First time I have seen $NVDA's Jensen having some issues in explaining $NVDA's moat vs ASICs like $GOOGL TPUs and $AMZN Trainium. Why does Anthropic use TPUs and Trainium? Jensen: "Anthropic is a unique instance, not a trend" I really got to tip my hat to Dwarkesh Patel; he... is one of the rare interviewers who first gets extremely well-prepared for an interview and, secondly, is true to his viewers more than the guests and asks tough questions, even though the interviewee might not like that.show more

@dwarkesh_sp He explained it all in the pod. They use it because when Anthropic came to Jensen for help—NVDA wasn’t (yet) in a position to deploy capital to them. So Anthropic turned to AMZN and GOOGL. Jensen also said it’s something he regrets.

@dwarkesh_sp Yes but exactly that argument is the problem. In this argument he is implicitly saying the hardware is not the real moat, the moat is who gives me the capital and then I buy his hardware.

@dwarkesh_sp Not what I took from it. Anthropic gave Jensen first right of refusal, then turned to GOOGL/AMZN silicon only after he declined. If Nvidia had deployed billions to Anthropic in 2022, its TPU/ASIC exposure would likely be much smaller but those commitments are locked in now.

I think the argument here is that nobody doubts that NVDA chips are better or surely were better than the ASICs at that point when Anthropic first looked at who to use but the question is are they today so much better that NVDA can charge a 75% margin, and here the arguments from Jensen for me from this just weren’t convincing enough.

@dwarkesh_sp Depends on whether you think Groq’s LPX fills the decode gap, since Nvidia GPUs still dominate prefill. If you can get both in one NVL system where NVDA takes 75% margin, why pay 65% ASIC margins to stitch multiple architectures together?

@dwarkesh_sp Dwarkesh is a fucking idiot. If I was Jensen I would just walk off, this street shitter grew a beard and now he’s like “I’m just like you guys!” Dude throwing extremes the entire time. MUH enriched uranium. Just shut the fuck up Patel

@dwarkesh_sp that question probably did more work than 20 earnings calls

@dwarkesh_sp More than anything, it’s the first time I’ve seen Jensen annoyed 😅 $NVDA

Your spin vs WTF 🤣 “Without Anthropic, why would there be any TPU growth at all? It's 100% Anthropic. Without Anthropic, why would there be Trainium growth at all? It's 100% Anthropic" WTF: Anthropic had to the trade inferior hardware for capital unburdening - the cheap, inferior stuff that no one else uses. Jensen stepped in once he realized VCs can’t fund the compute that Frontier labs need. And he did so to all labs. WTF #2: Even Google or Amazon external customers use Nvidia

@dwarkesh_sp what's more telling is that all of amazon / google chip business is essentially anthropic. and considering both those companies are major investors in anthropic.. that makes sense.

@dwarkesh_sp why does Apple use TPUs.

@dwarkesh_sp Apple uses TPUs to do what? To train it's models? that workload is a rounding error. and the results?? There are none. also - due to high demand and shortages of Nvidia chips, Apple likely used TPUs because of their availability on Google Cloud

@RihardJarc @dwarkesh_sp Meta uses TPU's as well, and OpenAI recently agreed to start utilizing them. You also have a variety of smaller deals with Salesforce and Samsung and Verizon, and smaller players themselves like Midjourney, Mistral, etc. Nvidia is still the big one, but TPUs are getting traction

@RihardJarc @dwarkesh_sp sure within google cloud they’ll use some TPUs. but that is largely driven by the lack of available nvidia compute. seen the market? there is no GPU compute avail. old GPUs running full tilt. this is fact left out of many of the arguments. demand is exploding and no capacity.

@RihardJarc @dwarkesh_sp That doesn't mean that there is no reason outside of capacity constraints to use them. Nvidia's chips are absolutely the most versatile, but some world class engineers have gone on record saying ASIC's can be more cost effective for very specific workloads

@RihardJarc @dwarkesh_sp ok. no argument. pretty sure nvidia grok deal plays to this. but very specific workloads etc.. its a huge market. did people expect nvidia to have 100% of it? it trades like a value stock is is growing like crazy. material positions in dozen future trillion companies.

If someone is whiling to give you billions, you accept the offer to give their asics a try. Jensen is being careful not to say things too blunt because they are his customers and showing respect. As for the debate on China, he is basically saying keep your friends close but enemies closer. There is a lot of wise reasons why this is important.

Jensen Huang just struggled to explain Nvidia's moat against Google's TPUs and Amazon's Trainium chips. His response: "Anthropic is a unique instance, not a trend." When the CEO of the most important chip company on earth dismisses a major customer's behavior as "unique," it usually means the behavior is exactly the trend he doesn't want to acknowledge. Here's the context. Anthropic runs Claude on three different types of compute: Nvidia GPUs, Google TPUs, and Amazon Trainium. The same models, the same workloads, running on fundamentally different chip architectures. This isn't hypothetical. It's production deployment. Anthropic serves Claude responses to millions of users across all three compute providers interchangeably. The reason this matters for Nvidia is that Anthropic proves the thing Jensen most needs to be false: that frontier AI models are portable across accelerator architectures. The investor thesis on Nvidia has always been that CUDA — Nvidia's software ecosystem — creates switching costs so high that customers are effectively locked in regardless of hardware performance. Anthropic demonstrates that assumption is wrong. You can, in fact, run the same model on different chips. It's operationally complex, but it's technically achievable. Jensen's framing that Anthropic is "unique" is partially accurate. Most AI labs don't have the engineering resources to deploy across multiple accelerator architectures. Anthropic has a specific incentive — they're deeply partnered with both Amazon (for Trainium via AWS) and Google (for TPUs via GCP), and both partnerships include preferential access to compute in exchange for using the providers' in-house chips. Anthropic gets compute they couldn't otherwise access. AWS and Google get proof points that their chips can handle frontier workloads. But the "unique" framing dodges the actual question. The question isn't whether Anthropic is unusual. The question is whether Anthropic's approach becomes the template for other AI labs. And the economic incentives for other labs to follow Anthropic's lead are enormous. Here's why. Nvidia's H100 GPUs currently cost approximately $25,000-$40,000 per unit. Google's TPU v5p generation costs Google itself far less to deploy internally, which translates to lower compute pricing for GCP customers. Amazon's Trainium chips are designed specifically for the cost structure needed to make AI inference economically viable at massive scale. The cost per token served on an ASIC optimized for AI workloads is structurally lower than the cost per token served on general-purpose GPUs. The reason is architectural. Nvidia GPUs are extraordinarily capable but general-purpose — they can run graphics rendering, scientific computing, cryptocurrency mining, and AI workloads. That flexibility comes with silicon area dedicated to capabilities AI inference doesn't need. ASICs designed specifically for AI can remove that unused silicon and optimize for exactly the operations frontier models perform most. The result: comparable performance at lower cost per operation. For frontier labs with massive compute bills, even a 20-30% cost reduction from running inference on ASICs represents hundreds of millions or billions of dollars in savings annually. OpenAI spends billions on Microsoft Azure compute. Meta spends comparable amounts on internal infrastructure. Anthropic's multi-accelerator deployment is generating real cost data that other labs are watching carefully. Jensen's response — "Anthropic is unique" — is technically true and strategically evasive. Anthropic is unique today. The pattern Anthropic is demonstrating — that frontier models can run on ASICs at meaningful scale — becomes the template for the industry if the cost savings prove sustainable. And the cost savings will prove sustainable because the underlying architectural advantage of ASICs for specific workloads is real. The broader threat to Nvidia's moat isn't that customers stop using GPUs entirely. It's that customers use GPUs for workloads that require general-purpose compute (training new models, research, experimentation) while using ASICs for workloads that are operationally well-understood (serving established models at scale). Training might remain GPU-dominated. Inference — which is the vast majority of total compute demand at scale — shifts toward ASICs. That bifurcation would be catastrophic for Nvidia's growth story. Training workloads are a fraction of total AI compute demand. Inference dominates. If inference migrates to specialized chips from Google, Amazon, and other providers developing their own AI silicon, Nvidia's addressable market shrinks even as overall AI compute demand grows. The company continues producing impressive training hardware while watching the much larger inference market capture value on competitor architectures. Google, Amazon, and Meta are all developing in-house AI chips specifically because they've run the math and concluded that not being dependent on Nvidia for inference compute is worth the massive investment required to develop competitive alternatives. The Anthropic deployment isn't unique. It's a leading indicator. The Dwarkesh interview caught Jensen in a moment of genuine difficulty because the question forced him to choose between acknowledging a legitimate threat and dismissing a proven counterexample. He chose dismissal. But the counterexample keeps running, successfully, in production, on the chips that aren't his. Nvidia's stock has been the single best large-cap investment of the past three years. Its moat has been assumed to be impregnable. Jensen's inability to cleanly articulate the moat against specific competitors suggests the moat is narrower than the valuation implies. The most dangerous moment for a dominant company isn't when competitors announce they're building alternatives. It's when the dominant CEO can't clearly explain why the alternatives don't work. Jensen just had that moment, on camera, in an interview most of the industry is watching. Anthropic is the unique instance today. If other labs follow the pattern — and the cost economics suggest many will — Anthropic becomes the leading indicator of what Nvidia's customer base looks like in 2028.

@dwarkesh_sp Anthropic uses TPUs to cut costs. Jensen says it's a special case. But cutting costs is what every AI lab will do. Not special - that is literally the trend.

@dwarkesh_sp yeah dwarkesh crushed it, jensen fumbled that one bad

@dwarkesh_sp Amazon and Google are invested heavily in Anthropic right? Combined about 25 to 30 percent ownership? Definitely the will provide these chips to Anthropic at cost. NVIDIA has 80% margins

@dwarkesh_sp @dwarkesh_sp rocks! really like him, keep it up buddy!

@dwarkesh_sp Yep, I don't usually watch youtubers or whatever you call them, but this dude generates insta respect 👌 And I agree, Jensen is in a very uncomfortable position answering that....

@dwarkesh_sp Agreed, @dwarkesh_sp is an excellent interviewer!

@dwarkesh_sp That’s cause he had to be diplomatic correct with…who’s that moron …oh yaa @dwarkesh_sp ….jensen could have told him off the frustration was real…very poorly done by patel!

@dwarkesh_sp Darkesh tandoori is a fakkot. Notice how he went in on Jensen, but other people he barely has the balls to ask anything important.

@dwarkesh_sp @dwarkesh_sp is one of the rare intellectually honest podcast interviewers

@dwarkesh_sp Jensen’s usual confidence seems to be trending down. Something I noticed in past earnings calls — he said, “You know, the benefits that AI produces for the world ultimately ‘has to’ generate revenues.” Maybe I’m reading too much, but I’d have expected him to say “will generate”.

@dwarkesh_sp 100% agree on @dwarkesh_sp. But gotta disagree on the first part. If you watch the full interview, it's the most convincing defense of $NVDA's moat I've seen from Jensen. Dwarkesh's high level also helped. (No $NVDA position btw)

@dwarkesh_sp Did we watch two different podcasts 😅😂

@dwarkesh_sp Point also is with TPUs, anthropic is still kicking ass and if anthropic is going to become Apple of AI, then Nvidia will get some hit

@dwarkesh_sp TPUs win on performance. The question is whether your scale justifies paying the ecosystem tax to get there. And Jensen kind of just dances around and misdirects since addressing it head on means admitting their only edge is "switching away from CUDA is expensive and annoying"

@dwarkesh_sp Did you listen to the same interview? You took Jensen’s answer completely out of context

@dwarkesh_sp NVDA cooked? Another Cisco in making? 🤔

not really, it says money comes with strings attached kinda also explains infra issues and unreliability codex usage is also seeing unprecedented growth. we are seeing claude-code moment, but st @thsottiaux does rate limit resets for us reliable infra and no capacity constrains i'm also thinking now if Dario had alternative motives now to campaign against nvidia sale of chips to china such a g8 scientist and leader, yet he forgot / failed to understand that, last time when android and app store was blocked, true alternative huawei became very viable not only in china but also in large parts of the world includiing the west, i clearly see it in germany also constrain creates crazy innovation, look into origins of mooncake (kv-cache) project, that alone is a huge blow to nvidia gpu requirements

@dwarkesh_sp Do you foresee $NVDA losing market share due to this?

@dwarkesh_sp The whole pie is going to grow much bigger but yeah I expect NVDA to have a smaller market can than the one they have today and the margins will also come down over the years.

@dwarkesh_sp Makes sense that Thiel sold his shares off then…

@dwarkesh_sp I’m not a fan of this pod. He’s a contentious and borderline disrespectful host. There were no issues with Jensen here.

@dwarkesh_sp

@dwarkesh_sp making @tylercowen proud

@dwarkesh_sp It's quite unique to see a 25 year old butt heads with one of the smartest and richest guys on the planet....all in a very rational discourse!

@dwarkesh_sp In order to be true to his viewers he’s going to have to consider that top guests won’t want to come on in future… it’s a balancing act

@dwarkesh_sp Maybe shut up And let him talk

@dwarkesh_sp Nvidia is a software company.

@dwarkesh_sp Nah only when you are talking to someone which doesn’t understand the tech and the industry, Jensen was doing this before this man even knew what a gpu was

@dwarkesh_sp And that is why Anthropic had that big problems serving their customers?