正在加载视频...

视频加载失败

返回首页

Compute Wars: OpenAI vs Anthopic. Why was Opus 4.5 such a breakthrough? Anthropic got lots more compute from AWS Madison and New Carlisle sites likely more than doubling their capacity. This got Anthropic got close to OpenAI's total capacity, and probably much higher effective capacity available for new model...

petergostev's profile picture

Peter Gostev

26,262 subscribers

158,528 次观看 • 6 个月前 •via X (Twitter)

科学技术
LIVE

Anya Rossi• Live Now

Private livecam show

24 条评论

Peter Gostev 的头像
Peter Gostev6 个月前

Interactive visualisation: Github:

Peter Gostev 的头像
Peter Gostev6 个月前

A lot of the starting point for data came from @EpochAIResearch's excellent resource with some additional research (mostly backward looking) and some extrapolations

Ganesh Kompella | Fractional CTO 的头像
Ganesh Kompella | Fractional CTO6 个月前

The compute race is the arms race nobody outside the industry tracks closely enough. Models don't get better because of one clever paper. They get better because someone got access to another 50,000 GPUs six months ago. Anthropic closing the compute gap with AWS explains the Opus jump better than any architectural innovation. Follow the GPUs, not the press releases.

CIPHER 的头像
CIPHER6 个月前

the compute race is the new arms race everyone talks about model architecture but infrastrcture decides who wins anthropic locking in aws capacity before openai could react was a chess move

Michel Justen 的头像
Michel Justen6 个月前

great visual, thanks for making this

𝘿𝙖𝙫𝙞𝙙 ✦ 𝙈𝙂𝙏 的头像
𝘿𝙖𝙫𝙞𝙙 ✦ 𝙈𝙂𝙏6 个月前

as an end user who runs opus 4.6 all day the compute constraints are obvious. rate limits hit hard even on max plan. but the quality when you have capacity is unmatched. its frustrating because the model is clearly capable of more, its just capacity gated

Kfir Gollan 的头像
Kfir Gollan6 个月前

@grok fact check this. Also, are there some public figures related to the distribution of compute usage? For example, let's say OpenAI allocates half of its resources to video models while Anthropic allocates 0, this will significantly change the status here

Max Ziebell 的头像
Max Ziebell6 个月前

Assuming they don’t run out of steam… or a breakthrough pushes acceptable consumer compute to edge devices.

Fernando Gonzalez 的头像
Fernando Gonzalez6 个月前

What assumptions for projection?

Jacquess Williams 的头像
Jacquess Williams6 个月前

More "compute" or a "faster" AI doesn't necessarily mean a "smarter" AI.

JC Christian 的头像
JC Christian6 个月前

the 6 month lag between getting capacity and shipping a model is the part people always forget when comparing these companies

Cassiopeia 的头像
Cassiopeia6 个月前

OpenAI 拥有强大的计算能力，能够利用包括 AWS 和 Cerebras 在内的多个服务器资源。 此外，还有 Epoch AI 未曾记录的其他 Stargate 站点。总体来说open ai拥有更多的算力

All Things Heroik 的头像
All Things Heroik6 个月前

Is this adjusted for them bailing on ram purchases?

AJ 的头像
AJ6 个月前

Great analysis. The compute gap matters for training, but the Stanford paper showing 6x performance gap from harness engineering alone suggests deployment-side architecture might matter even more than raw training FLOPS for end-user impact.

the meme jihad 的头像
the meme jihad6 个月前

Hate to break it to you cuz but nobody is getting any more compute while the entire oil infrastructure of the middle east is in peril.

Thomas Tao 的头像
Thomas Tao6 个月前

Yeah, and the hidden lag is systems work. More compute helps, but training stability and eval loops usually eat months first.

Sam Cohen 的头像
Sam Cohen6 个月前

How does Google compare over 26’ 27’?

StripSlashes 的头像
StripSlashes6 个月前

it was never a war

ReeseFang 的头像
ReeseFang6 个月前

compute is capex with better marketing. the underrated story is who converts FLOPS to capability most efficiently, not who has the most GPUs. Anthropic punching above weight on that ratio is the real signal here.

Alex at Libertify 📄➝ 🎥 的头像
Alex at Libertify 📄➝ 🎥6 个月前

The real advantage comes from how efficiently companies turn that compute into better training runs, faster iteration, and stronger products.

Adriana Sobota 的头像
Adriana Sobota6 个月前

The compute story explains so much of what happened in the last 6 months.

Joe Ward 的头像
Joe Ward6 个月前

Capacity is a good signal but if it were the dominate one then why isn’t gemini tops across the board? They arguably have more than both

Not Spacewear 的头像
Not Spacewear6 个月前

how does this compare to xAI. great chart

RelativelySmart 的头像
RelativelySmart6 个月前

They are just starting to train new models on Blackwell setups, most of what we see now was never trained initially on the latest series. Will likely see them in Q2 and Q3.

相关视频

Dylan on why Anthropic leading on models doesn't mean OpenAI loses. Even the tier 2 and 3 labs will sell out of tokens: "Everyone's like Anthropic won. They had Mythos, Opus 4.7 and their revenue is adding $10 billion a month all before OpenAI's alleged Spud release. So clearly Anthropic is in the lead and OpenAI is cooked. Anthropic has such bounds on compute, they can only grow it so fast. Dario used to gloat about how OpenAI was being too aggressive on compute and Anthropic was more sensible in their scaling. Now Anthropic is like I wish we had a lot more compute. OpenAI just raised a ton of money to get incremental compute in addition to the irresponsible levels of compute that they were buying. By the end of the year, let's say a 4.6 Opus tier model, the economy would spend $100 billion on. Anthropic won't have enough compute to do that. And presumably OpenAI and Google will hit that tier soon. Anthropic may get to charge 70%+ gross margins but if OpenAI hits it next, they charge 50% gross margins. They still get all of this incremental demand. There is such demand for these tokens and such limitations on compute. It's pretty clear even the tier 2 or tier 3 lab are going to be sold out of tokens. The economic value that the best model can deliver is growing faster than our ability to actually serve those tokens to people via the infrastructure."
2:33

Dylan on why Anthropic leading on models doesn't mean OpenAI loses. Even the tier 2 and 3 labs will sell out of tokens: "Everyone's like Anthropic won. They had Mythos, Opus 4.7 and their revenue is adding $10 billion a month all before OpenAI's alleged Spud release. So clearly Anthropic is in the lead and OpenAI is cooked. Anthropic has such bounds on compute, they can only grow it so fast. Dario used to gloat about how OpenAI was being too aggressive on compute and Anthropic was more sensible in their scaling. Now Anthropic is like I wish we had a lot more compute. OpenAI just raised a ton of money to get incremental compute in addition to the irresponsible levels of compute that they were buying. By the end of the year, let's say a 4.6 Opus tier model, the economy would spend $100 billion on. Anthropic won't have enough compute to do that. And presumably OpenAI and Google will hit that tier soon. Anthropic may get to charge 70%+ gross margins but if OpenAI hits it next, they charge 50% gross margins. They still get all of this incremental demand. There is such demand for these tokens and such limitations on compute. It's pretty clear even the tier 2 or tier 3 lab are going to be sold out of tokens. The economic value that the best model can deliver is growing faster than our ability to actually serve those tokens to people via the infrastructure."

Patrick OShaughnessy

295,307 次观看 • 5 个月前

SF Compute CEO evan conrad on his contrarian thesis that OpenAI and Anthropic won't eat the world, everyone will become OpenAI and Anthropic instead: "Our impression is not that OpenAI and Anthropic will eat the world, it's that everyone will look like OpenAI and Anthropic. That's simply the way that companies work now." "Every single company in the world that is of meaningful scale will look like a frontier lab. There is this belief that the large labs have won it all, and they've got this compounding advantage. The default assumption of that is assuming nobody else will do anything." "The bottleneck was there's 100 people, and they know how to train the models. Or you thought, the problem is compute. The problem of compute is credit. And OpenAI and Anthropic are not unique to having credit. In fact, they don't have as good credit as other people." "If you assume the problem is not the researchers, that that problem's basically gone away the moment you got to auto research, then what you should assume is that pretty much every other company is going to swap the way they work to look like a frontier lab. It won't be OpenAI and Anthropic take the whole world. It's that everyone's OpenAI and Anthropic now." San Francisco Compute
1:32

SF Compute CEO evan conrad on his contrarian thesis that OpenAI and Anthropic won't eat the world, everyone will become OpenAI and Anthropic instead: "Our impression is not that OpenAI and Anthropic will eat the world, it's that everyone will look like OpenAI and Anthropic. That's simply the way that companies work now." "Every single company in the world that is of meaningful scale will look like a frontier lab. There is this belief that the large labs have won it all, and they've got this compounding advantage. The default assumption of that is assuming nobody else will do anything." "The bottleneck was there's 100 people, and they know how to train the models. Or you thought, the problem is compute. The problem of compute is credit. And OpenAI and Anthropic are not unique to having credit. In fact, they don't have as good credit as other people." "If you assume the problem is not the researchers, that that problem's basically gone away the moment you got to auto research, then what you should assume is that pretty much every other company is going to swap the way they work to look like a frontier lab. It won't be OpenAI and Anthropic take the whole world. It's that everyone's OpenAI and Anthropic now." San Francisco Compute

MTS

135,154 次观看 • 1 个月前

.Dylan Patel thinks Anthropic will get to 5–6 GW by year end. But at a far higher cost than it would have had to pay if it had gone crazy on compute commitments early like OpenAI did. Ant will get the compute either through cloud partners like Bedrock and Vertex who take ~50% gross margin on revenue, or shorter-term deals where rates have climbed well above what long-term contracts locked in a year or two ago cost. They’ll probably also have to destroy demand by pushing up model prices.
2:16

.Dylan Patel thinks Anthropic will get to 5–6 GW by year end. But at a far higher cost than it would have had to pay if it had gone crazy on compute commitments early like OpenAI did. Ant will get the compute either through cloud partners like Bedrock and Vertex who take ~50% gross margin on revenue, or shorter-term deals where rates have climbed well above what long-term contracts locked in a year or two ago cost. They’ll probably also have to destroy demand by pushing up model prices.

Dwarkesh Patel

55,447 次观看 • 6 个月前

Baseten's Charlie O'Neill on why the gap to the frontier is just compute and data: "Despite OpenAI and Anthropic and the closed source labs having more compute and more data, the recipe is the same. We're all scaling the same recipe here, and we're at slightly different points of the scaling curve. It's not like OpenAI or Anthropic have any secret sauce that differentiates them scaling up their multi-trillion parameter models from the rest of the open source ecosystem." "It really is just a matter of taking the core underlying architecture and training recipe, and throwing more compute and data at it, which is really promising." "Of course, these closed source labs probably have a long tail of optimizations they've made, which we're gonna continue to try and figure out in the open source community. But the core mechanism is there." "K3 is a great example of, well, if you do scale this recipe up, we're gonna get a pretty good model that's very, very close to the frontier, if not frontier itself." Charlie O'Neill Baseten
1:51

Baseten's Charlie O'Neill on why the gap to the frontier is just compute and data: "Despite OpenAI and Anthropic and the closed source labs having more compute and more data, the recipe is the same. We're all scaling the same recipe here, and we're at slightly different points of the scaling curve. It's not like OpenAI or Anthropic have any secret sauce that differentiates them scaling up their multi-trillion parameter models from the rest of the open source ecosystem." "It really is just a matter of taking the core underlying architecture and training recipe, and throwing more compute and data at it, which is really promising." "Of course, these closed source labs probably have a long tail of optimizations they've made, which we're gonna continue to try and figure out in the open source community. But the core mechanism is there." "K3 is a great example of, well, if you do scale this recipe up, we're gonna get a pretty good model that's very, very close to the frontier, if not frontier itself." Charlie O'Neill Baseten

MTS

17,404 次观看 • 2 个月前

Chamath Palihapitiya, one of the most connected investors in tech and his warning is the clearest framing of the AI compute crisis anyone has put into words. "It is a five alarm fire for them. They need to have land, power, shell." He's talking about Anthropic and OpenAI and the threat he's describing is called the Friendster effect. Friendster was the dominant social network before MySpace and Facebook and it didn't lose because it had a bad product but rather it lost because it couldn't keep the site up. Demand outpaced infrastructure, the experience degraded, and users left for one that actually worked. Chamath's argument is that OpenAI and Anthropic are approaching exactly that moment. The numbers are already showing it, Anthropic is growing so fast that it had to cut Claude's thinking depth during peak hours, cap agentic sessions, and test removing Claude Code from its $20 plan entirely. GitHub Copilot paused new signups, paying enterprise customers are hitting usage walls they've never seen before. Dario Amodei himself admitted there is "no hedge on earth" against the risk of over-purchasing compute meaning he's deliberately staying lean on capacity, even as the demand wall approaches. The core problem is structural, OpenAI and Anthropic grew up renting capacity from hyperscalers AWS, Azure, Google Cloud. That was fine when they were small but now they're so large that dependency is a strategic liability. Every token they sell runs on someone else's infrastructure and every capacity decision belongs to someone else. And when demand spikes faster than anyone planned, there's nothing they can do in real time except throttle. Building your own infrastructure takes 18 to 24 months minimum, you need to acquire land, secure power, construct shell and none of that happens fast. That's why Google's $40 billion commitment to Anthropic this week is about more than just money but rather about securing the land, power, and infrastructure that Anthropic needs to not become Friendster. Whoever controls the compute controls the frontier and right now, the AI labs with the best products are the most dependent on infrastructure they don't own.
0:31

Chamath Palihapitiya, one of the most connected investors in tech and his warning is the clearest framing of the AI compute crisis anyone has put into words. "It is a five alarm fire for them. They need to have land, power, shell." He's talking about Anthropic and OpenAI and the threat he's describing is called the Friendster effect. Friendster was the dominant social network before MySpace and Facebook and it didn't lose because it had a bad product but rather it lost because it couldn't keep the site up. Demand outpaced infrastructure, the experience degraded, and users left for one that actually worked. Chamath's argument is that OpenAI and Anthropic are approaching exactly that moment. The numbers are already showing it, Anthropic is growing so fast that it had to cut Claude's thinking depth during peak hours, cap agentic sessions, and test removing Claude Code from its $20 plan entirely. GitHub Copilot paused new signups, paying enterprise customers are hitting usage walls they've never seen before. Dario Amodei himself admitted there is "no hedge on earth" against the risk of over-purchasing compute meaning he's deliberately staying lean on capacity, even as the demand wall approaches. The core problem is structural, OpenAI and Anthropic grew up renting capacity from hyperscalers AWS, Azure, Google Cloud. That was fine when they were small but now they're so large that dependency is a strategic liability. Every token they sell runs on someone else's infrastructure and every capacity decision belongs to someone else. And when demand spikes faster than anyone planned, there's nothing they can do in real time except throttle. Building your own infrastructure takes 18 to 24 months minimum, you need to acquire land, secure power, construct shell and none of that happens fast. That's why Google's $40 billion commitment to Anthropic this week is about more than just money but rather about securing the land, power, and infrastructure that Anthropic needs to not become Friendster. Whoever controls the compute controls the frontier and right now, the AI labs with the best products are the most dependent on infrastructure they don't own.

Milk Road AI

260,300 次观看 • 5 个月前

Brad Gerstner, running the numbers on Elon's next 6 gigawatts of compute: "it's $50 billion per gigawatt to build minimum - so you're $300 billion" This is him explaining why the frontier labs will overpay 3-5X the market to get compute, why almost nobody but Elon can stand one of these up in a year, and why he thinks the whole thing turns violent the moment demand slips Tthey're willing to pay 3, 4, 5X market pricing in order to get at scale compute" "Probably only Elon can actually stand up that much in that time frame - like Jensen said to me on the pod, it's like nobody comes close - Microsoft doesn't come close - Google doesn't come close" "This year Anthropic and OpenAI combined, their starting total compute was like 5 gigawatts - so he's talking about incrementally adding more than they had as combined companies" "You're effectively backstopping people's ability to build ahead of their revenue - so it becomes much more violent if you ever see demand slippage" "Credit spreads are blowing - have continued to stay wide on these deals" Bookmark & watch the full conversation ↓
5:32

Brad Gerstner, running the numbers on Elon's next 6 gigawatts of compute: "it's $50 billion per gigawatt to build minimum - so you're $300 billion" This is him explaining why the frontier labs will overpay 3-5X the market to get compute, why almost nobody but Elon can stand one of these up in a year, and why he thinks the whole thing turns violent the moment demand slips Tthey're willing to pay 3, 4, 5X market pricing in order to get at scale compute" "Probably only Elon can actually stand up that much in that time frame - like Jensen said to me on the pod, it's like nobody comes close - Microsoft doesn't come close - Google doesn't come close" "This year Anthropic and OpenAI combined, their starting total compute was like 5 gigawatts - so he's talking about incrementally adding more than they had as combined companies" "You're effectively backstopping people's ability to build ahead of their revenue - so it becomes much more violent if you ever see demand slippage" "Credit spreads are blowing - have continued to stay wide on these deals" Bookmark & watch the full conversation ↓

Fireside Alpha

106,612 次观看 • 1 个月前

Here’s my BBC interview: I went over how the OpenAI deal validates $AMZN AWS GenAI capabilities & shared my takeaway from the $PLTR thesis post-earnings. Every story you see about $GOOGL, $MSFT, $META, $ORCL or OpenAI building new compute factories connects directly to why Palantir can report a Rule of 40 score of 114. The trillions in AI infrastructure CapEx are creating raw capacity but capacity alone doesn’t create productivity and Palantir sits exactly at that conversion point. So yet it trades rich but it’s the only pure play with leverage to every dollar the hyperscalers spend. You’re not paying for today’s revenue but you’re paying for the right to own the switchboard that will orchestrate trillions in future AI infrastructure.
3:13

Here’s my BBC interview: I went over how the OpenAI deal validates $AMZN AWS GenAI capabilities & shared my takeaway from the $PLTR thesis post-earnings. Every story you see about $GOOGL, $MSFT, $META, $ORCL or OpenAI building new compute factories connects directly to why Palantir can report a Rule of 40 score of 114. The trillions in AI infrastructure CapEx are creating raw capacity but capacity alone doesn’t create productivity and Palantir sits exactly at that conversion point. So yet it trades rich but it’s the only pure play with leverage to every dollar the hyperscalers spend. You’re not paying for today’s revenue but you’re paying for the right to own the switchboard that will orchestrate trillions in future AI infrastructure.

Shay Boloor

62,985 次观看 • 11 个月前

🚨 Sam Altman: there's a very important thing happening. I did not expect 3 or 6 months ago to be at this point we're at now” We now have a confirmed reason why OpenAI pulled out of Sora and it was as most of us expected - because they needed to concentrate all of their compute and product capacity into their next generation of models and automated agents. Sam says: "We have a few times in our history realized something really important is working or about to work so well that we have to stop a bunch of other projects. In fact, this was the original thing happened with GPT-3. We had a whole portfolio of bets at the time... we shut down many projects that were working well like robotics... so that we could concentrate our compute, our researchers, our effort into this thing that we said, okay, there's a very important thing happening." "I did not expect 3 or 6 months ago to be at this point we're at now, where something very big and important is about to happen again with this next generation of models and the agents they can power." "We need to concentrate our compute and our product capacity into these next generation of automated researchers and companies... the note under the note is it's about compute. It's always about compute."
1:13

🚨 Sam Altman: there's a very important thing happening. I did not expect 3 or 6 months ago to be at this point we're at now” We now have a confirmed reason why OpenAI pulled out of Sora and it was as most of us expected - because they needed to concentrate all of their compute and product capacity into their next generation of models and automated agents. Sam says: "We have a few times in our history realized something really important is working or about to work so well that we have to stop a bunch of other projects. In fact, this was the original thing happened with GPT-3. We had a whole portfolio of bets at the time... we shut down many projects that were working well like robotics... so that we could concentrate our compute, our researchers, our effort into this thing that we said, okay, there's a very important thing happening." "I did not expect 3 or 6 months ago to be at this point we're at now, where something very big and important is about to happen again with this next generation of models and the agents they can power." "We need to concentrate our compute and our product capacity into these next generation of automated researchers and companies... the note under the note is it's about compute. It's always about compute."

Chris

243,817 次观看 • 6 个月前

GPT-6 Astra just destroyed its brother ChatGPT Pro for $200/month On September 10, OpenAI quietly closed new sign-ups for ChatGPT Pro at $200 a month. The reason wasn't price resistance. It was the opposite. Too many people wanted in, running the new Astra model, and that plan pulls more compute than the system could handle. If you already pay $200, nothing changes. Your access stays exactly the same. The $100 plan is still open to anyone who wants it. Only the top tier's front door just closed. Here's the part people miss. The $200 plan was built for the heaviest users on the platform, people who burn through far more than $200 worth of compute every single month. OpenAI wasn't protecting its margins. It was protecting the users already eating that capacity.
0:23

GPT-6 Astra just destroyed its brother ChatGPT Pro for $200/month On September 10, OpenAI quietly closed new sign-ups for ChatGPT Pro at $200 a month. The reason wasn't price resistance. It was the opposite. Too many people wanted in, running the new Astra model, and that plan pulls more compute than the system could handle. If you already pay $200, nothing changes. Your access stays exactly the same. The $100 plan is still open to anyone who wants it. Only the top tier's front door just closed. Here's the part people miss. The $200 plan was built for the heaviest users on the platform, people who burn through far more than $200 worth of compute every single month. OpenAI wasn't protecting its margins. It was protecting the users already eating that capacity.

Zentrix⌚️

42,492 次观看 • 18 天前

There's an economics theorem called Alchian-Allen. And it has the very interesting implication that AI labs will be able to charge *higher* margins on their best models as compute gets scarcer. As compute gets more expensive, the cost of running any model goes up. So you might as well pay a bit more to make sure you're running the very best model. Which means the economics of being at the frontier improve, because if you’re not running the very best model, then you’re underutilizing this very precious compute. This pushes the AI model market towards winner-take-all; if you're the best, you can get away with charging an even higher margin. Dylan Patel tells me that we’re already seeing this today: all the revenue in the industry is on the best models. That’s the Alchian-Allen effect. If there’s a cost increase that’s roughly the same for all products, then the relative difference in price between higher and lower quality goods actually goes down. Consumers become relatively more willing to pay for the premium product. And it means that as the compute shortage hits, AI labs can capture more margin - not less, as you might expect - because consumers are choosing premium models more often.
1:21

There's an economics theorem called Alchian-Allen. And it has the very interesting implication that AI labs will be able to charge *higher* margins on their best models as compute gets scarcer. As compute gets more expensive, the cost of running any model goes up. So you might as well pay a bit more to make sure you're running the very best model. Which means the economics of being at the frontier improve, because if you’re not running the very best model, then you’re underutilizing this very precious compute. This pushes the AI model market towards winner-take-all; if you're the best, you can get away with charging an even higher margin. Dylan Patel tells me that we’re already seeing this today: all the revenue in the industry is on the best models. That’s the Alchian-Allen effect. If there’s a cost increase that’s roughly the same for all products, then the relative difference in price between higher and lower quality goods actually goes down. Consumers become relatively more willing to pay for the premium product. And it means that as the compute shortage hits, AI labs can capture more margin - not less, as you might expect - because consumers are choosing premium models more often.

Dwarkesh Patel

323,500 次观看 • 6 个月前

$NBIS is showing just how tight AI compute still is as its first Blackwell capacity auction cleared 15% above the company’s previous record price and management says it could already sell out planned 2027 capacity today. That scarcity is translating directly into pricing power with long-term deals yielding ~$25M per MW while 6-month capacity is clearing at ~$50M per MW. That pricing power is now feeding directly into the model with Q2 revenue up 454% YoY, ARR reaching $3B backlog topping $40B and AI cloud margins near 50% while management raises contracted power to more than 5 GW and targets over 1 GW of new capacity annually from 2027.
1:02

$NBIS is showing just how tight AI compute still is as its first Blackwell capacity auction cleared 15% above the company’s previous record price and management says it could already sell out planned 2027 capacity today. That scarcity is translating directly into pricing power with long-term deals yielding ~$25M per MW while 6-month capacity is clearing at ~$50M per MW. That pricing power is now feeding directly into the model with Q2 revenue up 454% YoY, ARR reaching $3B backlog topping $40B and AI cloud margins near 50% while management raises contracted power to more than 5 GW and targets over 1 GW of new capacity annually from 2027.

Shay Boloor

68,480 次观看 • 1 个月前

SITUATION EXPLAINED: Dwarkesh argues compute could get 10X more expensive. • Anthropic's revenue has been 10X-ing year over year while lab compute only 3X's • Three ways that gap can close: margins rise, compute gets more expensive, or labs shift compute to inference • All three are already happening, Anthropic went from 40% margins in 2025 to possibly 80%+ this year • But labs don't want the third one, heavy inference spend signals AI progress has stalled and you're now a cloud provider • Spot compute prices are up 40%+ since February, and labs pay well above spot for security and scale • Google is reportedly paying SpaceXAI $900 million a month for 110K GPUs, roughly 2X spot • Key claim: if a human-level software engineer ran on an H100, that H100 should rent for $250K a year, 15X today's price • The 3X annual compute growth is 1.4X Moore's Law, 1.2X new fabs, and 1.8X from AI taking wafer allocation from other devices • The fab piece is bottlenecked by EUV tool supply through 2030, and the wafer piece hits a wall by end of 2027 • If compute stays scarce, new labs need far more capital just to reach the same starting line • Compute gets cheap again only once robots can turn sand and copper into computers, which is gated on robotics, not RSI sof 𓋹: "This is what I'm most concerned about, a lack of innovation, not within companies, but a lack of new companies that can actually do something substantially new, because of the scarcity of compute." Theo Jaffee: "The biggest companies in the world by revenue are Amazon and Walmart, at 743 billion and 725 billion. If Anthropic makes 100 billion by the end of this year, that puts them at Target. To go from Target to bigger than Walmart in a year would be very impressive."
16:11

SITUATION EXPLAINED: Dwarkesh argues compute could get 10X more expensive. • Anthropic's revenue has been 10X-ing year over year while lab compute only 3X's • Three ways that gap can close: margins rise, compute gets more expensive, or labs shift compute to inference • All three are already happening, Anthropic went from 40% margins in 2025 to possibly 80%+ this year • But labs don't want the third one, heavy inference spend signals AI progress has stalled and you're now a cloud provider • Spot compute prices are up 40%+ since February, and labs pay well above spot for security and scale • Google is reportedly paying SpaceXAI $900 million a month for 110K GPUs, roughly 2X spot • Key claim: if a human-level software engineer ran on an H100, that H100 should rent for $250K a year, 15X today's price • The 3X annual compute growth is 1.4X Moore's Law, 1.2X new fabs, and 1.8X from AI taking wafer allocation from other devices • The fab piece is bottlenecked by EUV tool supply through 2030, and the wafer piece hits a wall by end of 2027 • If compute stays scarce, new labs need far more capital just to reach the same starting line • Compute gets cheap again only once robots can turn sand and copper into computers, which is gated on robotics, not RSI sof 𓋹: "This is what I'm most concerned about, a lack of innovation, not within companies, but a lack of new companies that can actually do something substantially new, because of the scarcity of compute." Theo Jaffee: "The biggest companies in the world by revenue are Amazon and Walmart, at 743 billion and 725 billion. If Anthropic makes 100 billion by the end of this year, that puts them at Target. To go from Target to bigger than Walmart in a year would be very impressive."

MTS

10,617 次观看 • 2 个月前

🚨 JUST NOW: JENSEN HINTED ANTHROPIC’S NEXT DEAL AND ITS LIKELY $NBIS On the $NVDA earnings call, Jensen said they’ve secured compute for Anthropic across Azure, AWS and $CRWV then added there’s a “whole list of others” they are bringing online. $NBIS is an NVIDIA Cloud Partner with early access to Vera Rubin, already deploying NVL72 across the US and Europe from H2 2026. The partnerships are in place. The infrastructure is being built. Expect a deal soon. If Jensen is flagging more capacity providers coming online for Anthropic, Nebius has everything in place to be that call.
0:48

🚨 JUST NOW: JENSEN HINTED ANTHROPIC’S NEXT DEAL AND ITS LIKELY $NBIS On the $NVDA earnings call, Jensen said they’ve secured compute for Anthropic across Azure, AWS and $CRWV then added there’s a “whole list of others” they are bringing online. $NBIS is an NVIDIA Cloud Partner with early access to Vera Rubin, already deploying NVL72 across the US and Europe from H2 2026. The partnerships are in place. The infrastructure is being built. Expect a deal soon. If Jensen is flagging more capacity providers coming online for Anthropic, Nebius has everything in place to be that call.

Jordan

159,593 次观看 • 4 个月前

Here's my full TBPN interview from yesterday, where I explain why AI chip stocks are going to soar over the next 6–9 months due to the imminent RSI advances from OpenAI and Anthropic. RSI will drive a dramatic increase in AI model capabilities, which will consume exponentially more compute. Much of the recent FUD handwringing is misplaced, just like DeepSeek last year, especially when we're about to go vertical from RSI. Note: I was last on TBPN on March 30th, where I pitched CPU and HBM memory stocks on the agentic AI megatrend. They more than doubled in the months that followed.
31:20

Here's my full TBPN interview from yesterday, where I explain why AI chip stocks are going to soar over the next 6–9 months due to the imminent RSI advances from OpenAI and Anthropic. RSI will drive a dramatic increase in AI model capabilities, which will consume exponentially more compute. Much of the recent FUD handwringing is misplaced, just like DeepSeek last year, especially when we're about to go vertical from RSI. Note: I was last on TBPN on March 30th, where I pitched CPU and HBM memory stocks on the agentic AI megatrend. They more than doubled in the months that followed.

tae kim

47,184 次观看 • 2 个月前

Key fact claims for anyone watching the AI boom/bubble! They may contradict the Bridgewater AI slides I posted recently. Ed Zitron According to Dylan, both OpenAI and Anthropic have shifted from being venture-funded loss-making companies to generating massive positive gross margins (01:39 - 02:37). Key points regarding their profitability per gigawatt include: Revenue vs. Cost: While the base cost of compute is roughly $10 million to $15 million per megawatt, these labs have significantly increased their ability to monetize that capacity (02:29 - 03:00). Margin Improvement: OpenAI’s newer models (like 5.6) and Anthropic’s models (such as Opus 5 or Mythos) have pushed revenue generation well beyond the incremental cost of compute (02:54 - 03:02). Monetization Scaling: In the case of Anthropic, revenue has reached as high as $50 million per megawatt, allowing them to take the profit from that inference capacity and reinvest it heavily into training (03:07 - 03:23). Listen further for incredible projections: OAI + Ant will control most of the computing power on Earth within next few years?
1:46

Key fact claims for anyone watching the AI boom/bubble! They may contradict the Bridgewater AI slides I posted recently. Ed Zitron According to Dylan, both OpenAI and Anthropic have shifted from being venture-funded loss-making companies to generating massive positive gross margins (01:39 - 02:37). Key points regarding their profitability per gigawatt include: Revenue vs. Cost: While the base cost of compute is roughly $10 million to $15 million per megawatt, these labs have significantly increased their ability to monetize that capacity (02:29 - 03:00). Margin Improvement: OpenAI’s newer models (like 5.6) and Anthropic’s models (such as Opus 5 or Mythos) have pushed revenue generation well beyond the incremental cost of compute (02:54 - 03:02). Monetization Scaling: In the case of Anthropic, revenue has reached as high as $50 million per megawatt, allowing them to take the profit from that inference capacity and reinvest it heavily into training (03:07 - 03:23). Listen further for incredible projections: OAI + Ant will control most of the computing power on Earth within next few years?

steve hsu

14,230 次观看 • 1 个月前

Software may be starting to show early signs of life but I still think the “hard” side of AI has the cleaner lane for 2026 and probably most of 2027. We're nowhere close to done with the physical AI buildout since compute demand is still ahead of supply, power is becoming the bottleneck and every layer of AI infrastructure still needs more capacity. I own some software names but I still see this as more of an oversold bounce than a true leadership rotation away from the companies building the AI infrastructure layer. Software probably becomes a cleaner leadership group once inference costs fall, enterprise AI budgets expand and AI features start flowing through margins but that still feels more like a 2027 story to me.
0:29

Software may be starting to show early signs of life but I still think the “hard” side of AI has the cleaner lane for 2026 and probably most of 2027. We're nowhere close to done with the physical AI buildout since compute demand is still ahead of supply, power is becoming the bottleneck and every layer of AI infrastructure still needs more capacity. I own some software names but I still see this as more of an oversold bounce than a true leadership rotation away from the companies building the AI infrastructure layer. Software probably becomes a cleaner leadership group once inference costs fall, enterprise AI budgets expand and AI features start flowing through margins but that still feels more like a 2027 story to me.

Shay Boloor

122,716 次观看 • 4 个月前

“Elon and Dario should do a deal tomorrow.” - Chamath explains how power constraints are giving Elon massive leverage to make AI deals “ To the extent that OpenAI missed, I think what that is, is an insight to not enough compute capacity today, and that problem is only getting worse. If you look at the actual amount of gigawatts that are under construction, we have a huge mismatch now. People have announced all these projects, but less than half of it is actually being built. Less than half. Most of it is stuck in red tape. There's no credible strategy to turn any of this stuff on. Who will this hurt? It will hurt Anthropic and OpenAI the most. Who will this benefit? It will benefit the hyperscalers, specifically Oracle, Amazon, Meta, Microsoft, and Google. And now what you're going to see is a negotiation and a trade back and forth. How much equity do I have to give up? How much control do I have to give up to get access to the compute? How badly will I miss my growth forecasts if I don't? That's a huge lane for Grok and SpaceX to run through because they have a ton of excess capacity. And so I think the Cursor deal was the appetizer. If I were Elon now, I'd be running all over this market because if the models catch up in quality, I think he could also do something really crazy with Anthropic. He and Dario should do a deal tomorrow.”
1:20

“Elon and Dario should do a deal tomorrow.” - Chamath explains how power constraints are giving Elon massive leverage to make AI deals “ To the extent that OpenAI missed, I think what that is, is an insight to not enough compute capacity today, and that problem is only getting worse. If you look at the actual amount of gigawatts that are under construction, we have a huge mismatch now. People have announced all these projects, but less than half of it is actually being built. Less than half. Most of it is stuck in red tape. There's no credible strategy to turn any of this stuff on. Who will this hurt? It will hurt Anthropic and OpenAI the most. Who will this benefit? It will benefit the hyperscalers, specifically Oracle, Amazon, Meta, Microsoft, and Google. And now what you're going to see is a negotiation and a trade back and forth. How much equity do I have to give up? How much control do I have to give up to get access to the compute? How badly will I miss my growth forecasts if I don't? That's a huge lane for Grok and SpaceX to run through because they have a ton of excess capacity. And so I think the Cursor deal was the appetizer. If I were Elon now, I'd be running all over this market because if the models catch up in quality, I think he could also do something really crazy with Anthropic. He and Dario should do a deal tomorrow.”

The All-In Podcast

389,819 次观看 • 5 个月前

Why did Anthropic hold back Mythos? David Sacks: “ Marc Andreessen pointed out that one of the reasons why Anthropic might've wanted to hold back Mythos, is they simply didn't have the compute to serve it. The model was huge and very expensive to serve, something like 10 or 20x the token cost of, say, Opus. They knew Opus 4.7 was coming out, right? So they hold it back knowing that they don't have the compute to serve it anyway, and they save their compute for the next iteration of Opus. And then by holding it back, they create this impression of scarcity and altruism, and it turns into this gigantic marketing event for their product, because everyone in the government's like, 'Oh wow, they're holding it back because it's so amazing.' Now look, I think it may have been genuinely altruistic as well in the sense that Mythos does reveal coding vulnerabilities that people didn't know about before. And it does make sense to give time to companies with large code bases to patch these dormant bugs and vulnerabilities. But, it's looking more and more like Anthropic could not have offered that model commercially anyway, because it was just too big and expensive, and they needed to create space for Opus 4.7. So it's an interesting theory on what actually happened there.”
1:03

Why did Anthropic hold back Mythos? David Sacks: “ Marc Andreessen pointed out that one of the reasons why Anthropic might've wanted to hold back Mythos, is they simply didn't have the compute to serve it. The model was huge and very expensive to serve, something like 10 or 20x the token cost of, say, Opus. They knew Opus 4.7 was coming out, right? So they hold it back knowing that they don't have the compute to serve it anyway, and they save their compute for the next iteration of Opus. And then by holding it back, they create this impression of scarcity and altruism, and it turns into this gigantic marketing event for their product, because everyone in the government's like, 'Oh wow, they're holding it back because it's so amazing.' Now look, I think it may have been genuinely altruistic as well in the sense that Mythos does reveal coding vulnerabilities that people didn't know about before. And it does make sense to give time to companies with large code bases to patch these dormant bugs and vulnerabilities. But, it's looking more and more like Anthropic could not have offered that model commercially anyway, because it was just too big and expensive, and they needed to create space for Opus 4.7. So it's an interesting theory on what actually happened there.”

The All-In Podcast

170,866 次观看 • 5 个月前

I think I figured out how SpaceX / Elon Musk are going to pay for the ~8GW of DC capacity SA has forecasted them to deploy next year… as Brad Gerstner mentioned in this clip, $NVDA shareholders do not want NVIDIA backstopping $100’s of billions of dollars worth of compute- it will terrify the market AND, it seems highly unlikely that the Hypers will pay directly for that much compute either since it will balloon their spending next year + at the economics needed here, it would all be going to Frontier labs (hurting the diversification strategy they all are trying to pursue) So… what I think is likely to happen: Brad Gerstner laid out that in the past, the financial structure of compute infra was you get 25% payback per year, you recoup the infra cost by end of year 4, then year 5 profit & year 6 is the kicker to bump IRR up nicely That’s now changed as the demand has become so intense at the Frontier labs that the payback period on infra has compressed maybe into <1yr basically, this $300-$400B in CapEx is going to show up in 3-6 month term lease agreements between SpaceX & Anthropic / OpenAI with enough margin to pay off the compute portion entirely (or vast majority) over the lifespan of the agreement money will trade hands basically up front from ANT/OAI to SpaceX then to NVIDIA. NVIDIA will be paid in full for the GPUs up front and they’ll commit to supporting Elon / SpaceX to stand up the compute in the timeframe specified (SpaceX will ultimately be on the hook for timelines) The kicker would be if NVIDIA captures any durable rent here for facilitating this… a perpetual dividend or something like that
3:36

I think I figured out how SpaceX / Elon Musk are going to pay for the ~8GW of DC capacity SA has forecasted them to deploy next year… as Brad Gerstner mentioned in this clip, $NVDA shareholders do not want NVIDIA backstopping $100’s of billions of dollars worth of compute- it will terrify the market AND, it seems highly unlikely that the Hypers will pay directly for that much compute either since it will balloon their spending next year + at the economics needed here, it would all be going to Frontier labs (hurting the diversification strategy they all are trying to pursue) So… what I think is likely to happen: Brad Gerstner laid out that in the past, the financial structure of compute infra was you get 25% payback per year, you recoup the infra cost by end of year 4, then year 5 profit & year 6 is the kicker to bump IRR up nicely That’s now changed as the demand has become so intense at the Frontier labs that the payback period on infra has compressed maybe into <1yr basically, this $300-$400B in CapEx is going to show up in 3-6 month term lease agreements between SpaceX & Anthropic / OpenAI with enough margin to pay off the compute portion entirely (or vast majority) over the lifespan of the agreement money will trade hands basically up front from ANT/OAI to SpaceX then to NVIDIA. NVIDIA will be paid in full for the GPUs up front and they’ll commit to supporting Elon / SpaceX to stand up the compute in the timeframe specified (SpaceX will ultimately be on the hook for timelines) The kicker would be if NVIDIA captures any durable rent here for facilitating this… a perpetual dividend or something like that

Nick Dorsey

113,333 次观看 • 1 个月前

Elon Musk says Anthropic puts more care into AI safety than OpenAI and explains how rival companies should test one another’s models. “All the AI companies have a test harness, a series of tests that you give to any model to see if it’s going to build bioweapons or nuclear bombs or be deliberately deceptive. Everyone applying everyone else’s test harness is probably the best thing we can do to ensure our safety. We should try to do that as soon as possible.” “It’s the only thing I can think of that we could probably get all parties, including China, to agree to. You basically provide API access in advance of the model release. If they don’t solve what’s problematic, competitors can go public with the fact that they think the model being released is unsafe. The legal liability would be enormous.” “The two leading AI companies are Anthropic and OpenAI. Their models are quite close in capability, so it’s difficult for either one to slow down without handing the lead to the other. On balance, I think Anthropic puts more care into their safety than OpenAI. But even Anthropic acknowledges they are worried about their models. Many people from Anthropic have publicly voiced concern that their models are scaring them, that they’re getting scary smart.”
8:41

Elon Musk says Anthropic puts more care into AI safety than OpenAI and explains how rival companies should test one another’s models. “All the AI companies have a test harness, a series of tests that you give to any model to see if it’s going to build bioweapons or nuclear bombs or be deliberately deceptive. Everyone applying everyone else’s test harness is probably the best thing we can do to ensure our safety. We should try to do that as soon as possible.” “It’s the only thing I can think of that we could probably get all parties, including China, to agree to. You basically provide API access in advance of the model release. If they don’t solve what’s problematic, competitors can go public with the fact that they think the model being released is unsafe. The legal liability would be enormous.” “The two leading AI companies are Anthropic and OpenAI. Their models are quite close in capability, so it’s difficult for either one to slow down without handing the lead to the other. On balance, I think Anthropic puts more care into their safety than OpenAI. But even Anthropic acknowledges they are worried about their models. Many people from Anthropic have publicly voiced concern that their models are scaring them, that they’re getting scary smart.”

DogeDesigner

6,110,336 次观看 • 19 天前