is often cited as 40-100x more "hungry" than traditional AI due to its continuous, 24/7 operation and complex workflows. This stems from factors like chain-of-thought reasoning (multiple LLM calls per query), API/tool interactions, memory management, and orchestration loops, which can generate 10-100x more tokens and require real-time responsiveness. For example, a single agentic query might trigger 5-20 model inferences, making it 10-20x more compute-intensive than simple chatbots, and the always-on nature compounds this to 40-100x overall. Nvidia's CEO has highlighted this as driving "easily 100x more computation" for inference in agentic/reasoning setups. AMD's EPYC Venice (6th Gen EPYC, codenamed "Venice") and Intel's Xeon 7 Diamond Rapids represent the pinnacle of server CPU technology in 2026, both targeting high-performance data center workloads like AI inference, agentic AI orchestration, cloud computing, and HPC. Venice builds on AMD's Zen 6 architecture, emphasizing core density and efficiency, while Diamond Rapids leverages Intel's Panther Cove P-cores for balanced performance. Both chips adopt similar advancements like 16-channel DDR5 memory and PCIe Gen 6, but differ in core counts, process nodes, and overall design philosophy. Intel has faced acute supply constraints across its Xeon lineup, including legacy nodes (Intel 7/3) and the ramping 18A process for next-gen parts. Intel shortage is expected with lead times up to 6 months or longer. 1. AMD EPYC Venice vs Intel Xeon 7 Diamond Rapids Architecture AMD: Zen 6 chiplet design with 8 CCDs and dual IODs Intel: Panther Cove P-cores; multi-die architecture with 4 compute tiles Core/Thread Count AMD: Up to 256 cores / 512 threads (Zen 6c variant) Intel: Up to 192 cores / 192 threads Process Node AMD: TSMC N2 (2nm) Intel: Intel 18A (1.8nm-class); in-house fab Memory Support AMD: 16-channel DDR5; up to 1.6 TB/s bandwidth. Intel: 16-channel DDR5 ; up to 1.6 TB/s bandwidth I/O and Connectivity AMD: PCIe Gen 6 (up to 128 lanes); twice the CPU-to-GPU bandwidth Intel: PCIe Gen 6 (up to 128 lanes); LGA 9324 socket Power (TDP) AMD: Starting 400-500W, potentially lower due to efficiency gains from TSMC 2nm Intel: Starting 400-500W, as it targets competitive efficiency Performance Projections AMD: Up to 70% uplift vs. 5th Gen Turin (1.7x in multi-threaded/AI tasks) Intel: ~40% faster than Granite Rapids (Xeon 6, 128-core). Lags AMD in per-core perf and 40-50% behind Venice core-for-core comp Target Workloads AMD: AI inference/orchestration, HPC, cloud virtualization. Partnerships Intel: Hyperscale AI, general enterprise. Custom silicon Pricing: AMD: estimated $10k-$20k for top SKUs Intel: estimated $8-$18k Availability: AMD: Significant Ramp H2 2026 due to higher allocation from TSMC Intel: H1-H2 2026 delayed, but trying to catch up Overall: ~Venice's 256 cores provide a 33% edge over Diamond Rapids' 192, making it superior for massively parallel tasks like AI training/inference or virtualization ~TSMC's N2 vs. Intel 18A debates rage on which is "better," but AMD's mature chiplet approach yields better density ( 32 cores/CCD vs. Intel's 48/tile). Venice's redesign reduces latency, aiding agentic AI where CPUs handle orchestration ~ Early projections show Venice widening AMD's lead matching or exceeding Diamond Rapids' perf with fewer watts in multi-threaded benchmarks. Intel's no-SMT design (to prioritize AI) handicaps it vs. AMD's 512 threads, though Clearwater Forest (E-core) could compete in density-focused niches. ~Power & Cooling: Both push above 400-500W, demanding liquid cooling. ~AMD been taking market share now above 40%. AMD EPYC Venice emerges as the superior choice in 2026 for most server workloads. Its higher core/thread count (256/512 vs. 192/192), stronger per-core performance, and architecture optimized for AI-driven tasks (agentic orchestration with GPU integration) provide decisive advantages in throughput, scalability, and efficiency. Projections indicate Venice delivering 1.7x the performance of prior gens while widening the gap over Intel ( 40-70% leads in multi-threaded benchmarks). AMD's fabless model with TSMC ensures reliable scaling, and its ecosystem ( open ROCm) appeals to AI adopters. Intel's Diamond Rapids is competitive in single-threaded enterprise apps and custom hyperscale ( NVLink), with potential fab advantages for supply/security. However, without SMT and lower density, it falls short in core-for-core battles—exposing Intel to another generation of AMD dominance unless 18A yields surprise efficiency gains. For data centers prioritizing raw compute ( AI, HPC), Venice wins; for Intel-centric ecosystems or specialized I/O, Diamond Rapids holds ground. Real benchmarks post-launch will confirm, but logic points to AMD pulling ahead. 2. Market size , Potential Revenue and Supply Global Agentic AI market size is projected to be $3-$5 Trillion by 2030 according to McKinsey, where consensus points to 40-50% CAGR driven by small to large enterprise demand. I also wrote a full thread on how and why Agentic AI is so explosive that AMD will blow all anlaysts estimate for subscribers. Link below if you are interested. AMD's data center segment hit a record $5.4B in Q4 2025 (up 39% YoY), with EPYC shipments ramping due to agentic demand. With 2GW of deployment in H2 2026, AMD AI data center revenue has $40-$50B+ at the lowest or most conservative projection; or Total Revenue in the $77-$94B For FY2026. However, Agentic AI massive demand spike could send EPYC revenue 3x to 4x in the next few years, potentially surpassing MI series GPU demand as enterprises prioritize CPU-dense Rack setups. This is pushing $NVDA Jensen to rush a CPU design and acquired Groq, a new CPU player due to this massive TAM. Noted that this is just popping just in weeks, highlighting we are just so early in this AI Supercycle and the pace of adoption is insane, and clearly productivity will skyrocket. Why? Because Agentic AI is 24/7 Smart AI agent working for you or your businesses is a mad compelling, and it is estimated to be 40-100x more Inference Hugnry! Many experts already said it is impossible to project this kind of Inference Demand. AI CapEx is expected to ramp up even more in 2027-2028-2029 and 2030 as Global Agentic AI is going to scale to $3-$5 Trillion TAM by 2030. The nature of Agentic is driving higher CPU/GPU ratio, with CPUs handling 50-90% of Agentic workflows. For example, The current Helios Rack: 18 compute trays per rack with 72 GPUs + 18 CPUs. The beauty of this $META and $AMD long term partnership is, that it is absolutely flexible to adjust racks to higher CPU rato or equal to service different needs. Helios rack can be easily swap to 2 GPUs 2CPUs or even CPUs only trays for dedicated orchestration/head nodes. You see, the beauty of this open rack-scale is flexibility and evolvability. If Agentic AI demand pushes much higher, AMD should be able to adjust variant trays without abandoning Heilos Rack. We can't talk just about massive Agentic AI demand without talking about the Supply side or TSMC. TSMC, AMD's primary foundry for advanced nodes ( Zen 6/Venice on N2/2nm), is addressing AI-driven shortages through massive expansions. TSMC accelerates fab construction with up to 10 facilities targeted for 2026. TSMC is accelerating its domestic manufacturing expansion, with industry sources indicating that as many as ten fabs could be under construction or preparing to begin operations across Taiwan’s major science parks. TSMC Capex: $52-56B in 2026 (up 37% YoY), with $45B already approved for new/upgraded capacities. 70-80% for advanced processes (2nm/A16), 10-20% for packaging (CoWoS quadrupling to 120-140K wafers/month by late 2026). In addition, Taiwanese companies (led by TSMC) commit to at least $250B in direct investments in US-based advanced semiconductor, AI, and energy production/innovation capacity.Taiwan provides $250B in government credit guarantees to facilitate additional investments and build a full US semiconductor ecosystem (including industrial parks). TSMC completed a second land purchase in Arizona (January 2026) for gigafab scaling, with an additional $100B+ (potentially four more modules) to further expand and qualify for tariff exemptions. AMD with secured 12GW from OpenAI and $META and massive Agentic AI will mean higher priority acess to 20-30% more wafers on TSMC advanced nodes, as TSMC has multi-year agreements with AMD for AI chips. Dr. C. C. Wei, CEO of TSMC quote: "I spend a lot of time in the last three or four months talking to my customer and then customers. Customer. I want to make sure that my customers demand are real. I talk to those cloud service providers, all of them. Their answer is. I'm quite satisfied with their answer. Actually they show me the evidence that the AI really help their business. So they grow their business successfully and he or she in their financial return. So I also double check their financial status. They are very rich." Amid shortages, the US buildout ensures AMD can ramp production of Instinct GPUs and EPYC CPUs without the constraints hitting competitors like Intel. By diversifying away from Taiwan (85% of advanced nodes today), the agreement mitigates supply disruptions, ensuring stable flows for AMD's chips. Scaling production and securing supply will matter for AMD the most in the next 5-10 years growth. The growth could be 80-100% YoY or higher; or it could be in the 60%. The aggressive TSMC supply ramp is reassuring the higher growth point. Conclusion: AMD stands at a pivotal inflection point in 2026, where the explosive rise of agentic AI demanding 40-100x more inference compute through its 24/7, multi-step orchestration positions the company to potentially triple its EPYC CPU revenue to $45-60B+ by 2028 while scaling Instinct GPUs to tens of billions annually by 2027. Agentic AI demand could push AI CapEx closer to $1 Trillion in 2027, far higher than most estimates. Dr. Lisa Su, AMD's visionary CEO, is masterfully securing supply to harness this massive demand by prioritizing operational execution and deep TSMC collaboration, ensuring readiness for the second-half 2026 AI ramp. Dr. Su has explicitly called out surging EPYC demand for agentic tasks where CPUs power head nodes and traditional workloads alongside GPUs while guiding for data center dominance through proactive capacity planning and partnerships like Nutanix ($150M investment for open agentic platforms) or providing tens of millions CPUs for OpenAI, $META, $ORCL, $AMZN, $MSFT, $GOOGL and others. Her strategy includes multi-year TSMC agreements for advanced nodes (N2 for Venice CPUs and future Instincts), diversifying beyond Taiwan to mitigate risks, and unveiling innovations like the MI455X GPU at CES 2026, which she touted as enabling "the next trillion-dollar market opportunity" in physical AI. Dr. Su's forward-looking vision predicting AI reaching 5 billion users emphasizes "AI everywhere," backed by hardware like Ryzen AI chips, all while declaring demand "going through the roof" and committing to scale without bottlenecks. TSMC's aggressive ramp-up, fueled by $52-56B in 2026 capex (up 37% YoY) and 10+ new fabs across Taiwan, the US (Arizona cluster expanding to 6+ modules with $165B+ investment), Japan, and Europe, provides profound reassurance for AMD's supply stability. The January 2026 US-Taiwan agreement committing $250B in investments and credit guarantees for US reshoring accelerates this, granting tariff relief (15% rates with 1.5-2.5x exemptions) tied to capacity buildouts, enabling TSMC to potentially double output over the decade to meet AI wafer hunger. This translates to 20-30% higher wafer allocations on key nodes, sidestepping Intel-like shortages and empowering Dr. Su's team to deliver on hyperscaler demands without disruption. Ultimately, this synergy cements AMD's leadership in the agentic era, promising sustained growth, $5T+ valuations at scale, and a resilient path forward as AI reshapes the world. This is NOT Financial Advice! Video source: AMD CES 2026

Mike

44,460 次观看 • 6 个月前

New episode on the optimal exercise intensity, duration, and frequency to prevent and reverse heart aging! In this podcast episode, Dr. Benjamin Levine discusses his groundbreaking research, which reveals how three weeks of bed rest can have a more detrimental impact on fitness than 30 years of aging. Dr. Levine details his research findings that show how a structured exercise regimen can reverse up to 20 years of heart aging by improving both shrinkage and compliance, as well as enhancing aspects of vascular age by 15 years. He also discusses how resistance training and aerobic training have profound differences on the heart, what risks are linked to high-intensity exercise, why recovery is key for the heart, how exercise duration and intensity affect coronary calcium levels, what exercise dose increases Afib risk, and so much more. This episode is a must! Available on YouTube, Spotify, X, and everywhere else. Links in comment. Timestamps: 0:00 - Introduction 1:31 - Bed rest vs. 30 years of aging 5:18 - Recovering from bed rest 6:49 - Does exercise protect against long COVID? 11:27 - Bed rest as a model for space flight 12:24 - How bed rest affects heart size 13:52 - Why a brand-new rubber band mimics a lifetime of endurance training 17:23 - The exercise dose that preserves youthful cardiovascular structure 19:32 - Reversing 20 years of heart aging 23:14 - Reversing vascular age by 15 years 28:38 - Why start an exercise regimen in your 70s? 34:26 - High-intensity exercise risks 37:51 - Balancing high- & moderate-intensity training 42:49 - Training for health vs. training for performance 43:57 - Why muscle mass & cardiorespiratory fitness are like retirement funds 45:12 - Make exercise part of your personal hygiene 46:16 - Why VO2 max correlates with longevity 53:43 - Cardiorespiratory fitness & mortality 59:21 - How does change in fitness over time affect mortality? 1:01:34 - Exercise non-responders 1:05:23 - Limiting factors for VO2 max improvements 1:08:20 - How marathon training affects heart size 1:12:34 - Heart adaptations in purely strength-trained vs. endurance athletes 1:18:23 - Why pure strength-trainers should incorporate endurance training 1:22:07 - How strength training affects blood pressure 1:26:41 - How exercise influences cardiac output 1:28:39 - Does CrossFit count as endurance training? 1:31:04 - Exercise for improving blood pressure 1:36:11 - Lifestyle strategies for treating hypertension 1:38:40 - Why recovery is key 1:42:36 - The best indicator of being overtrained 1:43:36 - Estimating training zones 2-5 1:50:00 - Why HRV is a poor recovery indicator 1:55:16 - Why men are faster runners than women 1:58:49 - Can women achieve similar aerobic exercise benefits doing 2x less? 2:00:21 - Possible cardiovascular benefits of HRT in women 2:02:12 - Defining “extreme exercise” 2:04:00 - How exercise volume affects coronary plaque calcification 2:10:50 - How exercise duration & intensity affect coronary calcium levels 2:14:03 - Why high exercise duration & intensity increases Afib risk 2:16:33 - What exercise dose increases Afib risk? 2:17:59 - Managing stroke risk in athletes prone to Afib 2:21:14 - Why you shouldn’t become an endurance athlete to “live longer”
2:24:58

New episode on the optimal exercise intensity, duration, and frequency to prevent and reverse heart aging! In this podcast episode, Dr. Benjamin Levine discusses his groundbreaking research, which reveals how three weeks of bed rest can have a more detrimental impact on fitness than 30 years of aging. Dr. Levine details his research findings that show how a structured exercise regimen can reverse up to 20 years of heart aging by improving both shrinkage and compliance, as well as enhancing aspects of vascular age by 15 years. He also discusses how resistance training and aerobic training have profound differences on the heart, what risks are linked to high-intensity exercise, why recovery is key for the heart, how exercise duration and intensity affect coronary calcium levels, what exercise dose increases Afib risk, and so much more. This episode is a must! Available on YouTube, Spotify, X, and everywhere else. Links in comment. Timestamps: 0:00 - Introduction 1:31 - Bed rest vs. 30 years of aging 5:18 - Recovering from bed rest 6:49 - Does exercise protect against long COVID? 11:27 - Bed rest as a model for space flight 12:24 - How bed rest affects heart size 13:52 - Why a brand-new rubber band mimics a lifetime of endurance training 17:23 - The exercise dose that preserves youthful cardiovascular structure 19:32 - Reversing 20 years of heart aging 23:14 - Reversing vascular age by 15 years 28:38 - Why start an exercise regimen in your 70s? 34:26 - High-intensity exercise risks 37:51 - Balancing high- & moderate-intensity training 42:49 - Training for health vs. training for performance 43:57 - Why muscle mass & cardiorespiratory fitness are like retirement funds 45:12 - Make exercise part of your personal hygiene 46:16 - Why VO2 max correlates with longevity 53:43 - Cardiorespiratory fitness & mortality 59:21 - How does change in fitness over time affect mortality? 1:01:34 - Exercise non-responders 1:05:23 - Limiting factors for VO2 max improvements 1:08:20 - How marathon training affects heart size 1:12:34 - Heart adaptations in purely strength-trained vs. endurance athletes 1:18:23 - Why pure strength-trainers should incorporate endurance training 1:22:07 - How strength training affects blood pressure 1:26:41 - How exercise influences cardiac output 1:28:39 - Does CrossFit count as endurance training? 1:31:04 - Exercise for improving blood pressure 1:36:11 - Lifestyle strategies for treating hypertension 1:38:40 - Why recovery is key 1:42:36 - The best indicator of being overtrained 1:43:36 - Estimating training zones 2-5 1:50:00 - Why HRV is a poor recovery indicator 1:55:16 - Why men are faster runners than women 1:58:49 - Can women achieve similar aerobic exercise benefits doing 2x less? 2:00:21 - Possible cardiovascular benefits of HRT in women 2:02:12 - Defining “extreme exercise” 2:04:00 - How exercise volume affects coronary plaque calcification 2:10:50 - How exercise duration & intensity affect coronary calcium levels 2:14:03 - Why high exercise duration & intensity increases Afib risk 2:16:33 - What exercise dose increases Afib risk? 2:17:59 - Managing stroke risk in athletes prone to Afib 2:21:14 - Why you shouldn’t become an endurance athlete to “live longer”

Dr. Rhonda Patrick

266,327 次观看 • 2 年前

$AMD $5 Trillion MC Is Inevitable Long Term👑 This thread will focus more on Inference! 2026 EPYC "Venice" $TSM 2nm to save Large GW Scale Inference by 40% more than Prior Turin gen. Context: EPYC Turin achieves ~$0.001 per million tokens for batch inference vs $0.02-$0.12/ million tokens as I wrote the thread below. Venice is going to lower cost down to $0.0005-$0.0006/Million Tokens. OpenAI spent roughly $20B on Inference and Training, where 80-90% of that was for Inference per Analysts. AKA Renting Compute is Expensive AF! In this thread, I want to focus on why most analysts and investors are underestimating the role EPYC "Venice" and future Gen on overall Data center revenue. And $TSM ramping up 2nm supply early is a confirmation that AMD will be a major buyer long term. I will also link the thread the Gap between AMD Analysts & Reality and 2nm Ramp Thread so you have more comprehensive view of what I'm writing here. Before I go into detail this is my 2026 Projection: AI GPUs: $35-$50B EPYC Data Center: $15B-$17B Client Segment: $12-$13B Gaming: $6B Embedded: $4B-$5B Total Revenue $70-$100B Non-GAAP net income $18B-$25B Non-GAAP EPS $10.97-$15.40 Foward P/E 55x-70x= $603-$1,078 AMD's Analysts are projecting $0 Revenue for MI450 and sluggish EPYC Growth. Meaning, all analysts are either full of 💩 or Sexist, you decide! Analysts are also projecting 0% growth on AMD "Secret Weapon" Chip as $MSFT said we are at significant Windows refresh and upgrade cycle. Do you think TSMC would allocate more 2nm supply to $AMD at $0 MI450 revenue and sluggish EPYC? 1. EPYC is going to be the leader in lowest Inference! Current Turin cost saving is 95% vs $NVDA or 98-99% on Inference cost when you factor in renting Inference compute from Amazon Web Services, Microsoft Azure, or $NVDA Neocloud pets. TSMC claimed: 10-15% higher performance at iso-power, 25-30% lower power at iso-speed, and ~15% higher transistor density compared to 3nm. This reduces operational expenses (energy, cooling) while increasing throughput per chip. EPYC Turin achieves ~$0.001 per million tokens for batch inference (via vLLM on models like Llama 3 70B), driven by high core counts and low hardware costs. EPYC Venice offers ~1.7x overall performance and up to 70% more compute capability per core, with up to 256 cores (512 threads). Enhanced vector/AI instructions and open-source firmware (openSIL) optimize for inference workloads. AMD Incorporates AI Engines (now part of AMD's XDNA) for on-chip acceleration, improving efficiency for low-latency and edge inference. This reduces reliance on discrete GPUs, lowering system complexity and TCO. Venice SKUs are projected at $3,000-$15,000 ($5,000 for 256-core flagship), far below NVIDIA Rubin ($50,000-$90,000) or AMD's own MI450 GPUs ($40,000-$50,000). High memory bandwidth (up to 1.6 TB/s) supports efficient batch inference. Venice is designed exactly for Large customers that want to lower Inference Cost and MI450 Helios is for Customers that want Training at lowest TCO, TDP as well as lower Upfront 1GW scale(Full build $35-$40B vs $NVDA $55B-$80B). 2. Real World Example: OpenAI's 2025 inference spend reached ~$20B, escalating to even higher total compute rental (mostly inference) amid token volume growth(from video generating). By 2026, with usage doubling (consistent with industry trends: token demand grows 2-5x YoY), assume OpenAI processes ~1,800 billion million-tokens annually $NVDA Blackwell at $0.02-$0.12 is $36B(most optimized) Rubin is projected to be at $0.01/million tokens or $18B annual Inference Cost vs $AMD Venice $0.0005/million tokens or $0.9B annual Inference Cost => Massive saving for OpenAI or anyone that are paying 80-90% Annual Bill for Inference compute. In short, it is unsustainable to pay this much rent vs owning for all current AI players for the medium to long term. Rubin excels in low-latency decode (if Groq integration from $20B deal in 2027-2028), but Venice dominates batch (80% of inference by 2030). Actual savings depend on deployment scale (OpenAI's 6GW AMD plans), electricity rates, and software maturity. If Rubin only hits $0.03, savings swell to $53.1B vs. $17.1B. 3. Will running Inference on Venice and future Gen slow down response generation in 2026 and beyond? Human perception of "fast enough" for chat, agents, search augmentation, summarization, coding assistance is roughly Meaning, EPYC may generate $100B a year on data center revenue, Hence $MSFT $AMZN $META $GOOGL OpenAI xAI and 42+ Countries are leaning AMD for Inference, because the cost saving is MASSIVE! 4. Regular users (you, me, people using ChatGPT, Claude, Gemini, Grok, Perplexity...) are extremely unlikely to notice any slowdown and in many cases might even experience slightly faster or more consistent response times if the industry heavily shifts toward AMD EPYC for inference. What actually happens when companies save massively on inference? When OpenAI , Anthropic , Gemini , Grok Meta .... save billions on the batch/enterprise/RAG layer using EPYC Venice, they typically do one or more of these things with the savings, none of which make your chat slower but enhancing their bottom line(Profit) ~Keep prices the same → make more profit ~Lower subscription prices / increase free tier limits ~Train bigger & better models more frequently ~Offer longer context windows ~Add more reasoning steps / tool calls / agents per query ~Improve multimodal capabilities ~Build more data centers / reduce throttling during peaks In practice the consumer experience usually gets better, not worse, when inference becomes dramatically cheaper. Prime example is $META leaning AMD heavily or currently AMD largest customer. or Grok 2 to Grok 3 heavily used AMD for Inference saving. And most Grok Users reported Groke responses snappier, not slower. 5. What does this mean for potential Revenue? Noted that TSMC is massively ramping 2nm supply for $AMD both MI450 and EPYC. EPYC Conservative projection: FY2025: $10.5B(best Est) FY2026: $16B FY2027: $29B FY2028: $49B FY2029: $75B FY2030: $100B Large customers: $META OpenAI $MSFT $AMZN $GOOGL xAI (Apple?) Smaller customer: $DELL $HPE $SMCI and 42+ other countries. The roadmap to $5 Trillion is very much inevitable as Inference Cost from Renting or owning $NVDA are too high, but $NVDA will still dominate Training market share, where MI families are likely to take 15-20% market share, but the TAM is also expanding Rapidly. Most Institutions are projecting $2-$3Trillion TAM by 2030. $NVDA said $4 Trillion. Dr. Lisa Su said $1 Trillion+ by 2030. So you decide on how much TAM. If you enjoy this kind of analysis, Slap the Like/Repost and Bookmark to please the X Algo as it is Free.99! If you want to support my work further, consider subscribe to see more in-depth analysis! Alright, that is it. Not Financial Advice!
3:09

$AMD $5 Trillion MC Is Inevitable Long Term👑 This thread will focus more on Inference! 2026 EPYC "Venice" $TSM 2nm to save Large GW Scale Inference by 40% more than Prior Turin gen. Context: EPYC Turin achieves ~$0.001 per million tokens for batch inference vs $0.02-$0.12/ million tokens as I wrote the thread below. Venice is going to lower cost down to $0.0005-$0.0006/Million Tokens. OpenAI spent roughly $20B on Inference and Training, where 80-90% of that was for Inference per Analysts. AKA Renting Compute is Expensive AF! In this thread, I want to focus on why most analysts and investors are underestimating the role EPYC "Venice" and future Gen on overall Data center revenue. And $TSM ramping up 2nm supply early is a confirmation that AMD will be a major buyer long term. I will also link the thread the Gap between AMD Analysts & Reality and 2nm Ramp Thread so you have more comprehensive view of what I'm writing here. Before I go into detail this is my 2026 Projection: AI GPUs: $35-$50B EPYC Data Center: $15B-$17B Client Segment: $12-$13B Gaming: $6B Embedded: $4B-$5B Total Revenue $70-$100B Non-GAAP net income $18B-$25B Non-GAAP EPS $10.97-$15.40 Foward P/E 55x-70x= $603-$1,078 AMD's Analysts are projecting $0 Revenue for MI450 and sluggish EPYC Growth. Meaning, all analysts are either full of 💩 or Sexist, you decide! Analysts are also projecting 0% growth on AMD "Secret Weapon" Chip as $MSFT said we are at significant Windows refresh and upgrade cycle. Do you think TSMC would allocate more 2nm supply to $AMD at $0 MI450 revenue and sluggish EPYC? 1. EPYC is going to be the leader in lowest Inference! Current Turin cost saving is 95% vs $NVDA or 98-99% on Inference cost when you factor in renting Inference compute from Amazon Web Services, Microsoft Azure, or $NVDA Neocloud pets. TSMC claimed: 10-15% higher performance at iso-power, 25-30% lower power at iso-speed, and ~15% higher transistor density compared to 3nm. This reduces operational expenses (energy, cooling) while increasing throughput per chip. EPYC Turin achieves ~$0.001 per million tokens for batch inference (via vLLM on models like Llama 3 70B), driven by high core counts and low hardware costs. EPYC Venice offers ~1.7x overall performance and up to 70% more compute capability per core, with up to 256 cores (512 threads). Enhanced vector/AI instructions and open-source firmware (openSIL) optimize for inference workloads. AMD Incorporates AI Engines (now part of AMD's XDNA) for on-chip acceleration, improving efficiency for low-latency and edge inference. This reduces reliance on discrete GPUs, lowering system complexity and TCO. Venice SKUs are projected at $3,000-$15,000 ($5,000 for 256-core flagship), far below NVIDIA Rubin ($50,000-$90,000) or AMD's own MI450 GPUs ($40,000-$50,000). High memory bandwidth (up to 1.6 TB/s) supports efficient batch inference. Venice is designed exactly for Large customers that want to lower Inference Cost and MI450 Helios is for Customers that want Training at lowest TCO, TDP as well as lower Upfront 1GW scale(Full build $35-$40B vs $NVDA $55B-$80B). 2. Real World Example: OpenAI's 2025 inference spend reached ~$20B, escalating to even higher total compute rental (mostly inference) amid token volume growth(from video generating). By 2026, with usage doubling (consistent with industry trends: token demand grows 2-5x YoY), assume OpenAI processes ~1,800 billion million-tokens annually $NVDA Blackwell at $0.02-$0.12 is $36B(most optimized) Rubin is projected to be at $0.01/million tokens or $18B annual Inference Cost vs $AMD Venice $0.0005/million tokens or $0.9B annual Inference Cost => Massive saving for OpenAI or anyone that are paying 80-90% Annual Bill for Inference compute. In short, it is unsustainable to pay this much rent vs owning for all current AI players for the medium to long term. Rubin excels in low-latency decode (if Groq integration from $20B deal in 2027-2028), but Venice dominates batch (80% of inference by 2030). Actual savings depend on deployment scale (OpenAI's 6GW AMD plans), electricity rates, and software maturity. If Rubin only hits $0.03, savings swell to $53.1B vs. $17.1B. 3. Will running Inference on Venice and future Gen slow down response generation in 2026 and beyond? Human perception of "fast enough" for chat, agents, search augmentation, summarization, coding assistance is roughly Meaning, EPYC may generate $100B a year on data center revenue, Hence $MSFT $AMZN $META $GOOGL OpenAI xAI and 42+ Countries are leaning AMD for Inference, because the cost saving is MASSIVE! 4. Regular users (you, me, people using ChatGPT, Claude, Gemini, Grok, Perplexity...) are extremely unlikely to notice any slowdown and in many cases might even experience slightly faster or more consistent response times if the industry heavily shifts toward AMD EPYC for inference. What actually happens when companies save massively on inference? When OpenAI , Anthropic , Gemini , Grok Meta .... save billions on the batch/enterprise/RAG layer using EPYC Venice, they typically do one or more of these things with the savings, none of which make your chat slower but enhancing their bottom line(Profit) ~Keep prices the same → make more profit ~Lower subscription prices / increase free tier limits ~Train bigger & better models more frequently ~Offer longer context windows ~Add more reasoning steps / tool calls / agents per query ~Improve multimodal capabilities ~Build more data centers / reduce throttling during peaks In practice the consumer experience usually gets better, not worse, when inference becomes dramatically cheaper. Prime example is $META leaning AMD heavily or currently AMD largest customer. or Grok 2 to Grok 3 heavily used AMD for Inference saving. And most Grok Users reported Groke responses snappier, not slower. 5. What does this mean for potential Revenue? Noted that TSMC is massively ramping 2nm supply for $AMD both MI450 and EPYC. EPYC Conservative projection: FY2025: $10.5B(best Est) FY2026: $16B FY2027: $29B FY2028: $49B FY2029: $75B FY2030: $100B Large customers: $META OpenAI $MSFT $AMZN $GOOGL xAI (Apple?) Smaller customer: $DELL $HPE $SMCI and 42+ other countries. The roadmap to $5 Trillion is very much inevitable as Inference Cost from Renting or owning $NVDA are too high, but $NVDA will still dominate Training market share, where MI families are likely to take 15-20% market share, but the TAM is also expanding Rapidly. Most Institutions are projecting $2-$3Trillion TAM by 2030. $NVDA said $4 Trillion. Dr. Lisa Su said $1 Trillion+ by 2030. So you decide on how much TAM. If you enjoy this kind of analysis, Slap the Like/Repost and Bookmark to please the X Algo as it is Free.99! If you want to support my work further, consider subscribe to see more in-depth analysis! Alright, that is it. Not Financial Advice!

Mike

102,223 次观看 • 8 个月前

Just in $AMD Anush "Speed is the moat"|ROCm🎙️ In the race to define the future of AI, what's the one advantage that truly lasts? It's not proprietary tech, argues Anush Elangovan Elangovan, VP of AI Software at AMD , but the sustainable speed of innovation. He explains why AMD is rejecting the "walled garden" model for its open source ROCm stack, betting that an open community flywheel is the key to victory. Listen to understand how this open strategy is designed to out-innovate closed systems by empowering developers to solve everything from frontier-model challenges to the mundane, everyday problems that define the "last mile" of AI. AMD ROCm Software: Part 1 Transcript [00:00:00] Andrew Zigler: Joining me is Anush Elangovan, VP of AI software at AMD. And when people talk about AI compute, the conversation often stops at hardware specs, but it's more than just physical chips that win the game. It's also the software ecosystems supporting them. [00:00:18] Andrew Zigler: The prevailing strategy in the industry has been to build something like a walled garden. You know, something closed, proprietary locks, developers in. But AMD is betting on an entirely different play, open source acceleration, and with rock, their open source AI software stack. AMD is building not just hardware parity, but an innovation flywheel that's powered by the community with interoperability and the freedom to scale without all of that pesky lockin. [00:00:48] Andrew Zigler: And in this world, speed is your moat and how fast you can innovate while your platform remains open, flexible, and standardize across all of its applications. That's what we're gonna explore [00:01:00] today. So Anush, I'm really excited to have you here. Welcome to Dev Interrupted. [00:01:04] Anush Elangovan: Thanks for having me. Uh, super excited to chat about it. [00:01:07] Andrew Zigler: Amazing. Well, let's go ahead and dive right in with kind of what I laid it out with in the beginning, the idea of the moat and it being about speed. I wanna unpack that a bit because that came from you when you and I first spoke. And I, and I want to know, you know, how do you define speed inside of AMD beyond just things like hardware, benchmarks. [00:01:27] Anush Elangovan: Yeah, that's a very good question. So when we typically talk about speed, everyone's like, Hey, hardware benchmark specs, right? Like, uh, memory bandwidth or, or flops. And that is one important part of it, uh, AMD does very well. With that, we do have, a, a very good history of executing on that axis. [00:01:47] Anush Elangovan: But when I say speed is the moat, it is about, uh, how we prepare, how we build the muscle to run the race for a long time and run it fast. And it is [00:02:00] not about a single point in time that you've, you've beat some you know, benchmark and, and you declare victory. It's about building the ability to consistently develop and deliver. [00:02:13] Anush Elangovan: Both hardware and software innovation at scale and do it fast, right? Like, you know, we we're increasingly getting to a point where models come out and they're, uh, you know, a year or two ago it was like, Hey, they work on AMD on day zero, which is great, but now they are performing on AMD the day it releases, right? [00:02:32] Anush Elangovan: So, what does it take to Prefetch where the industry is going? Be prepared to intercept. At that point is what you know, I, I refer to as you know, the, the speed factor in, in creating this mode, right? And the mode is just shed all things that hold you back and run as fast as you can. [00:02:53] Anush Elangovan: Uh, because the pace of innovation that is, uh, being seen in, in AI [00:03:00] industries is just. Amazing. Right? And it's like, it's transformational at at how you generate electricity. It's transformational as at how you build data centers. It's transformational at how you deploy compute, networking. It's transformational at what kind of use cases you, you know, uh, use AI for. [00:03:17] Anush Elangovan: Uh, and for that, you need to be prepared to, see what comes tomorrow and be prepared to run the race tomorrow. [00:03:23] Andrew Zigler: Yeah, it's a really great perspective because it highlights that it's not just like a checkpoint that you run through. I like how you called out, like it's not just hitting that benchmark or being the best in class at that moment, in that snapshot, it's about having a. The throughput and about having that dedication to the idea and continuing to deliver on it. [00:03:43] Andrew Zigler: It's not just crossing the threshold, but it's also being the engine. And that's what, that's what protects a business. That is the moat, because the moat is that innovation layer, the faster and more, uh, future forward. That you can work and think, [00:04:00] you know, the better. Uh, we, we talk a lot about like future forward work styles. [00:04:04] Andrew Zigler: Like what are the things I could be doing right now today that are gonna be like, way more useful tomorrow? Let, let's abandon those, workflows that are older and that kind of like, that translates into. An advantage when you work that way. You know, what kind of things have you learned working with, uh, like across all spectrums of people who would use ROCm, right? [00:04:23] Andrew Zigler: You have like the developers, but then you also have the enterprises and you have this large span of adoptees, right? So what is the, what does that look like that you learn? [00:04:32] Anush Elangovan: Yeah, so, so the way I look at it is there are gonna be pockets of different, uh, you know, cadences, right? Like, so people who are deploying in enterprises, for example, right? The validation and how long it takes for them to deploy an LLM that's secure. It's, with guardrails, et cetera, maybe longer. [00:04:52] Anush Elangovan: but you still have to go through the process and you have to be prepared to like, walk that walk to deploy an enterprises. That doesn't mean it's [00:05:00] not fast, that's as fast as you can do for that industry, right? And if you are deploying AI in healthcare, right, it's, it's got its own, uh, cycle. [00:05:07] Anush Elangovan: but in each one of these, you want to see how, like, go down to the essence of what is it that you actually have to do. And, you know, I, I, I like how you framed it. It's like it's, you shed your prior assumptions of how things are done, right. And, and you kind of build up from a, uh, first principles, uh, approach to say, this is how I could use AI to unlock, whatever I'm doing. [00:05:33] Anush Elangovan: And, and, some of it, you know, it's good to really step back and look at. Just question every part of it, right? Like right now you're getting chat GPT and, Gemini competing for like, math, olympiads and, and, uh, college, uh, reasoning, uh, tests. Right? And, and those are like that, that is amazing and increasingly like complex tasks that they're trying to do. [00:05:58] Anush Elangovan: But there may also be like. [00:06:00] More mundane things that AI could, could get applied to. Right? And, and so when we think about shedding old ways, you wanna shed it not just in like the tip of the spear. It's like, you know, I'm gonna see what's the frontier model. It's also, it could be something as simple as. [00:06:18] Anush Elangovan: How do you choose a, a movie, uh, you know, like a recommendation system, right? Or, or, uh, an automated, uh, flight, uh, rebooking system. So the moment, you know, your flight is late, uh, right now it's a notification, right? It's like, oh, you got a text message saying your flight's late. And I got that like three times this week. [00:06:38] Anush Elangovan: But anyway, uh, and, and, and, and, I was just like, okay, so if I were to rethink this. All this MCPs that we have that should be hooked up into an MCP that says, your flight's delayed. Here are your options. If you want, you know, these are the paid options. Yeah. Here are the free options. This will get you back into your you know, Toronto airport [00:07:00] tonight. [00:07:00] Anush Elangovan: Or if you stay, here's a hotel plus this, plus this, plus. It's just like, go ahead is all I should say. Versus now I'm like, okay, can someone, you know, can I call a travel agent? Can I do this? Can I go online and log into And you know, so we gotta fundamentally rethink even those like small, nuances of, things that we do that can be automated out and AI is really, really good at doing something like this, right? Maybe I just explained an AI startup idea right now. Somebody should just start that. [00:07:29] Andrew Zigler: I think you did. Yeah, you definitely did. Someone, one of our listeners is definitely going to lift that off of you. I, I, I, you know, I hate being on the receiving end of those. You feel a little helpless and then you have to like, follow the whole flow. So I know what you mean. Like I, I like how you called out that the build and this like. [00:07:45] Andrew Zigler: Where speed is your moat and the innovation layer is protecting you, is what makes you better than your competitors. How you scale that and you bring that to market. So by understanding the problems that you're solving, uh, throwing away those older assumptions, but also [00:08:00] recognizing that like. We're building every single day, new things and new ways of using stuff that we're still figuring out the implications of. [00:08:08] Andrew Zigler: And so when you have a lot of velocity and you're introducing a lot of new ideas, and maybe you have that workflow now that automatically rebook your flight off of your late flight text message, and uh, I know I would certainly use it, but you know, what kind of philosophies guide the way that y'all think about building this ecosystem to manage that stability while letting folks. [00:08:29] Andrew Zigler: Play with the speed and the assumptions and the airplane re bookings. [00:08:34] Anush Elangovan: so, so I think, you know, we need to peel one layer down, right? and the philosophy is, Hey, we, we just discovered electricity, right? And you know what we're gonna do? We are gonna make motors, uh, or dynamos, right? Like engines. Uh, sure. We don't know if it's gonna be a Ferrari that you're gonna make, or it's a a a a dump truck. [00:08:57] Anush Elangovan: That's good for doing this. But let's [00:09:00] let, which is also required, right? You need a dump truck. You need a garbage truck. And, [00:09:04] Andrew Zigler: Yeah. You need the [00:09:04] Anush Elangovan: course you need, uh, a Ferrari for a midlife crisis, right? So, [00:09:09] Andrew Zigler: precisely. [00:09:10] Anush Elangovan: But, but my, uh, point is what do we build next? And, uh, and this is what I meant by like, okay, let's, let's take those baby steps to build the. [00:09:20] Anush Elangovan: Infrastructure that's required that we know we'll have to use, right? So, so if I just discovered electricity, okay, great. Now one, how do I save this electricity and how do I use it? So there's battery technology, so you need to do something like that, right? Like so. But then you also want to make it into an actionable thing. [00:09:37] Anush Elangovan: You want to make it for like automobiles, or you wanna use it for, you know, powering, uh, entire cities. So it is that transformational. So, uh, AI is that transformational. So, if you distill down, it'll, it'll come down to how do we think about, what we can do with this this fundamental technology that, We may not be aware of what it [00:10:00] is gonna unlock next, but at least you know the next step is clear, right? It's like a dense fog, you know, it's gonna be like, it, it's the right path. You see the light, but it's kind of like out there and, and the steps you're taking are concrete and you're like, okay, this is good. [00:10:16] Anush Elangovan: I, this is better than where I was or where we were. So we are moving forward. So you can build with the. Intuition from what you see in the short term and a tactical view, but towards what you think the future is gonna be. [00:10:28] Andrew Zigler: Right. You almost like we're all in this like fog of war, right? And like you said, you're reaching out and you're trying to step through it. You could think of it too, as like you're in the dark and your hands are up in front of you and you know that. You're, you're not gonna run your face into a wall because your hands are out in front of you, but you're not gonna maybe do much better than that. [00:10:45] Andrew Zigler: So that's kind of like, I think the eco, the, the industry, the world that we find ourselves in, uh, and we all have to, then this becomes the power of an ecosystem, of a group of people working together to create that layer of, [00:11:00] uh, of establishing the [00:11:01] Anush Elangovan: exactly. And I, I, I just, instead of, you know, saying fog of war I describe it as like, you're in this. Beautiful valley with like a morning, uh, fog that's in. You can smell the flowers. You, you hear the birds. You are like, okay, it's, we are in like, uh, utopian paradise and yes, I just need to like, continue the walk, right? [00:11:24] Anush Elangovan: and then move forward with that, conviction that you're in the right spot. [00:11:27] Andrew Zigler: Yeah. So let's talk about that ecosystem world. This nice, I love how you describe it, this grassy side of a hill in the morning that's covered in some mist and maybe we can't see 30 feet in one direction, but it sure is a beautiful hill and it smells nice. And so we're all here. And why is, in that world, why is. [00:11:44] Andrew Zigler: You know, open source, their strategic advantage that y'all are going for in the AI hardware market. And, and then how does like ROCm turn that into wins for people within that ecosystem? [00:11:56] Anush Elangovan: you know, the, the way we look at it is this, is kind of like how I view [00:12:00] AI and the ecosystem, right? But, but it is for everyone to enjoy. Uh, and so we do want to make sure that. You know, it is, uh, beneficial for everyone. [00:12:09] Anush Elangovan: The ecosystem can come in and, and innovate. It's an open innovation engine. and uh, it is very different from, you know, having a walled garden with, Hey, only I know how to do this and I'm gonna do it and throw it over the fence and you can use it or keep walking, right? So we'd like to be good citizens that way, but also. [00:12:30] Anush Elangovan: Uh, it is self-fulfilling in a way, right? Like it, the, the pace at which we innovate with open source is unmatched. Like, you know, our serving engines are like VLLM and, and sg l. Those things, uh, those frameworks are like super, super aggressive in terms of how fast they come out with features and how fast they can you know, get performant models out. [00:12:52] Anush Elangovan: And that compared with what, uh, you'd get from, you know, the likes of like T-R-T-L-L-M or something is always lagging, right? Because you [00:13:00] just can't keep up with you know, 200 commits a week just on one particular model to get that model really performant [00:13:06] Andrew Zigler: And, and, and in that world where, you know, everyone can enjoy the winds of this, what kind of customer stories or innovation stories have really stood out to you and excite you about building and creating this place for developers? [00:13:19] Anush Elangovan: Yeah. So I think the parts that are super exciting for me are when when we get to see a customer that is first skeptical. Then they start a little like, okay, fine, we'll give you a chance. Uh, we do a simple, uh, POC and then they're like, huh, this seems to work. Yeah, we told you it works. [00:13:42] Anush Elangovan: You don't have to change one line of code. Really? Yes, no need to change one line of code. Okay, let's try a production workload. So then they try it. Oh, you're more performant than the competition. Yes. We're more performant than, than the competition. So how much does it cost? And we're like, oh, it's your TCO is better with, uh, [00:14:00] AMD. [00:14:00] Anush Elangovan: So again, they're like, wow, okay, good. So now how do we deploy at scale? And then we go deploy it at scale. And when they give a thumbs up on that and they say, this is good, right? That's when you know, you, you see it go full circle from like, oh, we, we've never heard about AMD to like actually deploy to tens of thousands of GPUs In the order of a few months, right? It, it, it really is fascinating to see and very exciting and invigorating to [00:14:28] Andrew Zigler: Yeah. At like a great exposure to a lot of interesting problems. And, and then people using the infrastructure, the, the technology available to solve those problems. Really specific problems by the way, that's often why they're bringing their data and AI to it, uh, is because it is really specific and important for them. [00:14:45] Andrew Zigler: And there's a, a lot I think that other engineering orgs can learn and even emulate from AMD's success and, and having this open source ecosystem and it causing this acceleration within. You [00:15:00] know, uh, customers and enterprises that use and adopt the tools and, and, and that creates an advantage. And that goes back to why we're talking and like the real thesis of our conversation today. [00:15:10] Andrew Zigler: So how do you think engineering leaders that are listening to this and obviously tapping into this great success AMD has from an open source flywheel, how do you think other, other folks building in the same space can foster that open, first, that open source oriented culture in order to, you know, accelerate their innovation goals? [00:15:29] Anush Elangovan: Yeah, that's a very good question. So the startup that um, was acquired by AMD we, we built, I mean, we started off doing iot stuff and you know, smart ring and all that, right? But in the, the end of like, uh, and not the end, the last six years of the company was building ML compilers. [00:15:47] Anush Elangovan: And ml, ML compilers are like super, uh, complicated, sophisticated, advanced algorithms, dah, dah, dah. but it was all open source, right? So our VCs were like, wait, what do you mean your core [00:16:00] IP is open source? And um, the speed is the moat applied even then, right? It was just like, yes, if you have an idea that. [00:16:08] Anush Elangovan: Because someone saw this idea that you are, they're gonna be able to catch up, then you probably have the wrong idea anyway. But if they are, you know, you execute and they're gonna catch up, that you should assume they're gonna catch up. Right? So you gotta move forward. So keeping it open source is super important. [00:16:25] Anush Elangovan: But also to your question on like, you know, the learnings from an AMD standpoint, right? If there are, hard problems, I'd say dig in and work through it, right? Like there's no way but through it, right? That should be the simple mentality. And more, uh, frequently than not. you'll see that you'll just make it through in a, in, in good form. [00:16:52] Anush Elangovan: But if you doubt it and you're like, oh, I don't know if I should commit, if I'm, I, you know, what should just commit to do the right thing [00:17:00] every step, right? Every step, and just keep taking one step in front of the other. And in no time you'll see that you'll be running. Right. And, and yes, the first few steps will be like, yeah, everyone's complaining about your software quality. [00:17:15] Anush Elangovan: Everyone's complaining about this and that, and it doesn't work. And, and a few steps in, you know, you get, you get the hang of all the complaints that are coming in. You get the feedback loop. You're like, okay, what, what are you prioritizing again? One step in front of the other, right? You just keep knocking that out and then you get to a point where you're, it just becomes second nature, right? To do the, to do the right thing. And, and then yes, if someone gives you two options, you'll be like, fine. This is, uh, you know, there's always the resource trade off. There's always a human capital trade off, but what's the right thing to do? of course, I, I'm pragmatic about what we choose, but, but if the right thing for your long-term success is dig in, go first, principles, make it [00:18:00] happen. [00:18:00] Anush Elangovan: Well. Then just go for that. There's, there is no shortcut to [00:18:04] Andrew Zigler: acknowledging, you know, how it aligns with your mission, your core company goals, and what you're looking to achieve. And, and I, I love how you rightfully called out that in the open source world and you know, you have your technology that you've built, what you think is your moat upon, right? [00:18:22] Andrew Zigler: It's your code and, and to open source that, or to just make it where anyone could peer in is, you know. Scary in one regard, but two, it just kind of feels like you're handing away your throne room in some kind of sense, a very direct feeling sense. But the ultimately, you were really right to call out, and this is something I think about all the time, that the real power there is still the speed This the speed. [00:18:42] Andrew Zigler: That was the moat at the beginning of our conversation. It's the speed in combination with your. Very specific domain understanding of what you're building and what you're creating, and your new role as the steward of that world and how people plug into it, which [00:19:00] has frankly, a lot more influence and power than lording over a closed. [00:19:04] Andrew Zigler: You know, repository or an ecosystem, and like you said, like throwing things over the wall. Sure. There, there might be people always on the other side of that wall, but you're not gonna have a great connection with them. You're not gonna be able to really clearly understand them. I, I like your metaphor of the side of the field of the mountain a lot more. [00:19:23] Andrew Zigler: But, but in the, in this world, you know, where. That speed is, is the power and, and open source is just one way that you can harness that speed to get really far ahead and to innovate. , There's other parts of this equation that you can be experimenting with too, and I'd love to pick your brain about them as a software leader and, and, and one of them is about looking forward and kind of understanding that future that we're all building towards and beyond today's models and hardware. [00:19:48] Andrew Zigler: You know, what do you see as the next major bottleneck or opportunity in the AI compute space? As, as you know, enterprises and folks start to get a little more mature about what's available to [00:20:00] them. [00:20:00] Anush Elangovan: Yeah, I think, the bottleneck and opportunity is, uh, what I'd call, call walking the last mile of ai. Right. Uh, and like I I, I gave you an example, uh, previously, but, but it's similar to that. It's like there are cases where Humans have so many, uh, things to do in your day. You know, like the, if we sit down and actually had a customer focus like, okay, these customers lives, I'm gonna save four hours of this customer's life. And if you actually sit down and look at all of that, it'll be. Easily automatable, easily you know, uh, applicable, uh, for ai, right? [00:20:39] Anush Elangovan: Like, but then making it happen is gonna take a little bit, right? It's like maybe it's, uh, paying your utility bill, right? Or something like that, right? Or, or, your healthcare explanation of benefits. Uh, like, I'm sure you get an explanation of benefits, and I'm like, I, I don't even know what that thing is. [00:20:55] Anush Elangovan: It's just like EOB and like. [00:20:57] Andrew Zigler: it's a big, a big old PDF. Yeah, [00:21:00] exactly. [00:21:01] Anush Elangovan: Like, like, I'm like great straight to the, uh, shredder, right? And but that could be, you know, automated with the ai, right? It, it, it'd be like, Hey, the summary of this thing is you went and visited this day. Everything is okay. Everything is paid for, so don't worry, it's not a bill. [00:21:17] Anush Elangovan: That again, the same, uh, thing, but the sense of what that information overload is could be. Digested by ai, uh, accumulated over time and retrieved when you need it. Like, I don't, I actually don't even need to know this EOB right now, unless of course, whenever I need to know it, that maybe, you know, like for some benefits I need to figure out what do, what did I do over the past year and how do I apply it? Source:
31:35

Just in $AMD Anush "Speed is the moat"|ROCm🎙️ In the race to define the future of AI, what's the one advantage that truly lasts? It's not proprietary tech, argues Anush Elangovan Elangovan, VP of AI Software at AMD , but the sustainable speed of innovation. He explains why AMD is rejecting the "walled garden" model for its open source ROCm stack, betting that an open community flywheel is the key to victory. Listen to understand how this open strategy is designed to out-innovate closed systems by empowering developers to solve everything from frontier-model challenges to the mundane, everyday problems that define the "last mile" of AI. AMD ROCm Software: Part 1 Transcript [00:00:00] Andrew Zigler: Joining me is Anush Elangovan, VP of AI software at AMD. And when people talk about AI compute, the conversation often stops at hardware specs, but it's more than just physical chips that win the game. It's also the software ecosystems supporting them. [00:00:18] Andrew Zigler: The prevailing strategy in the industry has been to build something like a walled garden. You know, something closed, proprietary locks, developers in. But AMD is betting on an entirely different play, open source acceleration, and with rock, their open source AI software stack. AMD is building not just hardware parity, but an innovation flywheel that's powered by the community with interoperability and the freedom to scale without all of that pesky lockin. [00:00:48] Andrew Zigler: And in this world, speed is your moat and how fast you can innovate while your platform remains open, flexible, and standardize across all of its applications. That's what we're gonna explore [00:01:00] today. So Anush, I'm really excited to have you here. Welcome to Dev Interrupted. [00:01:04] Anush Elangovan: Thanks for having me. Uh, super excited to chat about it. [00:01:07] Andrew Zigler: Amazing. Well, let's go ahead and dive right in with kind of what I laid it out with in the beginning, the idea of the moat and it being about speed. I wanna unpack that a bit because that came from you when you and I first spoke. And I, and I want to know, you know, how do you define speed inside of AMD beyond just things like hardware, benchmarks. [00:01:27] Anush Elangovan: Yeah, that's a very good question. So when we typically talk about speed, everyone's like, Hey, hardware benchmark specs, right? Like, uh, memory bandwidth or, or flops. And that is one important part of it, uh, AMD does very well. With that, we do have, a, a very good history of executing on that axis. [00:01:47] Anush Elangovan: But when I say speed is the moat, it is about, uh, how we prepare, how we build the muscle to run the race for a long time and run it fast. And it is [00:02:00] not about a single point in time that you've, you've beat some you know, benchmark and, and you declare victory. It's about building the ability to consistently develop and deliver. [00:02:13] Anush Elangovan: Both hardware and software innovation at scale and do it fast, right? Like, you know, we we're increasingly getting to a point where models come out and they're, uh, you know, a year or two ago it was like, Hey, they work on AMD on day zero, which is great, but now they are performing on AMD the day it releases, right? [00:02:32] Anush Elangovan: So, what does it take to Prefetch where the industry is going? Be prepared to intercept. At that point is what you know, I, I refer to as you know, the, the speed factor in, in creating this mode, right? And the mode is just shed all things that hold you back and run as fast as you can. [00:02:53] Anush Elangovan: Uh, because the pace of innovation that is, uh, being seen in, in AI [00:03:00] industries is just. Amazing. Right? And it's like, it's transformational at at how you generate electricity. It's transformational as at how you build data centers. It's transformational at how you deploy compute, networking. It's transformational at what kind of use cases you, you know, uh, use AI for. [00:03:17] Anush Elangovan: Uh, and for that, you need to be prepared to, see what comes tomorrow and be prepared to run the race tomorrow. [00:03:23] Andrew Zigler: Yeah, it's a really great perspective because it highlights that it's not just like a checkpoint that you run through. I like how you called out, like it's not just hitting that benchmark or being the best in class at that moment, in that snapshot, it's about having a. The throughput and about having that dedication to the idea and continuing to deliver on it. [00:03:43] Andrew Zigler: It's not just crossing the threshold, but it's also being the engine. And that's what, that's what protects a business. That is the moat, because the moat is that innovation layer, the faster and more, uh, future forward. That you can work and think, [00:04:00] you know, the better. Uh, we, we talk a lot about like future forward work styles. [00:04:04] Andrew Zigler: Like what are the things I could be doing right now today that are gonna be like, way more useful tomorrow? Let, let's abandon those, workflows that are older and that kind of like, that translates into. An advantage when you work that way. You know, what kind of things have you learned working with, uh, like across all spectrums of people who would use ROCm, right? [00:04:23] Andrew Zigler: You have like the developers, but then you also have the enterprises and you have this large span of adoptees, right? So what is the, what does that look like that you learn? [00:04:32] Anush Elangovan: Yeah, so, so the way I look at it is there are gonna be pockets of different, uh, you know, cadences, right? Like, so people who are deploying in enterprises, for example, right? The validation and how long it takes for them to deploy an LLM that's secure. It's, with guardrails, et cetera, maybe longer. [00:04:52] Anush Elangovan: but you still have to go through the process and you have to be prepared to like, walk that walk to deploy an enterprises. That doesn't mean it's [00:05:00] not fast, that's as fast as you can do for that industry, right? And if you are deploying AI in healthcare, right, it's, it's got its own, uh, cycle. [00:05:07] Anush Elangovan: but in each one of these, you want to see how, like, go down to the essence of what is it that you actually have to do. And, you know, I, I, I like how you framed it. It's like it's, you shed your prior assumptions of how things are done, right. And, and you kind of build up from a, uh, first principles, uh, approach to say, this is how I could use AI to unlock, whatever I'm doing. [00:05:33] Anush Elangovan: And, and, some of it, you know, it's good to really step back and look at. Just question every part of it, right? Like right now you're getting chat GPT and, Gemini competing for like, math, olympiads and, and, uh, college, uh, reasoning, uh, tests. Right? And, and those are like that, that is amazing and increasingly like complex tasks that they're trying to do. [00:05:58] Anush Elangovan: But there may also be like. [00:06:00] More mundane things that AI could, could get applied to. Right? And, and so when we think about shedding old ways, you wanna shed it not just in like the tip of the spear. It's like, you know, I'm gonna see what's the frontier model. It's also, it could be something as simple as. [00:06:18] Anush Elangovan: How do you choose a, a movie, uh, you know, like a recommendation system, right? Or, or, uh, an automated, uh, flight, uh, rebooking system. So the moment, you know, your flight is late, uh, right now it's a notification, right? It's like, oh, you got a text message saying your flight's late. And I got that like three times this week. [00:06:38] Anush Elangovan: But anyway, uh, and, and, and, and, I was just like, okay, so if I were to rethink this. All this MCPs that we have that should be hooked up into an MCP that says, your flight's delayed. Here are your options. If you want, you know, these are the paid options. Yeah. Here are the free options. This will get you back into your you know, Toronto airport [00:07:00] tonight. [00:07:00] Anush Elangovan: Or if you stay, here's a hotel plus this, plus this, plus. It's just like, go ahead is all I should say. Versus now I'm like, okay, can someone, you know, can I call a travel agent? Can I do this? Can I go online and log into And you know, so we gotta fundamentally rethink even those like small, nuances of, things that we do that can be automated out and AI is really, really good at doing something like this, right? Maybe I just explained an AI startup idea right now. Somebody should just start that. [00:07:29] Andrew Zigler: I think you did. Yeah, you definitely did. Someone, one of our listeners is definitely going to lift that off of you. I, I, I, you know, I hate being on the receiving end of those. You feel a little helpless and then you have to like, follow the whole flow. So I know what you mean. Like I, I like how you called out that the build and this like. [00:07:45] Andrew Zigler: Where speed is your moat and the innovation layer is protecting you, is what makes you better than your competitors. How you scale that and you bring that to market. So by understanding the problems that you're solving, uh, throwing away those older assumptions, but also [00:08:00] recognizing that like. We're building every single day, new things and new ways of using stuff that we're still figuring out the implications of. [00:08:08] Andrew Zigler: And so when you have a lot of velocity and you're introducing a lot of new ideas, and maybe you have that workflow now that automatically rebook your flight off of your late flight text message, and uh, I know I would certainly use it, but you know, what kind of philosophies guide the way that y'all think about building this ecosystem to manage that stability while letting folks. [00:08:29] Andrew Zigler: Play with the speed and the assumptions and the airplane re bookings. [00:08:34] Anush Elangovan: so, so I think, you know, we need to peel one layer down, right? and the philosophy is, Hey, we, we just discovered electricity, right? And you know what we're gonna do? We are gonna make motors, uh, or dynamos, right? Like engines. Uh, sure. We don't know if it's gonna be a Ferrari that you're gonna make, or it's a a a a dump truck. [00:08:57] Anush Elangovan: That's good for doing this. But let's [00:09:00] let, which is also required, right? You need a dump truck. You need a garbage truck. And, [00:09:04] Andrew Zigler: Yeah. You need the [00:09:04] Anush Elangovan: course you need, uh, a Ferrari for a midlife crisis, right? So, [00:09:09] Andrew Zigler: precisely. [00:09:10] Anush Elangovan: But, but my, uh, point is what do we build next? And, uh, and this is what I meant by like, okay, let's, let's take those baby steps to build the. [00:09:20] Anush Elangovan: Infrastructure that's required that we know we'll have to use, right? So, so if I just discovered electricity, okay, great. Now one, how do I save this electricity and how do I use it? So there's battery technology, so you need to do something like that, right? Like so. But then you also want to make it into an actionable thing. [00:09:37] Anush Elangovan: You want to make it for like automobiles, or you wanna use it for, you know, powering, uh, entire cities. So it is that transformational. So, uh, AI is that transformational. So, if you distill down, it'll, it'll come down to how do we think about, what we can do with this this fundamental technology that, We may not be aware of what it [00:10:00] is gonna unlock next, but at least you know the next step is clear, right? It's like a dense fog, you know, it's gonna be like, it, it's the right path. You see the light, but it's kind of like out there and, and the steps you're taking are concrete and you're like, okay, this is good. [00:10:16] Anush Elangovan: I, this is better than where I was or where we were. So we are moving forward. So you can build with the. Intuition from what you see in the short term and a tactical view, but towards what you think the future is gonna be. [00:10:28] Andrew Zigler: Right. You almost like we're all in this like fog of war, right? And like you said, you're reaching out and you're trying to step through it. You could think of it too, as like you're in the dark and your hands are up in front of you and you know that. You're, you're not gonna run your face into a wall because your hands are out in front of you, but you're not gonna maybe do much better than that. [00:10:45] Andrew Zigler: So that's kind of like, I think the eco, the, the industry, the world that we find ourselves in, uh, and we all have to, then this becomes the power of an ecosystem, of a group of people working together to create that layer of, [00:11:00] uh, of establishing the [00:11:01] Anush Elangovan: exactly. And I, I, I just, instead of, you know, saying fog of war I describe it as like, you're in this. Beautiful valley with like a morning, uh, fog that's in. You can smell the flowers. You, you hear the birds. You are like, okay, it's, we are in like, uh, utopian paradise and yes, I just need to like, continue the walk, right? [00:11:24] Anush Elangovan: and then move forward with that, conviction that you're in the right spot. [00:11:27] Andrew Zigler: Yeah. So let's talk about that ecosystem world. This nice, I love how you describe it, this grassy side of a hill in the morning that's covered in some mist and maybe we can't see 30 feet in one direction, but it sure is a beautiful hill and it smells nice. And so we're all here. And why is, in that world, why is. [00:11:44] Andrew Zigler: You know, open source, their strategic advantage that y'all are going for in the AI hardware market. And, and then how does like ROCm turn that into wins for people within that ecosystem? [00:11:56] Anush Elangovan: you know, the, the way we look at it is this, is kind of like how I view [00:12:00] AI and the ecosystem, right? But, but it is for everyone to enjoy. Uh, and so we do want to make sure that. You know, it is, uh, beneficial for everyone. [00:12:09] Anush Elangovan: The ecosystem can come in and, and innovate. It's an open innovation engine. and uh, it is very different from, you know, having a walled garden with, Hey, only I know how to do this and I'm gonna do it and throw it over the fence and you can use it or keep walking, right? So we'd like to be good citizens that way, but also. [00:12:30] Anush Elangovan: Uh, it is self-fulfilling in a way, right? Like it, the, the pace at which we innovate with open source is unmatched. Like, you know, our serving engines are like VLLM and, and sg l. Those things, uh, those frameworks are like super, super aggressive in terms of how fast they come out with features and how fast they can you know, get performant models out. [00:12:52] Anush Elangovan: And that compared with what, uh, you'd get from, you know, the likes of like T-R-T-L-L-M or something is always lagging, right? Because you [00:13:00] just can't keep up with you know, 200 commits a week just on one particular model to get that model really performant [00:13:06] Andrew Zigler: And, and, and in that world where, you know, everyone can enjoy the winds of this, what kind of customer stories or innovation stories have really stood out to you and excite you about building and creating this place for developers? [00:13:19] Anush Elangovan: Yeah. So I think the parts that are super exciting for me are when when we get to see a customer that is first skeptical. Then they start a little like, okay, fine, we'll give you a chance. Uh, we do a simple, uh, POC and then they're like, huh, this seems to work. Yeah, we told you it works. [00:13:42] Anush Elangovan: You don't have to change one line of code. Really? Yes, no need to change one line of code. Okay, let's try a production workload. So then they try it. Oh, you're more performant than the competition. Yes. We're more performant than, than the competition. So how much does it cost? And we're like, oh, it's your TCO is better with, uh, [00:14:00] AMD. [00:14:00] Anush Elangovan: So again, they're like, wow, okay, good. So now how do we deploy at scale? And then we go deploy it at scale. And when they give a thumbs up on that and they say, this is good, right? That's when you know, you, you see it go full circle from like, oh, we, we've never heard about AMD to like actually deploy to tens of thousands of GPUs In the order of a few months, right? It, it, it really is fascinating to see and very exciting and invigorating to [00:14:28] Andrew Zigler: Yeah. At like a great exposure to a lot of interesting problems. And, and then people using the infrastructure, the, the technology available to solve those problems. Really specific problems by the way, that's often why they're bringing their data and AI to it, uh, is because it is really specific and important for them. [00:14:45] Andrew Zigler: And there's a, a lot I think that other engineering orgs can learn and even emulate from AMD's success and, and having this open source ecosystem and it causing this acceleration within. You [00:15:00] know, uh, customers and enterprises that use and adopt the tools and, and, and that creates an advantage. And that goes back to why we're talking and like the real thesis of our conversation today. [00:15:10] Andrew Zigler: So how do you think engineering leaders that are listening to this and obviously tapping into this great success AMD has from an open source flywheel, how do you think other, other folks building in the same space can foster that open, first, that open source oriented culture in order to, you know, accelerate their innovation goals? [00:15:29] Anush Elangovan: Yeah, that's a very good question. So the startup that um, was acquired by AMD we, we built, I mean, we started off doing iot stuff and you know, smart ring and all that, right? But in the, the end of like, uh, and not the end, the last six years of the company was building ML compilers. [00:15:47] Anush Elangovan: And ml, ML compilers are like super, uh, complicated, sophisticated, advanced algorithms, dah, dah, dah. but it was all open source, right? So our VCs were like, wait, what do you mean your core [00:16:00] IP is open source? And um, the speed is the moat applied even then, right? It was just like, yes, if you have an idea that. [00:16:08] Anush Elangovan: Because someone saw this idea that you are, they're gonna be able to catch up, then you probably have the wrong idea anyway. But if they are, you know, you execute and they're gonna catch up, that you should assume they're gonna catch up. Right? So you gotta move forward. So keeping it open source is super important. [00:16:25] Anush Elangovan: But also to your question on like, you know, the learnings from an AMD standpoint, right? If there are, hard problems, I'd say dig in and work through it, right? Like there's no way but through it, right? That should be the simple mentality. And more, uh, frequently than not. you'll see that you'll just make it through in a, in, in good form. [00:16:52] Anush Elangovan: But if you doubt it and you're like, oh, I don't know if I should commit, if I'm, I, you know, what should just commit to do the right thing [00:17:00] every step, right? Every step, and just keep taking one step in front of the other. And in no time you'll see that you'll be running. Right. And, and yes, the first few steps will be like, yeah, everyone's complaining about your software quality. [00:17:15] Anush Elangovan: Everyone's complaining about this and that, and it doesn't work. And, and a few steps in, you know, you get, you get the hang of all the complaints that are coming in. You get the feedback loop. You're like, okay, what, what are you prioritizing again? One step in front of the other, right? You just keep knocking that out and then you get to a point where you're, it just becomes second nature, right? To do the, to do the right thing. And, and then yes, if someone gives you two options, you'll be like, fine. This is, uh, you know, there's always the resource trade off. There's always a human capital trade off, but what's the right thing to do? of course, I, I'm pragmatic about what we choose, but, but if the right thing for your long-term success is dig in, go first, principles, make it [00:18:00] happen. [00:18:00] Anush Elangovan: Well. Then just go for that. There's, there is no shortcut to [00:18:04] Andrew Zigler: acknowledging, you know, how it aligns with your mission, your core company goals, and what you're looking to achieve. And, and I, I love how you rightfully called out that in the open source world and you know, you have your technology that you've built, what you think is your moat upon, right? [00:18:22] Andrew Zigler: It's your code and, and to open source that, or to just make it where anyone could peer in is, you know. Scary in one regard, but two, it just kind of feels like you're handing away your throne room in some kind of sense, a very direct feeling sense. But the ultimately, you were really right to call out, and this is something I think about all the time, that the real power there is still the speed This the speed. [00:18:42] Andrew Zigler: That was the moat at the beginning of our conversation. It's the speed in combination with your. Very specific domain understanding of what you're building and what you're creating, and your new role as the steward of that world and how people plug into it, which [00:19:00] has frankly, a lot more influence and power than lording over a closed. [00:19:04] Andrew Zigler: You know, repository or an ecosystem, and like you said, like throwing things over the wall. Sure. There, there might be people always on the other side of that wall, but you're not gonna have a great connection with them. You're not gonna be able to really clearly understand them. I, I like your metaphor of the side of the field of the mountain a lot more. [00:19:23] Andrew Zigler: But, but in the, in this world, you know, where. That speed is, is the power and, and open source is just one way that you can harness that speed to get really far ahead and to innovate. , There's other parts of this equation that you can be experimenting with too, and I'd love to pick your brain about them as a software leader and, and, and one of them is about looking forward and kind of understanding that future that we're all building towards and beyond today's models and hardware. [00:19:48] Andrew Zigler: You know, what do you see as the next major bottleneck or opportunity in the AI compute space? As, as you know, enterprises and folks start to get a little more mature about what's available to [00:20:00] them. [00:20:00] Anush Elangovan: Yeah, I think, the bottleneck and opportunity is, uh, what I'd call, call walking the last mile of ai. Right. Uh, and like I I, I gave you an example, uh, previously, but, but it's similar to that. It's like there are cases where Humans have so many, uh, things to do in your day. You know, like the, if we sit down and actually had a customer focus like, okay, these customers lives, I'm gonna save four hours of this customer's life. And if you actually sit down and look at all of that, it'll be. Easily automatable, easily you know, uh, applicable, uh, for ai, right? [00:20:39] Anush Elangovan: Like, but then making it happen is gonna take a little bit, right? It's like maybe it's, uh, paying your utility bill, right? Or something like that, right? Or, or, your healthcare explanation of benefits. Uh, like, I'm sure you get an explanation of benefits, and I'm like, I, I don't even know what that thing is. [00:20:55] Anush Elangovan: It's just like EOB and like. [00:20:57] Andrew Zigler: it's a big, a big old PDF. Yeah, [00:21:00] exactly. [00:21:01] Anush Elangovan: Like, like, I'm like great straight to the, uh, shredder, right? And but that could be, you know, automated with the ai, right? It, it, it'd be like, Hey, the summary of this thing is you went and visited this day. Everything is okay. Everything is paid for, so don't worry, it's not a bill. [00:21:17] Anush Elangovan: That again, the same, uh, thing, but the sense of what that information overload is could be. Digested by ai, uh, accumulated over time and retrieved when you need it. Like, I don't, I actually don't even need to know this EOB right now, unless of course, whenever I need to know it, that maybe, you know, like for some benefits I need to figure out what do, what did I do over the past year and how do I apply it? Source:

Mike

15,145 次观看 • 9 个月前

Alex Hormozi’s business advice has taken the world by storm. He’s published two books and has more than 9,000,000 followers on the Internet. This is the first interview he’s ever done all about his writing process. Some highlights: 1. Be loyal to the truth, not your own ideas. 2. Silence the world when you write: Alex closes the windows, wears earplugs, and uses noise-cancelling headphones to create an environment where the outside world ceases to exist. 3. His book, $100M Leads went through 19 drafts (so expect to rewrite, a lot). 4. There’s a big difference between becoming known and becoming respected. Don’t let an algorithm convince you otherwise. 5. The pain is the pitch: The more vividly you describe someone’s problem, the less you need to sell the solution. 6. People buy things from people who can describe their pain better than they can. They assume: “If you understand my problem that well, you must have the solution.” 7. Sell at the point of greatest deprivation, not satisfaction. Offer the steak when someone is starving, not when they’re already full. 8. When writing ads, capture moments of pain as specifically as possible. Don’t say “I was overweight” when you can say “I wore a cover-up at the beach and avoided photos.” 9. Structure your writing time around long, uninterrupted blocks of time. Alex shoots for six hour blocks, even if that means waking up at 5am. 10. Memory is unreliable, so capture your best stories. Alex has an Excel spreadsheet with more than 600 stories from his life. 11. The #1 creator mistake: They keep building new products for their audience rather than building more audience for their product. 12. Alex’s best ideas come from reconciling contradictions. He says: “I look for places where two things seem true, but seem to conflict.” 13. The back cover should function as a mini sales letter. Each bullet point should tease a lesson in a way that makes people curious​. 14. If an idea can’t be operationalized, it’s useless – “What does this change about what someone actually does?” If the answer is nothing, cut it. I’ve shared the full interview with Alex Hormozi below. And if audio is your thing, I’ve linked to Apple, Spotify, YouTube in the reply tweets. — — Timestamps Below 00:00:00 Intro 00:00:20 Hormozi’s cave 00:05:35 Discovery process 00:12:32 Fix problems using MECE 00:13:43 Marketing “$100M Leads” 00:20:55 Writing for YouTube vs Books 00:27:13 The pain is the pitch 00:31:37 Break down the world into frameworks 00:34:09 10x more effort = 1000x results 00:38:24 Understanding the table of contents 00:49:27 The only app I use to read articles (Readwise Reader) 00:50:53 How to think about your audience 00:54:22 The #1 creator mistake 00:57:58 The business of Hormozi’s books 01:02:04 The love for writing as a kid 01:04:34 Hormozi’s ideal university writing class
1:09:04

Alex Hormozi’s business advice has taken the world by storm. He’s published two books and has more than 9,000,000 followers on the Internet. This is the first interview he’s ever done all about his writing process. Some highlights: 1. Be loyal to the truth, not your own ideas. 2. Silence the world when you write: Alex closes the windows, wears earplugs, and uses noise-cancelling headphones to create an environment where the outside world ceases to exist. 3. His book, $100M Leads went through 19 drafts (so expect to rewrite, a lot). 4. There’s a big difference between becoming known and becoming respected. Don’t let an algorithm convince you otherwise. 5. The pain is the pitch: The more vividly you describe someone’s problem, the less you need to sell the solution. 6. People buy things from people who can describe their pain better than they can. They assume: “If you understand my problem that well, you must have the solution.” 7. Sell at the point of greatest deprivation, not satisfaction. Offer the steak when someone is starving, not when they’re already full. 8. When writing ads, capture moments of pain as specifically as possible. Don’t say “I was overweight” when you can say “I wore a cover-up at the beach and avoided photos.” 9. Structure your writing time around long, uninterrupted blocks of time. Alex shoots for six hour blocks, even if that means waking up at 5am. 10. Memory is unreliable, so capture your best stories. Alex has an Excel spreadsheet with more than 600 stories from his life. 11. The #1 creator mistake: They keep building new products for their audience rather than building more audience for their product. 12. Alex’s best ideas come from reconciling contradictions. He says: “I look for places where two things seem true, but seem to conflict.” 13. The back cover should function as a mini sales letter. Each bullet point should tease a lesson in a way that makes people curious​. 14. If an idea can’t be operationalized, it’s useless – “What does this change about what someone actually does?” If the answer is nothing, cut it. I’ve shared the full interview with Alex Hormozi below. And if audio is your thing, I’ve linked to Apple, Spotify, YouTube in the reply tweets. — — Timestamps Below 00:00:00 Intro 00:00:20 Hormozi’s cave 00:05:35 Discovery process 00:12:32 Fix problems using MECE 00:13:43 Marketing “$100M Leads” 00:20:55 Writing for YouTube vs Books 00:27:13 The pain is the pitch 00:31:37 Break down the world into frameworks 00:34:09 10x more effort = 1000x results 00:38:24 Understanding the table of contents 00:49:27 The only app I use to read articles (Readwise Reader) 00:50:53 How to think about your audience 00:54:22 The #1 creator mistake 00:57:58 The business of Hormozi’s books 01:02:04 The love for writing as a kid 01:04:34 Hormozi’s ideal university writing class

David Perell

33,676 次观看 • 1 年前

$AMD| The FOMO to buy AMD Chips is NOW 🧵 Not Financial Advice! DYOR! Research Purpose Only! The Inference Queen is the biggest winner in Agentic AI where all other CPUs are struggling to compete with a 2yr old EPYC Turin and EPYC Venice is in mass production phase. AMD stresses deployability today on standard x86 platforms (no proprietary architectures required), full software compatibility, and open standards. This positions Venice + Helios as a practical, high-density alternative to competing solutions while underscoring that agentic AI shifts the balance toward CPU-rich racks alongside GPUs, and most importantly, lowering the cost of token to accelerate adoption and innovation. Context: The Wall Street Journal yesterday came out with an article that OpenAI is condiering drasstically lowering the token prices to win more customers from Anthropic. The narrative "they" are trying to exacerbate the current AI selloff won't last long. This is a fundamental misunderstanding of what is going on, or what I already discussed for months and years. Followers and Subscribers already knew this for years, that this day would come, where token cost will bcome the central discussion among enterprises as there is no such thing as unlimited budget or Tokenmaxxing when they use $NVDA chips or In-house Hyperscalers chips. I will link various threads if you are interested in understanding the full picture from supply chain to recent TSMC Rapid 2nm expansion up to 12 Fabs total by 2027/2028. Hyperscalers and AI natives effectively have no choice but to buy more AMD system for Agentic AI as leadership in economical, power-aware, high-volume internal + agentic use. However, due to supply constraints where Supply is far behind Demand, this makes multi-vendor reality along with in-house chips drive faster industry progress, lower overall costs, and better sustainability. NVIDIA’s Vera Rubin cannot compete with a 2 years old EPYC Turin, but AMD under Dr. Lisa Su has engineered the lowest cost-per-million-tokens, highly competitive energy-efficient solutions, and superior CPU orchestration for agentic AI at scale with Helios. Dr. Su has championed this shift since at least 2023, foreseeing the rise of agentic workflows that demand far more orchestration, parallel agents, and balanced compute well before the industry fully embraced it. Her long-term vision of AI moving from simple prompts to always on, multi-agent systems has driven AMD’s investments in high-core EPYC CPUs and integrated rack-scale solutions, perfectly positioning the company for today’s realities. The OpenAI-AMD 1GW Helios deployment (starting H2 2026) represents a pivotal vertical integration move that directly supercharges the inference economics. This isn't incremental; it's a structural shift toward ownership of massive, optimized rack-scale capacity, enabling the lowest token costs and triggering the enterprise adoption flywheel. We need to be honest, $AMD is the only company that made a big bet on Inference since the day Chatgpt became sensational where $NVDA and others were betting big on Training. At the end of the day, Token bill from Anthropic has to obey economics. Meaning the bills rise, companies have to get more out of it to justify the cost. It cannot be an unlimited inference budget, and it has to show up on efficiency, profitability and operating leverage. 1. Tokenomics After you understand this, you will understand why Citi cited Anthropic is likely to sign a deal with $AMD along with Hyperscalers, AI Labs, Sovereign AI like Softbank 5GW in France and many other countries. However, OpenAI and $META are now wanting faster deployment, and they are AMD shareholders now, they have prioritized allocation. Anthropic and Hyperscalers just cannot compete when Helios Rack lower token cost to$0.0003–$0.0005 per million tokens at GW scale. Cost to build 1GW data center 1GW Helios Rack full build is estimated $30-$35B 1GW Rubin Rack full build is estimated $45-$55B Inference (Cost per Million Tokens) ~$NVDA B200 / HGX: ~$0.02–$0.08 on optimized workloads (FP4/MXFP4, speculative decoding). Significant improvement over Hopper but still premium-priced. GB200 NVL72 rack-scale: $0.05–$0.25+ ~$AMD Helios Racks: $0.0003-$0.0005 per M tokens, dramatically lower than NVIDIA equivalents in owned infra. MI355X node-level: Up to 40% more tokens per dollar vs. competing solutions ( B200), driven by higher memory capacity (up to 288GB+ HBM), strong bandwidth, and lower acquisition costs. Training ~$NVDA Rubin Rack is estimated $0.7-$1.2/M Tokens ~$AMD Helios Rack is estimated $0.65-$1.0/M Tokens Now, OpenAI, META and Hyperscalers can lower Inference cost even further with $AMD EPYC Venice "dense rack" or Agentic AI Rack. AMD published a detailed technical blog emphasizing that the future of agentic AI autonomous, multi-step AI systems requiring heavy orchestration, databases, caching, APIs, and control planes demands massive CPU-dense rack-scale infrastructure, not just GPUs. The catalyst prominently positions their upcoming 6th Gen EPYC "Venice" processors as the key enabler for next-generation dense racks, delivering leadership throughput under real-world power, cooling, and density constraints. ~EPYC Venice (Zen 6 architecture, up to 256 cores / 512 threads per socket) is projected to deliver exceptional rack-level performance. In AMD’s modeled 100 kW rack comparisons, Venice-powered systems are expected to achieve ~3.30x the throughput of NVIDIA’s Vera (88-core Olympus) baseline across a broad mix of agentic-supporting workloads. ~This builds on current-generation 5th Gen EPYC "Turin" (up to 192 cores), which already delivers ~2.37x rack throughput vs. Vera and ~1.6x vs. Intel’s Xeon 6980P (128 cores). ~ Liquid-cooled Turin deployments already support >27,000 CPU cores per rack today. Venice is architected to push this beyond 36,000 cores in the same rack class, dramatically increasing concurrent agent capacity and overall infrastructure efficiency. 2. Ownership vs renting compute from Hyperscalers matter to OpenAI and only owning $AMD chips can meaningfully lower token cost for enterprises. ~Eliminates cloud overhead: No provider margins, utilization buffers, or egress fees. Direct control over power contracts, cooling, scheduling, and orchestration at dedicated facilities. ~Helios optimizations at GW scale: Rack-level density (1.4+ exaFLOPS FP8 per rack), high HBM4 bandwidth, EPYC orchestration for agentic workloads, and superior TCO/TDP. AMD's long-standing focus on tokens per dollar/watt shines here 20-40%+ efficiency edges in inference-heavy scenarios. ~At 1GW+ optimized deployment, inference hits $0.0003–$0.0005 per million tokens (community/analyst models tied to Helios metrics). This is dramatically lower than typical rented/cloud equivalents, especially for high-volume output tokens in agentic flows. High token bills today, enterprises running heavy agentic/coding/analysis workloads can face $50-100M+/month at current API rates (flagship models $5-30+/M output, scaled to massive volumes). Post-Helios compression, same volume will drop to $10-15M/month (or better) via lower underlying costs passed through as pricing flexibility, volume tiers, caching, or batch discounts. ROI thresholds collapse. More companies greenlight pilots → production → massive scaling. Agentic AI (autonomous workflows) multiplies token demand exponentially, but affordability removes the friction. OpenAI gains flexibility, Unlike more cloud-dependent rivals (Anthropic), they can lower effective pricing, offer aggressive enterprise bundles, or absorb volume without margin destruction directly tackling "high token bill" complaints while maintaining profitability as usage explodes. 3. Agentic AI Models shifted CPU:GPU Ratio to 1:1 toward 3-5:1 with Explosively Token-Hungry Workloads Agentic AI (autonomous, multi-step agents with planning, tool use, iteration, and self-correction) is fundamentally more compute and token intensive than conversational or single-turn generative AI. Agentic AI. autonomous, multi-step workflows with orchestration, tool use, parallel agents, data movement, and enterprise integration has dramatically increased the importance of strong host CPUs alongside GPUs. This shifts the CPU-to-GPU ratio higher and makes balanced systems critical toward 1:1 to 5:1 as enterprises testing more than 5-10 agents. AMD EPYC Venice excels ~Leadership core density (up to 256 Zen 6 cores per socket) for running many agents in parallel, orchestration layers, and high-throughput control-plane tasks. ~Superior performance-per-core and power efficiency ( up to 2.1x higher perf/core and 2.26x better SPECpower vs. NVIDIA Grace in benchmarks). ~Tight integration in Helios: One Venice CPU + multiple MI450 GPUs per node, enabling efficient data feeding to GPUs ("zero-copy"), parallel execution, and full rack utilization for complex agentic loops. Hyperscalers (Meta, Microsoft, Amazon, Google, Softbank) and AI natives (OpenAI, Anthropic...) are adopting high-core EPYC at scale specifically for these agentic demands, as CPUs now handle a larger share of non-model work (orchestration, policy enforcement, tool calls). This complements AMD’s lower-cost GPUs for overall TCO wins. ~Agents often generate 10–100x+ more tokens per task due to iterative reasoning chains, multiple tool calls, verification loops, and long-context orchestration. ~Goldman Sachs forecasts token consumption multiplying 24x by 2030 (to 120 quadrillion tokens/month) largely driven by agentic adoption in consumer and enterprise. ~Enterprise data shows agent-pattern workloads growing at 680% annualized rates, projected to surpass conversational AI in token volume by Q3 2026. ~Daily enterprise agent token consumption is already in the billions, with complex workflows (coding, workflows, analysis) amplifying this dramatically. 4. Competitive Edge: Winning Customers from Anthropic Anthropic’s Claude models (especially Opus/Sonnet) excel in complex reasoning and agentic coding, commanding premium positioning. However, their higher underlying costs (heavier reliance on third-party cloud with margins) limit pricing flexibility compared to OpenAI’s owned Helios capacity. Anthropic is on track to generate $10.9 billion in Q2 revenue. The company expects to achieve its first-ever quarterly adjusted operating profit of $559 million. However, sustaining full-year profitability remains challenging due to immense computing and model training costs The truth is, Anthropic has no choice but to buy as much $AMD chips as possible if they want to compete with OpenAI or get investors attention. This 5% adjusted operating profit to revenue ratio is just pathetic. Current pricing dynamics (2026): OpenAI already undercuts on many tiers ( flagship output tokens significantly cheaper than equivalent Claude Opus). Nano/mini models offer 5–10x advantages for volume work. Anthropic holds edges in long-context flat pricing and certain reasoning quality. OpenAI after Helios Rack Ownership, At $0.0003–$0.0005/M effective costs, OpenAI gains massive headroom to: ~Aggressively discount high-volume agentic tiers or bundles. ~Offer “unlimited” enterprise plans or usage-based models that Anthropic struggles to match without margin erosion. ~Target cost-sensitive, high-throughput agent deployments (dev tools, automation platforms) where token bills explode. Enterprises facing $ millions in monthly agentic bills will migrate to the provider delivering better economics at scale. OpenAI’s combination of strong models (o-series reasoning) + lowest TCO positions it to erode Anthropic’s enterprise share, especially as agentic becomes the dominant token consumer. Cheaper tokens expand the total addressable market dramatically. This feeds the data/model improvement loop, justifying further capex. AMD benefits from proven scale pulling in more customers (Meta, Oracle, Microsfot, Amazon, Softbank, TensorWave, LumaAI ... already aligned on Helios). Conclusion: Dr. Lisa Su has been laser focused on inference economics since at least 2022–2023, repeatedly emphasizing that the real battleground for AI scalability would be TCO, power efficiency (TDP), and ultimately tokens per dollar and per watt not just raw training FLOPS. While many viewed inference as a secondary, commoditized workload, Dr. Su architected AMD’s roadmap around rack-scale systems optimized for high-volume, sustained inference that would dominate as models matured and usage exploded. Helios represents the culmination of that multi-year bet: a fully integrated, open platform designed precisely for the economics of massive token throughput. This deep, strategic partnership with OpenAI starting with the 1GW Helios deployment in H2 2026 and scaling to 6GW, is the embodiment of that shared vision. Both companies foresaw a future where agentic AI models evolve to become extraordinarily token-hungry: autonomous agents executing complex, iterative workflows with planning, tool use, verification loops, and long-context reasoning. These workloads can consume 100x+ more tokens per task than traditional chat or single-turn generation, driving exponential demand as capabilities improve and enterprises deploy them at scale. By owning and optimizing this massive Helios capacity at GW scale, OpenAI achieves inference costs as low as $0.0003–$0.0005 per million tokens. This structural cost advantage allows OpenAI to absorb the coming token explosion profitably, dramatically lower effective pricing for enterprises, and win high-volume agentic workloads from higher-cost competitors like Anthropic. What was once a prohibitive monthly token bill becomes an affordable accelerator for productivity and innovation. The OpenAI-AMD alliance validates Dr. Su’s prescient strategy and turns the Agentic flywheel into reality: Collapsing inference costs → explosive token consumption → richer data and better models → accelerate greater demand. This partnership doesn’t just address today’s economics, it positions both leaders at the center of the infrastructure buildout that will power AI’s next decade. By delivering the lowest inference economics at scale, OpenAI not only solves enterprise bill pain but gains a decisive weapon to win share from higher-cost rivals like Anthropic. And that is why OpenAI and $META will deploy EPYC Dense Rack Not Financial Advice! DYOR! Research Purpose Only!
4:46

$AMD| The FOMO to buy AMD Chips is NOW 🧵 Not Financial Advice! DYOR! Research Purpose Only! The Inference Queen is the biggest winner in Agentic AI where all other CPUs are struggling to compete with a 2yr old EPYC Turin and EPYC Venice is in mass production phase. AMD stresses deployability today on standard x86 platforms (no proprietary architectures required), full software compatibility, and open standards. This positions Venice + Helios as a practical, high-density alternative to competing solutions while underscoring that agentic AI shifts the balance toward CPU-rich racks alongside GPUs, and most importantly, lowering the cost of token to accelerate adoption and innovation. Context: The Wall Street Journal yesterday came out with an article that OpenAI is condiering drasstically lowering the token prices to win more customers from Anthropic. The narrative "they" are trying to exacerbate the current AI selloff won't last long. This is a fundamental misunderstanding of what is going on, or what I already discussed for months and years. Followers and Subscribers already knew this for years, that this day would come, where token cost will bcome the central discussion among enterprises as there is no such thing as unlimited budget or Tokenmaxxing when they use $NVDA chips or In-house Hyperscalers chips. I will link various threads if you are interested in understanding the full picture from supply chain to recent TSMC Rapid 2nm expansion up to 12 Fabs total by 2027/2028. Hyperscalers and AI natives effectively have no choice but to buy more AMD system for Agentic AI as leadership in economical, power-aware, high-volume internal + agentic use. However, due to supply constraints where Supply is far behind Demand, this makes multi-vendor reality along with in-house chips drive faster industry progress, lower overall costs, and better sustainability. NVIDIA’s Vera Rubin cannot compete with a 2 years old EPYC Turin, but AMD under Dr. Lisa Su has engineered the lowest cost-per-million-tokens, highly competitive energy-efficient solutions, and superior CPU orchestration for agentic AI at scale with Helios. Dr. Su has championed this shift since at least 2023, foreseeing the rise of agentic workflows that demand far more orchestration, parallel agents, and balanced compute well before the industry fully embraced it. Her long-term vision of AI moving from simple prompts to always on, multi-agent systems has driven AMD’s investments in high-core EPYC CPUs and integrated rack-scale solutions, perfectly positioning the company for today’s realities. The OpenAI-AMD 1GW Helios deployment (starting H2 2026) represents a pivotal vertical integration move that directly supercharges the inference economics. This isn't incremental; it's a structural shift toward ownership of massive, optimized rack-scale capacity, enabling the lowest token costs and triggering the enterprise adoption flywheel. We need to be honest, $AMD is the only company that made a big bet on Inference since the day Chatgpt became sensational where $NVDA and others were betting big on Training. At the end of the day, Token bill from Anthropic has to obey economics. Meaning the bills rise, companies have to get more out of it to justify the cost. It cannot be an unlimited inference budget, and it has to show up on efficiency, profitability and operating leverage. 1. Tokenomics After you understand this, you will understand why Citi cited Anthropic is likely to sign a deal with $AMD along with Hyperscalers, AI Labs, Sovereign AI like Softbank 5GW in France and many other countries. However, OpenAI and $META are now wanting faster deployment, and they are AMD shareholders now, they have prioritized allocation. Anthropic and Hyperscalers just cannot compete when Helios Rack lower token cost to$0.0003–$0.0005 per million tokens at GW scale. Cost to build 1GW data center 1GW Helios Rack full build is estimated $30-$35B 1GW Rubin Rack full build is estimated $45-$55B Inference (Cost per Million Tokens) ~$NVDA B200 / HGX: ~$0.02–$0.08 on optimized workloads (FP4/MXFP4, speculative decoding). Significant improvement over Hopper but still premium-priced. GB200 NVL72 rack-scale: $0.05–$0.25+ ~$AMD Helios Racks: $0.0003-$0.0005 per M tokens, dramatically lower than NVIDIA equivalents in owned infra. MI355X node-level: Up to 40% more tokens per dollar vs. competing solutions ( B200), driven by higher memory capacity (up to 288GB+ HBM), strong bandwidth, and lower acquisition costs. Training ~$NVDA Rubin Rack is estimated $0.7-$1.2/M Tokens ~$AMD Helios Rack is estimated $0.65-$1.0/M Tokens Now, OpenAI, META and Hyperscalers can lower Inference cost even further with $AMD EPYC Venice "dense rack" or Agentic AI Rack. AMD published a detailed technical blog emphasizing that the future of agentic AI autonomous, multi-step AI systems requiring heavy orchestration, databases, caching, APIs, and control planes demands massive CPU-dense rack-scale infrastructure, not just GPUs. The catalyst prominently positions their upcoming 6th Gen EPYC "Venice" processors as the key enabler for next-generation dense racks, delivering leadership throughput under real-world power, cooling, and density constraints. ~EPYC Venice (Zen 6 architecture, up to 256 cores / 512 threads per socket) is projected to deliver exceptional rack-level performance. In AMD’s modeled 100 kW rack comparisons, Venice-powered systems are expected to achieve ~3.30x the throughput of NVIDIA’s Vera (88-core Olympus) baseline across a broad mix of agentic-supporting workloads. ~This builds on current-generation 5th Gen EPYC "Turin" (up to 192 cores), which already delivers ~2.37x rack throughput vs. Vera and ~1.6x vs. Intel’s Xeon 6980P (128 cores). ~ Liquid-cooled Turin deployments already support >27,000 CPU cores per rack today. Venice is architected to push this beyond 36,000 cores in the same rack class, dramatically increasing concurrent agent capacity and overall infrastructure efficiency. 2. Ownership vs renting compute from Hyperscalers matter to OpenAI and only owning $AMD chips can meaningfully lower token cost for enterprises. ~Eliminates cloud overhead: No provider margins, utilization buffers, or egress fees. Direct control over power contracts, cooling, scheduling, and orchestration at dedicated facilities. ~Helios optimizations at GW scale: Rack-level density (1.4+ exaFLOPS FP8 per rack), high HBM4 bandwidth, EPYC orchestration for agentic workloads, and superior TCO/TDP. AMD's long-standing focus on tokens per dollar/watt shines here 20-40%+ efficiency edges in inference-heavy scenarios. ~At 1GW+ optimized deployment, inference hits $0.0003–$0.0005 per million tokens (community/analyst models tied to Helios metrics). This is dramatically lower than typical rented/cloud equivalents, especially for high-volume output tokens in agentic flows. High token bills today, enterprises running heavy agentic/coding/analysis workloads can face $50-100M+/month at current API rates (flagship models $5-30+/M output, scaled to massive volumes). Post-Helios compression, same volume will drop to $10-15M/month (or better) via lower underlying costs passed through as pricing flexibility, volume tiers, caching, or batch discounts. ROI thresholds collapse. More companies greenlight pilots → production → massive scaling. Agentic AI (autonomous workflows) multiplies token demand exponentially, but affordability removes the friction. OpenAI gains flexibility, Unlike more cloud-dependent rivals (Anthropic), they can lower effective pricing, offer aggressive enterprise bundles, or absorb volume without margin destruction directly tackling "high token bill" complaints while maintaining profitability as usage explodes. 3. Agentic AI Models shifted CPU:GPU Ratio to 1:1 toward 3-5:1 with Explosively Token-Hungry Workloads Agentic AI (autonomous, multi-step agents with planning, tool use, iteration, and self-correction) is fundamentally more compute and token intensive than conversational or single-turn generative AI. Agentic AI. autonomous, multi-step workflows with orchestration, tool use, parallel agents, data movement, and enterprise integration has dramatically increased the importance of strong host CPUs alongside GPUs. This shifts the CPU-to-GPU ratio higher and makes balanced systems critical toward 1:1 to 5:1 as enterprises testing more than 5-10 agents. AMD EPYC Venice excels ~Leadership core density (up to 256 Zen 6 cores per socket) for running many agents in parallel, orchestration layers, and high-throughput control-plane tasks. ~Superior performance-per-core and power efficiency ( up to 2.1x higher perf/core and 2.26x better SPECpower vs. NVIDIA Grace in benchmarks). ~Tight integration in Helios: One Venice CPU + multiple MI450 GPUs per node, enabling efficient data feeding to GPUs ("zero-copy"), parallel execution, and full rack utilization for complex agentic loops. Hyperscalers (Meta, Microsoft, Amazon, Google, Softbank) and AI natives (OpenAI, Anthropic...) are adopting high-core EPYC at scale specifically for these agentic demands, as CPUs now handle a larger share of non-model work (orchestration, policy enforcement, tool calls). This complements AMD’s lower-cost GPUs for overall TCO wins. ~Agents often generate 10–100x+ more tokens per task due to iterative reasoning chains, multiple tool calls, verification loops, and long-context orchestration. ~Goldman Sachs forecasts token consumption multiplying 24x by 2030 (to 120 quadrillion tokens/month) largely driven by agentic adoption in consumer and enterprise. ~Enterprise data shows agent-pattern workloads growing at 680% annualized rates, projected to surpass conversational AI in token volume by Q3 2026. ~Daily enterprise agent token consumption is already in the billions, with complex workflows (coding, workflows, analysis) amplifying this dramatically. 4. Competitive Edge: Winning Customers from Anthropic Anthropic’s Claude models (especially Opus/Sonnet) excel in complex reasoning and agentic coding, commanding premium positioning. However, their higher underlying costs (heavier reliance on third-party cloud with margins) limit pricing flexibility compared to OpenAI’s owned Helios capacity. Anthropic is on track to generate $10.9 billion in Q2 revenue. The company expects to achieve its first-ever quarterly adjusted operating profit of $559 million. However, sustaining full-year profitability remains challenging due to immense computing and model training costs The truth is, Anthropic has no choice but to buy as much $AMD chips as possible if they want to compete with OpenAI or get investors attention. This 5% adjusted operating profit to revenue ratio is just pathetic. Current pricing dynamics (2026): OpenAI already undercuts on many tiers ( flagship output tokens significantly cheaper than equivalent Claude Opus). Nano/mini models offer 5–10x advantages for volume work. Anthropic holds edges in long-context flat pricing and certain reasoning quality. OpenAI after Helios Rack Ownership, At $0.0003–$0.0005/M effective costs, OpenAI gains massive headroom to: ~Aggressively discount high-volume agentic tiers or bundles. ~Offer “unlimited” enterprise plans or usage-based models that Anthropic struggles to match without margin erosion. ~Target cost-sensitive, high-throughput agent deployments (dev tools, automation platforms) where token bills explode. Enterprises facing $ millions in monthly agentic bills will migrate to the provider delivering better economics at scale. OpenAI’s combination of strong models (o-series reasoning) + lowest TCO positions it to erode Anthropic’s enterprise share, especially as agentic becomes the dominant token consumer. Cheaper tokens expand the total addressable market dramatically. This feeds the data/model improvement loop, justifying further capex. AMD benefits from proven scale pulling in more customers (Meta, Oracle, Microsfot, Amazon, Softbank, TensorWave, LumaAI ... already aligned on Helios). Conclusion: Dr. Lisa Su has been laser focused on inference economics since at least 2022–2023, repeatedly emphasizing that the real battleground for AI scalability would be TCO, power efficiency (TDP), and ultimately tokens per dollar and per watt not just raw training FLOPS. While many viewed inference as a secondary, commoditized workload, Dr. Su architected AMD’s roadmap around rack-scale systems optimized for high-volume, sustained inference that would dominate as models matured and usage exploded. Helios represents the culmination of that multi-year bet: a fully integrated, open platform designed precisely for the economics of massive token throughput. This deep, strategic partnership with OpenAI starting with the 1GW Helios deployment in H2 2026 and scaling to 6GW, is the embodiment of that shared vision. Both companies foresaw a future where agentic AI models evolve to become extraordinarily token-hungry: autonomous agents executing complex, iterative workflows with planning, tool use, verification loops, and long-context reasoning. These workloads can consume 100x+ more tokens per task than traditional chat or single-turn generation, driving exponential demand as capabilities improve and enterprises deploy them at scale. By owning and optimizing this massive Helios capacity at GW scale, OpenAI achieves inference costs as low as $0.0003–$0.0005 per million tokens. This structural cost advantage allows OpenAI to absorb the coming token explosion profitably, dramatically lower effective pricing for enterprises, and win high-volume agentic workloads from higher-cost competitors like Anthropic. What was once a prohibitive monthly token bill becomes an affordable accelerator for productivity and innovation. The OpenAI-AMD alliance validates Dr. Su’s prescient strategy and turns the Agentic flywheel into reality: Collapsing inference costs → explosive token consumption → richer data and better models → accelerate greater demand. This partnership doesn’t just address today’s economics, it positions both leaders at the center of the infrastructure buildout that will power AI’s next decade. By delivering the lowest inference economics at scale, OpenAI not only solves enterprise bill pain but gains a decisive weapon to win share from higher-cost rivals like Anthropic. And that is why OpenAI and $META will deploy EPYC Dense Rack Not Financial Advice! DYOR! Research Purpose Only!

Mike

84,951 次观看 • 3 个月前

A beanie that reads your thoughts and turns them into text — no surgery required?@jason grills the co-founders on their noninvasive brain-computer interface, backed by Vinod Khosla, and calls cap on the whole thing (until he doesn't). This episode of This Week in Startups covers a lot of ground: Jason's tactical tip of the day on making everyone the CEO of their domain, a deep dive into Sabi's thought-to-text beanie, a live demo of AI-powered podcast sidebars built by the TWiST audience, the announcement of a new $5K bounty for an annotation tool, and Jason's big five wellness framework. 0:00 Intro & tactical tip: Make everyone the CEO of their domain 1:49 Matt Coffin's "CEO of X" management philosophy 3:04 Community ownership: Deputizing Ricky, Lawn, Bianca, Maddie, Kabir 4:02 Building the Noti Gang: X group chats and community flywheels 5:08 Founder takeaways: Activate your top 1%, make someone the CEO of it 6:18 The streamer trick, parasocial dynamics, and creator ethics 7:59 Plaud: If your work depends on conversations — interviews, meetings, calls — you need a Plaud NotePin. You can check it out at and use code TWIST for 10% off! 9:34 Guest intro: Rahul Chhabra of Sabi 10:01 What is Sabi? Noninvasive BCI in a beanie with 100,000 sensors 10:02 LinkedIn Jobs - Hire right, the first time. Post your first job and get $100 off towards your job post at 11:02 How it works: From fMRI to EEG, from hospital to hat 13:46 The brain foundation model and thought-to-text decoding 15:55 Vetting the founders: BITS Pilani, Stanford, athlete fatigue AI 17:22 Vinod Khosla's investment thesis: BCI must be noninvasive 19:30 Jason's challenge: Say "Calacanis" or it doesn't count 19:50 Northwest Registered Agent: Get more when you start your business with Northwest. In 10 clicks and 10 minutes, you can form your company and walk away with a real business identity — Learn more at 21:30 Reserve your device, release by end of 2026 22:26 Privacy concerns: Does the beanie read everything? 23:47 Jason's Big Five wellness framework: sleep, nutrition, exercise, meditation, socialization 25:41 Bounty #1: AI live sidebar contest — demos from the TWiST audience 26:02 What the bounty asked for: AI personas watching the show in real time 27:04 Oliver's breakdown: What's easy vs. hard about live AI commentary 28:40 Demo #1: Armchair (by Mark Colebrook) — fact-checker + troll personas, live 30:26 Render: Find out why 5 million developers are already using the all-in-one cloud platform, Render. Go to and apply for the Render Startup Program to get $500-$100,000 in free credits, depending on your stage and backers. 35:30 Live political violence test: The sidebar in real time on the WHCD shooting 38:45 Demo #2: Pod Commentators / SideCast — browser-based, Gemini-powered 40:35 Jason's revised Bounty #1 spec: Fact-checker + cynic, public stream or Zoom 42:41 Timeline: check-ins May 1, May 8; final winner May 15 44:51 Demo #3: BMD Pat (by Patrick Hughes) — instant URL, all-snarky personas 46:52 Bounty #2 announced: — a fair-use multimedia annotation tool 49:31 Annotated: the of media commentary 51:11 Contest rules: Jason owns the domain, winner gets $5K + potential ongoing work 53:13 Wrap-up: Bounty 1 = AI sidebar, Bounty 2 = Annotated 🎥 Watch the full episode here 👇
55:43

A beanie that reads your thoughts and turns them into text — no surgery required?@jason grills the co-founders on their noninvasive brain-computer interface, backed by Vinod Khosla, and calls cap on the whole thing (until he doesn't). This episode of This Week in Startups covers a lot of ground: Jason's tactical tip of the day on making everyone the CEO of their domain, a deep dive into Sabi's thought-to-text beanie, a live demo of AI-powered podcast sidebars built by the TWiST audience, the announcement of a new $5K bounty for an annotation tool, and Jason's big five wellness framework. 0:00 Intro & tactical tip: Make everyone the CEO of their domain 1:49 Matt Coffin's "CEO of X" management philosophy 3:04 Community ownership: Deputizing Ricky, Lawn, Bianca, Maddie, Kabir 4:02 Building the Noti Gang: X group chats and community flywheels 5:08 Founder takeaways: Activate your top 1%, make someone the CEO of it 6:18 The streamer trick, parasocial dynamics, and creator ethics 7:59 Plaud: If your work depends on conversations — interviews, meetings, calls — you need a Plaud NotePin. You can check it out at and use code TWIST for 10% off! 9:34 Guest intro: Rahul Chhabra of Sabi 10:01 What is Sabi? Noninvasive BCI in a beanie with 100,000 sensors 10:02 LinkedIn Jobs - Hire right, the first time. Post your first job and get $100 off towards your job post at 11:02 How it works: From fMRI to EEG, from hospital to hat 13:46 The brain foundation model and thought-to-text decoding 15:55 Vetting the founders: BITS Pilani, Stanford, athlete fatigue AI 17:22 Vinod Khosla's investment thesis: BCI must be noninvasive 19:30 Jason's challenge: Say "Calacanis" or it doesn't count 19:50 Northwest Registered Agent: Get more when you start your business with Northwest. In 10 clicks and 10 minutes, you can form your company and walk away with a real business identity — Learn more at 21:30 Reserve your device, release by end of 2026 22:26 Privacy concerns: Does the beanie read everything? 23:47 Jason's Big Five wellness framework: sleep, nutrition, exercise, meditation, socialization 25:41 Bounty #1: AI live sidebar contest — demos from the TWiST audience 26:02 What the bounty asked for: AI personas watching the show in real time 27:04 Oliver's breakdown: What's easy vs. hard about live AI commentary 28:40 Demo #1: Armchair (by Mark Colebrook) — fact-checker + troll personas, live 30:26 Render: Find out why 5 million developers are already using the all-in-one cloud platform, Render. Go to and apply for the Render Startup Program to get $500-$100,000 in free credits, depending on your stage and backers. 35:30 Live political violence test: The sidebar in real time on the WHCD shooting 38:45 Demo #2: Pod Commentators / SideCast — browser-based, Gemini-powered 40:35 Jason's revised Bounty #1 spec: Fact-checker + cynic, public stream or Zoom 42:41 Timeline: check-ins May 1, May 8; final winner May 15 44:51 Demo #3: BMD Pat (by Patrick Hughes) — instant URL, all-snarky personas 46:52 Bounty #2 announced: — a fair-use multimedia annotation tool 49:31 Annotated: the of media commentary 51:11 Contest rules: Jason owns the domain, winner gets $5K + potential ongoing work 53:13 Wrap-up: Bounty 1 = AI sidebar, Bounty 2 = Annotated 🎥 Watch the full episode here 👇

This Week in Startups

16,398 次观看 • 4 个月前