Dylan Patel

1. Key Themes

The Frontier Labs Are Now Profitable and Self-Funding Growth

Anthropic and OpenAI have crossed a critical inflection point from VC-funded loss machines to profitable businesses. The shift in gross margin per megawatt is the core signal.

"GPT-4 being served on NVIDIA Hopper GPUs was generating negative gross margin for OpenAI. But now, when OpenAI serves GPT-5.6 or Anthropic serves Opus 5 or Mythos, Fable 5, their revenue generation has passed well beyond the sort of incremental $10, $15 million per megawatt. In the case of Anthropic, the revenue has gone as high as $50 million per megawatt." — Dylan Patel 00:02:50

"Anthropic started turning a profit in Q2. It's believed at some point in Q3, OpenAI could potentially start turning a profit even with the big rise of Codex and 5.6 and all this." — Dylan Patel 00:01:48

Compute Is Rapidly Centralizing Into Two Labs

The share of incremental global compute flowing to Anthropic and OpenAI is already ~30% and is projected to hit 40–50% next year, with potential to exceed 70–80% by 2028.

"Anthropic OpenAI are taking as much as 40% to 50% of compute next year. And the centralization doesn't look like it's slowing down or stopping. In fact, it looks like it's only accelerating." — Dylan Patel 00:04:17

"By the time you're in, like, towards the end of 2028, if this trend continues, which I see nothing that's stopping it, you've got them just controlling most of the usable flops in the world on their own." — Dylan Patel 00:06:39

Labs Are Quietly Reallocating More Compute to R&D, Not Inference

A non-consensus view: as revenue per megawatt rises, labs are choosing to reinvest marginal compute into training and research rather than revenue-generating inference — accelerating toward AGI rather than maximizing near-term profit.

"I do believe that the labs are going to allocate less and less compute to inference over time, which I think is very non-consensus, right? Everyone's sort of the standard belief of most people's, oh, most compute will go to inference. Most of it will go to forward passes for training, not maybe necessarily revenue generating inference." — Dylan Patel 00:30:30

"In January, they added less compute than December. And yet their revenue adds skyrocketed, and then they've sort of plateaued. They're only adding, you know, they're not adding $25 billion ARR every month now. And so that means the marginal megawatt they're getting is going higher percentage to R&D than it is inference." — Dylan Patel 00:33:09

The AI CapEx Supercycle Will Trigger a Sovereign Debt and Interest Rate Crisis

The scale of AI infrastructure spending — projected at $11 trillion in CapEx from 2024–2029, with $5 trillion needing to be debt-financed — will crowd out all other borrowers and push interest rates significantly higher, with devastating consequences for developing nations and rate-sensitive industries.

"In the modeling that we do, we have about $11 trillion of CapEx from 2024 to 2029. Total. And if you do — you know, if you fund a lot of this with cash flows and as much as you can, you still end up with north of $5 trillion of credit that need to be issued for this $11 trillion-plus buildup." — Dylan Patel 00:54:26

"Meta's raised at like 5% to 6%. I don't see why they wouldn't pay 8%. Because they would happily pay 8% because the return from the compute that they're going to build is humongous. That makes everyone else in the economy also pay 250 bps more." — Dylan Patel 00:57:19

"Basil Hopper, who's a good friend and he's an economist, he made this point that we'll see a second Volcker shock. So in the 80s, to fight inflation, Fed Chair Paul Volcker raised interest rates like more than 5%, and that caused some 40 different countries, mostly Latin America, to default in that decade. And I think that will probably happen again." — Dwarkesh Patel 00:58:44

Safety Regulation Is Becoming the Primary Constraint on Lab Revenue Growth

The most significant near-term cap on lab monetization is not model capability or compute access — it is regulatorily-enforced withholding of best models from external deployment and even internal use.

"OpenAI not releasing Astra. OpenAI stopping training for two weeks. Anthropic not releasing what their safety assessment set is Model 2, which is widely believed to be the next version of Mythos. They're clearly not releasing their best models, and in which case their revenue per megawatt stalls or even can start to decline again because other models are competitive again." — Dylan Patel 00:16:21

"Anthropic had to stop giving Mythos to foreign employees for a bit... Like internally as well?" — Dwarkesh Patel 01:05:20

The Supply Chain Cannot React Fast Enough to Meet Demand — Creating Massive Arbitrage Opportunities

The bullwhip effect across the semiconductor and data center supply chain means there is a sustained multi-year gap between the economics that should be incentivizing supply expansion and the actual supply coming online.

"If anyone had $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one and wait, wait, wait, and then sell it for north of a billion dollars." — Dylan Patel 00:10:04

"Carl Zeiss, they're like, yeah, yeah, yeah, we need to make 100 EUV tools by the end of the decade. I think when we first had our episode earlier this year, they didn't even think they needed to make that many, enough mirrors to make 100 EUV tools a year." — Dylan Patel 00:10:28

The Value Capture Layer in AI Keeps Shifting — and Currently Sits at the Model Layer

Value in the AI stack has migrated from chips/fabs (2023) to the model layer (now), and end customers still capture far more than the labs charge.

"If we go back again a year ago, the hardware supply chain was generating all this gross margin while literally everyone else was losing money on it. OpenAI and Anthropic were just plowing VC money in... Initially in 2023, the memory guys were making no money off of HBM or memory for AI, even though theoretically their value they were delivering was humongous." — Dylan Patel 00:20:38

"Jane Street is capturing $300 million per megawatt or $500 million per megawatt" while paying Anthropic prices that reflect only a fraction of the value extracted. — Dylan Patel 01:14:48

AI Compute Centralization Is an Inexorable Force Toward Economic and Political Power Concentration

Every structural force — economies of scale in training, the compute premium for frontier models, RSI, continual learning from deployment data — points toward power concentrating in one or two entities.

"In all scenarios of the world, you know, there's 80,000 worlds and only one of them, Anthropic, doesn't own the whole world." — Dylan Patel 01:15:39

"Unless AI progress slows down, unless governments regulate the fuck out of it, this is all that happens. In which case, you know, we're headed for a world where either we have super concentration of resources and we pray that that one company gets everything right." — Dylan Patel 01:14:26

China Is Structurally Far Behind on Compute, But Will Hockey-Stick After 2027

Export controls have kept China under 10% of global incremental compute, but domestic fab capacity from SMIC and CXMT coming online in 2027–2028 will create a step-change.

"China today, domestically still continues to have sub 10% of incremental new compute. So in 2028, it might start to inflect up, I think. But it's pretty easy to say China will have like 30 gigawatts of AI compute or less. By 2028." — Dylan Patel 00:35:28

"In 2028, especially, fabs start to go up from SMIC and CXMT and such, where domestic production is actually reaching many millions of units a year. And now they're incrementally adding, you know, 5, 10 gigawatts in just 2028 of domestically produced chips." — Dylan Patel 00:36:14

2. Contrarian Perspectives

The Labs Will Voluntarily Reduce the Share of Compute Going to Inference as Revenue Per Megawatt Rises

The conventional wisdom is that inference is where most compute will go. Dylan argues the opposite: the higher the revenue per inference megawatt, the more incentive the labs have to redirect compute to training — because the ROI on making the next model better exceeds the ROI on selling tokens.

"If you now get to generating $60, $70 million per megawatt, do you still allocate 40% to inference and generate all this profit and then do dividends and share buybacks? Or do you go build AGI? And I think the obvious answer from Anthropic and OpenAI, and not just at the executive level, but also their board, is go build AGI because it's way more profitable." — Dylan Patel 00:30:58

Most of the World's Compute Will Be Owned by Two Private Companies by 2028

This would have sounded absurd even two years ago. The specific numbers: Anthropic and OpenAI are going from roughly 2 gigawatts each at the start of 2025 to taking ~40–50% of incremental compute in 2026, and potentially 70–80% by 2028.

"By the time you're in, like, towards the end of 2028, if this trend continues, which I see nothing that's stopping it, you've got them just controlling most of the usable flops in the world on their own." — Dylan Patel 00:06:39

AI Will Trigger a Sovereign Debt Crisis in Developing Countries Before Singularity

The higher interest rate environment driven by AI CapEx demand will cause multiple sovereign defaults — a second Volcker shock — well before any RSI or AGI transition occurs.

"I think that will probably happen again. In fact, okay, now we're getting into like singularity talk. So we've been talking about what happens if interest rates... I think this all happens before singularity, by the way." — Dwarkesh Patel 00:58:44

"Every country that is not involved in the production of AI defaults. Every stock that is not an AI stock is like worth basically zero because discounted cash flows are worth nothing." — Dwarkesh Patel 01:00:23

If You Are Truly AI-Pilled, Everything in the Economy Should Trade at 2–3x Earnings

The logic: if memory, chips, and AI infrastructure face a structurally high-discount-rate world, then all stable cash-flow businesses should be valued at historically low multiples. High memory stock valuations are internally inconsistent with a belief in AI-driven economic transformation.

"If you're really AI pilled, everything in the economy should trade at like two or three times earnings. And if you're not AI pilled, then sure, they're over earning. So it's sort of like an argument for why, like, I think memory is going to do great. But, you know, memory stocks shouldn't, you know, 10x or whatever again." — Dylan Patel 01:01:28

China's Compute Gap Is Larger Than It Appears Because Quality Adjustments Dramatically Reduce Effective Chinese Gigawatts

The raw gigawatt numbers overstate China's position. Chinese domestic chips are 3–5x worse per watt than frontier Western chips, meaning China's 30 gigawatts by 2028 is functionally far less.

"Those chips are definitely worse than the chips that NVIDIA will have in 28 or Google will have in 28 or OpenAI will have in 2028." — Dylan Patel 00:36:14

"So even the gigawatt number overstates things, you're saying. It's like 30 gigawatts, but it's really much worse chips." — Dwarkesh Patel 00:38:10 / "Implying that there's nothing to slow down the U.S. labs." — Dylan Patel 00:38:25

3. Companies Identified

Anthropic

Leading AI lab. Mentioned throughout as the clearest case study of the lab economics inflection: turned profitable in Q2, generating up to $50M per megawatt in revenue, approaching 5 gigawatts of compute by end of 2025, using TPUs purchased from Google deployed via FluidStack, and quietly allocated increasing compute share to R&D.

"In the case of Anthropic, the revenue has gone as high as $50 million per megawatt." — Dylan Patel 00:02:50

OpenAI

The other dominant lab. Believed to potentially turn profitable in Q3, operating GPT-5.6 and Codex, building its own chips, and also increasing compute centralization.

"It's believed at some point in Q3, OpenAI could potentially start turning a profit even with the big rise of Codex and 5.6 and all this." — Dylan Patel 00:01:48

SpaceX

Identified as the plausible third-largest compute hoarder — alongside Meta — building compute on its own balance sheet without pre-signed customers, then selling to labs at massive premiums ($25–40M per gigawatt).

"SpaceX is building a ton of compute. And they're actively going to lease quite a bit of it to Anthropic and OpenAI, most likely, because they're the ones who have the marginal capability to pay the highest price." — Dylan Patel 00:04:46 "Elon wouldn't have sold if it was 15. But he's selling because it's 25 plus." — Dylan Patel 00:27:00

Jane Street

Quantitative trading firm. Cited as one of Anthropic's largest customers and as a canonical example of end-user value capture vastly exceeding what the lab charges — generating potentially $300–500M per megawatt in value from tokens they buy at a fraction of that.

"Jane Street with their exclusive contract with OpenAI for GPT 5.6 Ultra Fast Mode or Jane Street where they're like one of Anthropic's biggest customers is generating way, way, way, way more value out of the tokens they're paying for than Anthropic is generating in terms of profit." — Dylan Patel 00:18:49 "Jane Street is capturing $300 million per megawatt or $500 million per megawatt." — Dylan Patel 01:14:48

ASML

Semiconductor lithography equipment monopolist. Cited as the critical bottleneck in EUV tool production — currently targeting ~100 EUV tools per year by end of decade, constrained by Carl Zeiss mirror supply.

"If anyone had $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one and wait, wait, wait, and then sell it for north of a billion dollars." — Dylan Patel 00:10:04

Carl Zeiss

Optics manufacturer supplying mirrors for ASML EUV machines. The deepest physical bottleneck in the entire AI compute supply chain.

"You go to talk to someone at Carl Zeiss, they're like, yeah, yeah, yeah, we need to make 100 EUV tools by the end of the decade. I think when we first had our episode earlier this year, they didn't even think they needed to make that many, enough mirrors to make 100 EUV tools a year." — Dylan Patel 00:10:28

NVIDIA

Cited as the dominant accelerator supplier, with Jensen Huang's exclusive arrangement with SpaceX/xAI noted, and as a company that could raise prices but hasn't fully yet.

"Jensen where he's now all of a sudden using Twitter. And, you know, Elon's saying they're exclusive to NVIDIA. But why doesn't Jensen raise his prices?" — Dylan Patel 00:26:30

TSMC

Fabrication partner for most frontier AI chips. Noted as raising prices slowly relative to how fast value has accrued to it, and as the manufacturer of chips that smuggled their way into Chinese supply chains.

"TSMC raising prices very slowly, but memory companies raising prices very quickly." — Dylan Patel 00:26:39

SK Hynix

HBM memory supplier. Called out alongside Micron and Samsung as companies that should be — and are beginning to — raise prices in response to value migration.

"SK Hynix and Micron and Samsung looked at NVIDIA and were like, well, why don't they raise their price?" — Dylan Patel 00:26:30

Micron

Memory company. Mentioned alongside SK Hynix and Samsung as a company that should trade at low multiples if you are truly AI-pilled.

"Why does Micron or Hynix or Kyoxya trade at two or three times earnings? And it's like, well, if you're really AI pilled, everything in the economy should trade at like two or three times earnings." — Dylan Patel 01:01:28

Google

Mentioned as TPU supplier to Anthropic (TPU v7), as a hyperscaler that bought SpaceX compute at $40B per gigawatt, and as a major compute infrastructure builder.

"Anthropic with TPUs that they're purchasing from Google and deploying with FluidStack." — Dylan Patel 00:04:46

FluidStack

Deployment infrastructure company used by Anthropic to deploy Google TPUs.

"Anthropic with TPUs that they're purchasing from Google and deploying with FluidStack." — Dylan Patel 00:04:46

ByteDance / Seed

Chinese AI lab. Called out as the outlier among Chinese labs with significantly more compute than peers like Kimi.

"ByteDance Seed being the one outlier where they have significantly more than that." — Dylan Patel 00:40:04

Moonshot AI (Kimi)

Chinese AI lab. Cited as an example of a Chinese frontier lab running on as little as a few hundred megawatts — a fraction of Anthropic's compute — yet producing competitive models.

"Kimi is not running, you know, a gigawatt or anywhere close to it." — Dylan Patel 00:40:35

SMIC

China's leading domestic semiconductor foundry. Cited as beginning to produce meaningful domestic chip volumes in 2027–2028.

"In 28, especially, fabs start to go up from SMIC and CXMT and such, where domestic production is actually reaching many millions of units a year." — Dylan Patel 00:36:14

CXMT

Chinese memory chip manufacturer. Mentioned alongside SMIC as contributing to China's domestic compute uplift post-2027.

"Fabs start to go up from SMIC and CXMT and such." — Dylan Patel 00:36:14

Broadcom

Mentioned as one of the semiconductor companies with balance sheet capacity to help fund infrastructure CapEx.

"There's some semiconductor companies like NVIDIA and Broadcom and the memory companies turning around and deciding to fund some of this capex." — Dylan Patel 00:46:09

Amazon / AWS

Hyperscaler mentioned as a major compute builder, Bedrock operator for Anthropic models, and debt-raiser for AI infrastructure.

"When Amazon is serving Bedrock Anthropic models, that counts as Anthropic compute in sort of our worldview." — Dylan Patel 00:14:06

Open Router

Inference marketplace. Mentioned as the distribution layer where anyone with a GB300 rack and open-weight models can immediately start generating revenue above compute cost.

"Go put it on Open Router. It's very simple. And you'll start generating more revenue than you're paying for the compute." — Dylan Patel 00:14:36

Antithesis

Deterministic software testing platform enabling time-travel debugging and perfect reproducibility in distributed systems. Mentioned as a sponsor.

Grok / xAI

Mentioned via Grokbot as a multi-agent automation tool used by Dwarkesh for recruiting workflows.

4. People Identified

Founder of SemiAnalysis. The primary expert guest providing granular, model-based forecasts of global compute deployment, lab economics, supply chain constraints, and geopolitical compute dynamics.

"In the modeling that we do, we have about $11 trillion of CapEx from 2024 to 2029. Total." — Dylan Patel 00:54:26

Dario Modeji (Dario)

CEO of Anthropic. Referenced in context of the centralization debate — Gavin Baker claimed Dario believes only one company will exist; Dario and Sholto disputed it.

"I think Gavin Baker was like — Dario believes that there's only going to be one company in the world. And then, you know, Sholto and Dario came out and were like, no, no, no, we didn't say that." — Dylan Patel 01:09:47

Sholto Douglas

AI researcher, apparently a mutual acquaintance of both Dwarkesh and Dylan, referred to as "our roommate Sholto" and connected to Anthropic leadership discussions.

"It's sort of what you and I believe we're in a world where models are capable of that... The limiter on AGI is not how fast can the research engineers like, you know, our roommate Sholto can crank the gears." — Dylan Patel 01:02:56

Basil Hopper

Economist, friend of Dwarkesh. Made the prediction of a second Volcker shock driven by AI CapEx crowding out sovereign borrowing — causing ~40 country defaults as happened in the 1980s.

"Basil Hopper, who's a good friend and he's an economist, he made this point that we'll see a second Volcker shock." — Dwarkesh Patel 00:58:44

Gavin Baker

Investor. Cited for the claim that Dario believes only one company will survive the AI race.

"I think Gavin Baker was like — Dario believes that there's only going to be one company in the world." — Dylan Patel 01:09:47

Damon Binder

Researcher. Cited for input-output table analysis showing how fast a fully automated economy could compound — potentially doubling annually.

"There's a researcher, Damon Binder, who's done great work on this. But basically, if you look at like input-output tables in a fully automated economy, just like what would it take to like double the entire stock of things in the economy?" — Dwarkesh Patel 00:59:41

Paul Volcker

Former Fed Chair. Referenced historically as the person who raised interest rates to fight inflation in the 1980s, causing ~40 sovereign defaults — the analogy for what AI CapEx-driven rate rises may cause.

"Fed Chair Paul Volcker raised interest rates like more than 5%, or it's like something like 8%, real interest rates 8%. And that caused some 40 different countries, mostly Latin America, to default in that decade." — Dwarkesh Patel 00:58:44

Jensen Huang

CEO of NVIDIA. Referenced as having an exclusive arrangement with Elon Musk/xAI, and as a person who should logically raise prices given demand.

"Jensen where he's now all of a sudden using Twitter. And, you know, Elon's saying they're exclusive to NVIDIA. But why doesn't Jensen raise his prices?" — Dylan Patel 00:26:30

5. Operating Insights

Compute Arbitrage Is Already Viable at Current Prices — Anyone Can Stand Up a Profitable Inference Business Today

The gap between compute cost and inference revenue is so wide right now that a technically competent person or small team can generate positive gross margin immediately using commodity hardware and open-weight models. This is a tactical window before compute prices reprice upward.

"Go get a GB300 rack. Go download the Kimi weights. Go download vLLM or SGLang. Set it up. You know, Codex and Fable can actually help you do this. It's pretty simple... Go put it on Open Router. It's very simple. And you'll start generating more revenue than you're paying for the compute." — Dylan Patel 00:14:36

When Computing as Infrastructure Is Bottlenecked, the Bottleneck Itself Is the Investable Asset

Dylan articulates a specific, actionable arbitrage in physical infrastructure: the value of a turbine, an EUV tool, or anything bottlenecking a data center is massively above its cost because the downstream revenue it unlocks is so large. Operators should think about where the physical constraint is and own it.

"You've seen people do funny arbitrages here where they buy, like, turbines and then they try and resell them. Because the value of a turbine is way more because it's the thing bottlenecking a data center. If anyone had $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one and wait, wait, wait, and then sell it for north of a billion dollars." — Dylan Patel 00:10:04

Build Compute Speculatively on Your Own Balance Sheet — Don't Wait for a Customer to Sign

The companies generating extraordinary optionality (Meta, SpaceX) did so by building compute without a pre-signed customer, giving them the power to either use it internally or sell it at peak market prices. Companies that wait for customers to sign first give away pricing leverage.

"Meta and SpaceX are saying, actually, I'm going to build the compute and I can start to rent it out for not 13. I can sell it for 25, 50 and more." — Dylan Patel 00:24:23

6. Overlooked Insights

Most Frontier Model Training Runs Use Only a Small Fraction of a Lab's Total Compute — The Rest Is Research

Dylan reveals a structural breakdown of lab compute budgets that is almost never discussed: even when a lab has multiple gigawatts, the actual training run for a frontier model uses less than 200 megawatts. The majority of compute is consumed by exploratory research — testing architectures, data mixes, hyperparameters — not the final training run. This has profound implications: Chinese labs with only 100–200 megawatts are not as compute-disadvantaged as the raw gigawatt comparison suggests, because the bottleneck to model quality is research compute, not training run scale. But it also means that as automated research arrives, this ratio will flip and labs with more compute will see their advantage multiply dramatically.

"When Anthropic trains Mythos, it's sub 200 megawatts, right? The pre-training... At most, the most they ever used at one point in time was maybe 200 megawatts. And then, in reality, they had multiple gigawatts, so most of their compute was going to the research, not the development of a model... 50% of the compute is research, like 10% of the compute is development, and then 40% is inference." — Dylan Patel 00:41:03

"As we get closer and as we get further and further down, implement automated coding, automated researcher, I actually expect the percentage of the compute budget that goes to research versus training to start to, like, become a lot more fuzzy or even higher for training." — Dylan Patel 00:42:30

The Total CapEx Figure Cited for AI Dramatically Understates Actual Investment Required Because It Excludes Lead-Time Infrastructure

When people cite $40–50 billion data center CapEx figures, they are counting only critical IT — servers, networking, fiber, transceivers. They are not counting the data center buildings or the power generation infrastructure, both of which must be funded and constructed one to two years ahead of the compute itself. This means the real CapEx burden is being systematically underestimated by everyone citing headline AI investment numbers, and the crowding-out effect on credit markets is correspondingly larger than most analysts model.

"When you talk about AI CapEx, people are saying $40, $50 billion. But that's really just the critical IT. Yeah. Right? The servers, the networking, the fiber, the transceivers, optical communications, all this sort of stuff. It doesn't account for the data center itself or the power plants themselves, which are being built ahead of time. So, if I'm building 100 gigawatts this year and 150 gigawatts next year, well, then all of the buildings for that 150 gigawatts need to be built in CapEx this year." — Dylan Patel 00:43:24