{
 "episode_id": "dwarkesh-podcast-2026-08-25-dylan-patel-anthropic-openai-will-have-most-of-the-world-s-c",
 "title": "Dylan Patel – Anthropic & OpenAI will have most of the world's compute by 2028",
 "podcast": "dwarkesh-podcast",
 "published_at": "2026-08-25",
 "speakers": [
  "Dwarkesh Patel",
  "Dylan Patel"
 ],
 "summary": "Dylan Patel (SemiAnalysis) returns to argue that compute is centralizing into two labs faster than almost anyone models: Anthropic and OpenAI go from ~2GW each at the start of 2026 to above 5GW by year-end, take 40-50% of all new compute next year, and — because each new watt is 3-5x more efficient than the last — end up controlling most of the world's usable flops by the end of 2028. The mechanism is a margin flip: serving GPT-4 on Hopper was gross-margin-negative, while Anthropic now clears as much as $50M of revenue per megawatt against a $10-15M cost, letting the labs outbid everyone for compute and recycle the profit into training. Patel's non-consensus call is that the labs will allocate a *declining* share of compute to inference over time — building AGI beats dividends and buybacks — and that they are already doing so. He argues the binding constraint on takeoff is not research speed but political economy: regulation (unreleased models, state data-center moratoriums), supply-chain bullwhip through ASML and the memory makers, and above all capital, with SemiAnalysis modelling ~$11T of CapEx through 2029 of which ~$5T must be raised as debt. Dwarkesh presses the resulting macro case — that AI's extraordinary returns push up the global rate of interest, crater the discounted cash flows of every non-AI equity, and trigger a second Volcker shock that defaults heavily-indebted countries like Pakistan and Nigeria. On China, Patel credits export controls with holding it under 10% of deployed watts and ~30GW by 2028, but expects a domestic-fab hockey stick from 2028-29. The episode closes on concentration: Dwarkesh's estimate that a single lab could hold more effective AI labor than there are people on earth by 2030, and Patel's flat refusal to trust either the government or the lab CEOs with it.",
 "topics": [
  {
   "name": "Lab economics and the margin flip",
   "timestamp": "0:36",
   "summary": "Patel walks through how the labs went from venture-funded losses to profit — Anthropic profitable from Q2 — as revenue per megawatt passed the $10-15M cost of compute, in Anthropic's case reaching $50M. That spread is what funds incremental training."
  },
  {
   "name": "How fast compute centralizes into two labs",
   "timestamp": "3:50",
   "summary": "Anthropic and OpenAI go from ~2GW each to above 5GW inside the year and take 40-50% of new compute next year. Patel argues half of the world's incremental compute goes to them by end of 2027, and that efficiency gains per watt compound the concentration into 2028."
  },
  {
   "name": "Supply-chain bottlenecks and the bullwhip",
   "timestamp": "7:01",
   "summary": "Dwarkesh argues a ~$6B fab CapEx outlay yields a gigawatt a year worth ~$100B of revenue, so capitalism should expand supply. Patel counters that the whip takes years to reach the tail — Carl Zeiss mirrors, ~100 ASML EUV tools for 2030 — and the world is capital-constrained."
  },
  {
   "name": "Where the value gets captured",
   "timestamp": "18:23",
   "summary": "Patel traces value capture as it migrated from the hardware supply chain to the model layer, and argues most of it still lands with end users like Jane Street and Meta rather than with OpenAI and Anthropic."
  },
  {
   "name": "Compute pricing and the Elon arbitrage",
   "timestamp": "22:18",
   "summary": "SpaceX sold a gigawatt to Google at ~$40B, well above the $13M/MW norm. Patel explains why balance-sheet holders (Meta, SpaceX) who build unfinanced compute now set the price, while everyone else must pre-contract a customer to raise capital."
  },
  {
   "name": "Regulation as the brake on revenue per megawatt",
   "timestamp": "28:31",
   "summary": "Unreleased frontier models, two-week training pauses, and state-level data-center moratoriums cap external progress. Patel expects Anthropic to keep its best model ~6 months ahead of what ships publicly, and sees the regulatory drag as inflationary."
  },
  {
   "name": "Shifting compute from inference to training",
   "timestamp": "30:30",
   "summary": "Patel's non-consensus call: the labs will dedicate a falling share of compute to inference and a rising share to research and training, because building AGI beats taking the profit — and he argues they have already been doing this for three months."
  },
  {
   "name": "China's compute trajectory and export controls",
   "timestamp": "34:14",
   "summary": "The US now deploys ~70% of AI watts against China's sub-10%. Patel expects China at ~30GW or less in 2028, a domestic-fab inflection from SMIC and CXMT, and a plausible 50GW year in 2029 — quality-weighted worth roughly 20GW of American chips."
  },
  {
   "name": "The CapEx bill and where the cash comes from",
   "timestamp": "44:48",
   "summary": "SemiAnalysis models ~$11T of CapEx from 2024 to 2029, ~$6T cash-funded and ~$5T debt-funded. Hyperscalers already spend everything and raise debt on top, so the marginal financier has to be infrastructure investors and the broader credit market."
  },
  {
   "name": "AI, interest rates, and a sovereign debt crisis",
   "timestamp": "47:27",
   "summary": "Dwarkesh argues AI's returns bid up the rate of interest, crowding out every other borrower: US debt service could pass 60% of tax revenue, non-AI equities crater on a higher discount rate, and heavily-indebted countries default in a second Volcker shock."
  },
  {
   "name": "Concentration of AI labor and power",
   "timestamp": "1:07:36",
   "summary": "Effective AI population at the frontier grows ~10x a year, so a single lab could plausibly hold more labor-equivalents than there are people on earth by 2030. Both agree every force points to centralization; neither trusts the government or the labs with it."
  }
 ],
 "entities": [
  {
   "name": "Dylan Patel",
   "type": "person",
   "salience": "central",
   "role_in_episode": "Guest; founder of SemiAnalysis; supplies the compute, pricing, and CapEx model behind the centralization thesis."
  },
  {
   "name": "Dwarkesh Patel",
   "type": "person",
   "salience": "central",
   "role_in_episode": "Host; presses the macro case that AI returns raise global interest rates and trigger sovereign defaults."
  },
  {
   "name": "Anthropic",
   "type": "company",
   "salience": "central",
   "role_in_episode": "One of the two labs absorbing the world's compute; profitable from Q2, clearing up to $50M revenue per megawatt."
  },
  {
   "name": "OpenAI",
   "type": "company",
   "salience": "central",
   "role_in_episode": "The other absorbing lab; building its own chips, possibly profitable in Q3, withholding Astra."
  },
  {
   "name": "SemiAnalysis",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Patel's research firm; source of the compute, pricing, and $11T CapEx models cited throughout."
  },
  {
   "name": "China",
   "type": "place",
   "salience": "central",
   "role_in_episode": "Subject of the export-control debate; sub-10% of deployed watts today, expected to hockey-stick on domestic fabs from 2028."
  },
  {
   "name": "United States",
   "type": "place",
   "salience": "discussed",
   "role_in_episode": "Deploys ~70% of AI watts; its tax base and debt service are the core of the sovereign-debt argument."
  },
  {
   "name": "NVIDIA",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Accelerator supplier whose pricing power is part of the supply-chain bullwhip."
  },
  {
   "name": "Google",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "TPU supplier to Anthropic and buyer of SpaceX compute at ~$40B a gigawatt."
  },
  {
   "name": "Meta",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Balance-sheet compute hoarder and plausible number three; also cited as an Anthropic customer and a debt issuer."
  },
  {
   "name": "SpaceX",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Builds unfinanced compute and resells it far above cost, resetting the market price per gigawatt."
  },
  {
   "name": "Amazon",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Hyperscaler raising debt for CapEx; serves Anthropic models via Bedrock."
  },
  {
   "name": "Microsoft",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Hyperscaler cited as heading toward debt-funded CapEx."
  },
  {
   "name": "ASML",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "EUV tool bottleneck; ~100 tools a year for 2030 is the supply-chain constraint."
  },
  {
   "name": "Carl Zeiss",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Mirror supplier whose slow capacity response illustrates the bullwhip effect."
  },
  {
   "name": "TSMC",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Foundry raising prices slowly relative to the memory makers; also made chips that reached Huawei."
  },
  {
   "name": "Jane Street",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Archetype of the end user capturing far more value per megawatt than the lab selling the tokens."
  },
  {
   "name": "SMIC",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Chinese foundry expected to drive the domestic compute inflection in 2028."
  },
  {
   "name": "CXMT",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Chinese memory maker paired with SMIC in the domestic-production ramp."
  },
  {
   "name": "Huawei",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Cited as an end recipient of TSMC chips routed through intermediaries."
  },
  {
   "name": "Micron",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Memory maker raising prices quickly as the supply chain rebalances."
  },
  {
   "name": "SK Hynix",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Memory supplier cited in the price-raising cascade behind NVIDIA."
  },
  {
   "name": "Samsung",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Memory supplier; also cited as a source of HBM reaching China."
  },
  {
   "name": "FluidStack",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Deployment partner for Anthropic's Google TPUs."
  },
  {
   "name": "Mythos",
   "type": "product",
   "salience": "discussed",
   "role_in_episode": "Anthropic frontier model used as the worked example of sub-200MW pre-training and of withheld/restricted deployment."
  },
  {
   "name": "Dario Amodei",
   "type": "person",
   "salience": "mentioned",
   "role_in_episode": "Named as one of the people Patel would not trust with centralized AI power."
  },
  {
   "name": "Sam Altman",
   "type": "person",
   "salience": "mentioned",
   "role_in_episode": "Named alongside Amodei in Patel's refusal to trust concentrated control."
  },
  {
   "name": "Elon Musk",
   "type": "person",
   "salience": "discussed",
   "role_in_episode": "Credited with the compute arbitrage that reset pricing from $15M to $25M+ per megawatt."
  },
  {
   "name": "Jensen Huang",
   "type": "person",
   "salience": "mentioned",
   "role_in_episode": "Referenced for the export-control debate and for NVIDIA's latitude to raise prices."
  },
  {
   "name": "Paul Volcker",
   "type": "person",
   "salience": "mentioned",
   "role_in_episode": "Historical anchor for the sovereign-default scenario Dwarkesh expects to repeat."
  },
  {
   "name": "AI compute",
   "type": "concept",
   "salience": "central",
   "role_in_episode": "The unit of account for the whole conversation — gigawatts, revenue per megawatt, and who owns them."
  },
  {
   "name": "AI capex",
   "type": "concept",
   "salience": "central",
   "role_in_episode": "The ~$11T through-2029 bill and the financing question it creates."
  },
  {
   "name": "AI regulation",
   "type": "concept",
   "salience": "central",
   "role_in_episode": "Unreleased models, training pauses, and state data-center moratoriums — the brake on external AI progress."
  },
  {
   "name": "Export controls",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Credited by both with materially holding back Chinese compute stock."
  },
  {
   "name": "Interest rates",
   "type": "concept",
   "salience": "central",
   "role_in_episode": "The transmission channel from AI returns to the rest of the world economy."
  },
  {
   "name": "Sovereign debt crisis",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The predicted second Volcker shock defaulting heavily-indebted developing countries."
  },
  {
   "name": "Recursive self-improvement",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The takeoff scenario whose timing determines whether China ever catches up."
  },
  {
   "name": "AGI",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The alternative use of compute that Patel argues beats dividends and buybacks."
  },
  {
   "name": "AI value capture",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Which layer — chips, models, apps, users — keeps the surplus AI generates."
  }
 ],
 "claims": [
  {
   "holder": "Dylan Patel",
   "claim": "The centralization of compute into OpenAI and Anthropic is accelerating rather than slowing.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute"
   ],
   "timestamp": "3:50",
   "quote": "And this centralization doesn't look like it's slowing down or stopping. In fact, it looks like it's only accelerating."
  },
  {
   "holder": "Dylan Patel",
   "claim": "By the end of 2028 OpenAI and Anthropic will control most of the world's usable flops on their own, because their share of incremental compute is also the newest and most efficient.",
   "stance": "prediction",
   "conviction": 0.75,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute"
   ],
   "timestamp": "5:52",
   "quote": "you've got them just controlling most of the usable flops in the world on their own."
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "The gap between fab CapEx and end AI revenue is so large (roughly 100x) that capitalism should force a much faster supply expansion than is happening.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "TSMC",
    "ASML",
    "AI capex"
   ],
   "timestamp": "9:11",
   "quote": "That still means there's 100X discrepancy between fab capex and end revenue generated."
  },
  {
   "holder": "Dylan Patel",
   "claim": "The supply chain will expand, but with a long bullwhip lag — the price signal takes years to reach suppliers like Carl Zeiss, so the constraint holds in the near term.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "ASML",
    "Carl Zeiss",
    "AI compute"
   ],
   "timestamp": "10:04",
   "quote": "yes, capitalism will cause these things to expand, but it's a whip, right? It takes a long time for the whip signal to get to the tail end."
  },
  {
   "holder": "Dylan Patel",
   "claim": "The labs cannot reach ~100GW by 2028 at today's compute prices — they would have to pay $25-50M per megawatt to take 70% of the world's compute.",
   "stance": "assertion",
   "conviction": 0.75,
   "subjects": [
    "AI compute",
    "OpenAI",
    "Anthropic"
   ],
   "timestamp": "16:03",
   "quote": "They do have to start paying 25, 30, 50 million dollars a megawatt to really gobble up 70% of the world's compute"
  },
  {
   "holder": "Dylan Patel",
   "claim": "The safety regulation the US labs advocate for slows them down far more than it slows open-source Chinese models.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "AI regulation",
    "OpenAI",
    "Anthropic",
    "China"
   ],
   "timestamp": "16:03",
   "quote": "This regulation that they advocate for is actually slowing down the labs a lot more than it slows down sort of the open source Chinese language models."
  },
  {
   "holder": "Dylan Patel",
   "claim": "Because the labs are withholding their best models, revenue per megawatt can stall or decline as competing models catch up.",
   "stance": "assertion",
   "conviction": 0.75,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI regulation"
   ],
   "timestamp": "16:03",
   "quote": "They're clearly not releasing their best models, and in which case, the revenue per megawatt stalls or even can start to decline again because other models are competitive again."
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "Full AGI would generate many hundreds of billions of dollars of value per gigawatt, far above today's ~$100B.",
   "stance": "prediction",
   "conviction": 0.6,
   "subjects": [
    "AGI",
    "AI compute"
   ],
   "timestamp": "18:18",
   "quote": "But it would be many hundreds of billions of dollars if you get like full AGI per gigawatt."
  },
  {
   "holder": "Dylan Patel",
   "claim": "Most of the value AI models generate is captured by their users, not by OpenAI and Anthropic.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI value capture",
    "Jane Street"
   ],
   "timestamp": "18:23",
   "quote": "most of the value that these models generate does not get given to OpenAI and Anthropic."
  },
  {
   "holder": "Dylan Patel",
   "claim": "State-level data-center restrictions will cut supply, raise costs, and pass those costs through — making AI regulation inflationary.",
   "stance": "prediction",
   "conviction": 0.75,
   "subjects": [
    "AI regulation",
    "AI compute"
   ],
   "timestamp": "28:31",
   "quote": "These sorts of things are going to decrease supply and increase cost and that's going to get passed on as well."
  },
  {
   "holder": "Dylan Patel",
   "claim": "In a takeoff scenario Anthropic would keep its best model roughly six months ahead of anything released externally, for both safety and competitive reasons.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "Anthropic",
    "AI regulation"
   ],
   "timestamp": "28:31",
   "quote": "why would Anthropic not have their best model six months ahead of what is externally available because of safety and regulation, but also the competitive advantage."
  },
  {
   "holder": "Dylan Patel",
   "claim": "The labs will allocate a declining share of compute to inference over time — a view he flags as explicitly non-consensus — because building AGI beats taking the profit.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute",
    "AGI"
   ],
   "timestamp": "30:30",
   "quote": "this is sort of what I personally believe that the labs are going to allocate less and less compute to inference over time, which I think is very non-consensus"
  },
  {
   "holder": "Dylan Patel",
   "claim": "The labs are already shifting marginal compute toward R&D rather than inference — visible in revenue adds plateauing while compute keeps growing.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "Anthropic",
    "OpenAI",
    "AI compute"
   ],
   "timestamp": "32:40",
   "quote": "they are factually increasing their compute towards R&D today."
  },
  {
   "holder": "Dylan Patel",
   "claim": "China will eventually hockey-stick its compute build because scaling manufacturing quickly is what it is best at.",
   "stance": "prediction",
   "conviction": 0.8,
   "subjects": [
    "China",
    "AI compute"
   ],
   "timestamp": "37:01",
   "quote": "China is definitely going to hockey stick if there's anything China's really good at is scaling manufacturing really, really quickly."
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "If China's compute stock really is that far behind by the time automated research arrives, US export controls will have been a notable policy success.",
   "stance": "opinion",
   "conviction": 0.65,
   "subjects": [
    "Export controls",
    "China",
    "United States"
   ],
   "timestamp": "38:37",
   "quote": "So, if that ends up being the case, that would have worked. I think that's actually a notable success."
  },
  {
   "holder": "Dylan Patel",
   "claim": "If takeoff is slower than the fast-RSI scenario implies, Chinese subsidies mean China catches up drastically on semiconductors and therefore on compute.",
   "stance": "prediction",
   "conviction": 0.7,
   "subjects": [
    "China",
    "Export controls",
    "AI compute"
   ],
   "timestamp": "39:25",
   "quote": "if takeoff is not as fast as you're implying, but actually takes longer, then ultimately China will catch up drastically on the semiconductor side"
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "The US will be fine fiscally because it can tax the data centers, provided they are built in America.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "United States",
    "AI capex"
   ],
   "timestamp": "47:27",
   "quote": "I think the US will be fine at the end of the day, because if the data centers are built in America, we can fundamentally just tax the data centers."
  },
  {
   "holder": "Dylan Patel",
   "claim": "The crowding-out effect on credit — not physics — is what stops an unbounded compute buildout.",
   "stance": "assertion",
   "conviction": 0.75,
   "subjects": [
    "Interest rates",
    "AI capex"
   ],
   "timestamp": "52:25",
   "quote": "So, this crowding out effect is actually the thing that I like, is the reason it's not like YOLO 1 billion gigawatts."
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "A higher discount rate from AI-driven interest rates will crater the valuation of stable-cash-flow, non-AI equities even if the index holds up.",
   "stance": "prediction",
   "conviction": 0.7,
   "subjects": [
    "Interest rates",
    "AI value capture"
   ],
   "timestamp": "57:55",
   "quote": "if interest rates rise, the discount rate increases, which means that the discount cash flows of all equities crater"
  },
  {
   "holder": "Dylan Patel",
   "claim": "If you take AI seriously, every non-AI asset should trade at two or three times earnings — meaning the stock market should crash.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "Interest rates",
    "AI value capture"
   ],
   "timestamp": "1:01:28",
   "quote": "if you're really AI-pilled, everything in the economy should trade it like two or three times earnings."
  },
  {
   "holder": "Dylan Patel",
   "claim": "The limiter on AGI is not researcher throughput but how much the rest of the world permits the reallocation of capital toward it.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "AGI",
    "AI regulation",
    "AI capex"
   ],
   "timestamp": "1:01:28",
   "quote": "the limiter on AGI is not how fast can the research engineers like our roommate Sholto can crank the gears. It's actually just like, how much does the rest of the world let that happen?"
  },
  {
   "holder": "Dylan Patel",
   "claim": "A slow takeoff remains possible even with capable models, because economic and regulatory friction — not capability — is the binding constraint.",
   "stance": "opinion",
   "conviction": 0.6,
   "subjects": [
    "Recursive self-improvement",
    "AI regulation"
   ],
   "timestamp": "1:03:00",
   "quote": "But slow takeoff is, at least in my hope, possible because everything in the economy and regulatory world"
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "By the end of this decade a single lab could plausibly hold more effective AI labor than there are people on earth.",
   "stance": "prediction",
   "conviction": 0.7,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AGI"
   ],
   "timestamp": "1:07:36",
   "quote": "That there's more AI labor, more effective population within a single lab, than there are people on earth."
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "Because most of the world's cognitive work would sit inside two labs, misalignment there is effectively misalignment of most of the world.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AGI"
   ],
   "timestamp": "1:07:36",
   "quote": "if these AIs are misaligned, then most of the world is misaligned basically, because most of the world's minds are there."
  },
  {
   "holder": "Dwarkesh Patel",
   "claim": "AI training has huge economies of scale — training effort amortizes across billions of sessions — which structurally compounds whoever is ahead.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "AI compute",
    "OpenAI",
    "Anthropic"
   ],
   "timestamp": "1:11:20",
   "quote": "I think the fundamental problem is that lab, AI training has huge economies of scale."
  },
  {
   "holder": "Dylan Patel",
   "claim": "Every structural force in the market — economies of scale, continual learning, RSI — points to centralization of compute and power.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "AI compute",
    "Recursive self-improvement"
   ],
   "timestamp": "1:12:22",
   "quote": "All of these things point to centralization."
  },
  {
   "holder": "Dylan Patel",
   "claim": "Neither the government nor the lab CEOs should be trusted with centralized control of AI.",
   "stance": "opinion",
   "conviction": 0.9,
   "subjects": [
    "Dario Amodei",
    "Sam Altman",
    "AI regulation"
   ],
   "timestamp": "1:12:47",
   "quote": "I don't trust the government and I don't trust Dario and I don't trust Sam."
  },
  {
   "holder": "Dylan Patel",
   "claim": "The one saving grace of the current arrangement is that Anthropic does not capture most of the value it creates — the rest of the economy does.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "Anthropic",
    "AI value capture",
    "Jane Street"
   ],
   "timestamp": "1:13:58",
   "quote": "the one positive thing here is that today, Anthropic does not capture most of the value."
  }
 ],
 "predictions": [
  {
   "holder": "Dylan Patel",
   "prediction": "By the end of 2027, half of the world's incremental new compute goes to just OpenAI and Anthropic.",
   "timeframe": "by end of 2027",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute"
   ],
   "timestamp": "3:50",
   "quote": "it's really by the end of next year, it's already half of the incremental compute is going to Anthropic and OpenAI.",
   "due": "2027-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "By the end of 2028, OpenAI and Anthropic together control most of the world's usable flops.",
   "timeframe": "by end of 2028",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute"
   ],
   "timestamp": "5:52",
   "quote": "by the time you're in 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.",
   "due": "2028-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "Most compute will still transact at under $20 billion per gigawatt at the end of 2027, despite the headline $40B SpaceX deal.",
   "timeframe": "by end of 2027",
   "subjects": [
    "AI compute",
    "SpaceX"
   ],
   "timestamp": "22:31",
   "quote": "I think most compute will still continue to transact at sub $20 billion a gigawatt.",
   "due": "2027-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "By the end of 2027 a leading lab's blended revenue per megawatt reaches $70-80M or higher.",
   "timeframe": "by end of 2027",
   "subjects": [
    "Anthropic",
    "OpenAI",
    "AI value capture"
   ],
   "timestamp": "25:55",
   "quote": "it's like 70, 80 million dollars a megawatt blended across the company, if not higher",
   "due": "2027-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "Global incremental AI compute additions reach 90-100 gigawatts in 2029.",
   "timeframe": "in 2029",
   "subjects": [
    "AI compute"
   ],
   "timestamp": "33:45",
   "quote": "30 this year, 50 next year, 70 and 28 29 should be on the order of 90 to 100",
   "due": "2029-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "China will have roughly 30 gigawatts of AI compute or less in 2028.",
   "timeframe": "in 2028",
   "subjects": [
    "China",
    "AI compute"
   ],
   "timestamp": "34:34",
   "quote": "it's pretty easy to say China will have like 30 gigawatts of AI compute or less.",
   "due": "2028-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "Chinese domestic fabs (SMIC, CXMT) add 5-10 gigawatts of domestically produced compute in 2028 alone.",
   "timeframe": "in 2028",
   "subjects": [
    "China",
    "SMIC",
    "CXMT"
   ],
   "timestamp": "35:46",
   "quote": "now they're incrementally adding 5, 10 gigawatts in just 2028 of domestically produced chips.",
   "due": "2028-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "China can add 50 gigawatts of incremental compute in 2029, though quality-weighted that is worth closer to 20 gigawatts of American chips.",
   "timeframe": "in 2029",
   "subjects": [
    "China",
    "AI compute"
   ],
   "timestamp": "37:49",
   "quote": "it's completely reasonable that China in 2029 can do 50 gigs.",
   "due": "2029-12-31"
  },
  {
   "holder": "Dwarkesh Patel",
   "prediction": "Incremental global AI CapEx approaches $10 trillion a year by the end of the decade.",
   "timeframe": "by 2030",
   "subjects": [
    "AI capex"
   ],
   "timestamp": "44:10",
   "quote": "So very plausibly, incremental capex every year is getting close to 10 trillion dollars.",
   "due": "2030-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "AI CapEx from 2024 to 2029 totals about $11 trillion, roughly $6T cash-funded and $5T debt-funded.",
   "timeframe": "2024-2029",
   "subjects": [
    "AI capex",
    "SemiAnalysis"
   ],
   "timestamp": "54:35",
   "quote": "In the modeling that we do, we have about $11 trillion of CapEx from 2024 to 2029",
   "due": "2029-12-31"
  },
  {
   "holder": "Dylan Patel",
   "prediction": "Meta's borrowing cost rises to around 8%, up from the 5-6% it raised at recently, dragging the whole economy's cost of credit up with it.",
   "timeframe": "by 2029",
   "subjects": [
    "Meta",
    "Interest rates"
   ],
   "timestamp": "56:48",
   "quote": "I don't see why they wouldn't pay 8 percent.",
   "due": "2029-12-31"
  },
  {
   "holder": "Dwarkesh Patel",
   "prediction": "A second Volcker shock repeats the 1980s pattern, defaulting a wave of heavily-indebted developing countries.",
   "timeframe": "unspecified",
   "subjects": [
    "Sovereign debt crisis",
    "Interest rates"
   ],
   "timestamp": "58:44",
   "quote": "he made this point that we'll see a second Volcker shock ... And I think that will probably happen again"
  },
  {
   "holder": "Dwarkesh Patel",
   "prediction": "Heavily-indebted, low-tax-revenue countries such as Pakistan and Nigeria are hit hardest by the new interest rate regime.",
   "timeframe": "unspecified",
   "subjects": [
    "Sovereign debt crisis",
    "Interest rates"
   ],
   "timestamp": "51:31",
   "quote": "Those countries like Pakistan or Nigeria or something, I think are just going to be very fucked in this new interest rate regime."
  },
  {
   "holder": "Dwarkesh Patel",
   "prediction": "In the 2030s the rate of interest reaches tens of percent as an automated economy potentially doubles every year.",
   "timeframe": "in the 2030s",
   "subjects": [
    "Interest rates",
    "AGI"
   ],
   "timestamp": "59:56",
   "quote": "So then we'll go into a world, I think in the 2030s, where the rate of interest is like tens of percent."
  },
  {
   "holder": "Dwarkesh Patel",
   "prediction": "By the end of this decade, a single frontier lab holds more effective AI labor than there are people on earth.",
   "timeframe": "by end of the decade",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AGI"
   ],
   "timestamp": "1:07:36",
   "quote": "And I think that's like a thing that is very plausible by the end of this decade.",
   "due": "2030-12-31"
  }
 ],
 "facts": [
  {
   "statement": "AI infrastructure accounted for most of America's GDP growth last year.",
   "subjects": [
    "United States",
    "AI capex"
   ],
   "timestamp": "0:36",
   "quote": "most of GDP growth in America was just AI infrastructure",
   "source": "Dylan Patel"
  },
  {
   "statement": "About a third of the compute coming online this year is for OpenAI and Anthropic as end customers.",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute"
   ],
   "value": "~1/3 of new compute",
   "timestamp": "0:36",
   "quote": "about a third of the compute coming online is for the labs, for Open AI and Anthropic",
   "source": "Dylan Patel"
  },
  {
   "statement": "Global AI CapEx is a little over $1 trillion this year and is projected above $2 trillion in 2028.",
   "subjects": [
    "AI capex"
   ],
   "value": "$1T (2026) → $2T+ (2028)",
   "timestamp": "0:36",
   "quote": "we're at a little bit over a trillion dollars of CapEx this year. As we go out into 28, it's going to be more than two trillion dollars.",
   "source": "Dylan Patel"
  },
  {
   "statement": "Anthropic started turning a profit in Q2.",
   "subjects": [
    "Anthropic"
   ],
   "timestamp": "0:36",
   "quote": "Anthropic started turning a profit in Q2.",
   "source": "Dylan Patel"
  },
  {
   "statement": "The base cost of compute is roughly $10-15 million per megawatt.",
   "subjects": [
    "AI compute"
   ],
   "value": "$10-15M per MW",
   "timestamp": "0:36",
   "quote": "The base cost of compute tends to be around 10 or 13 or 15 million dollars per megawatt.",
   "source": "Dylan Patel"
  },
  {
   "statement": "Anthropic's revenue per megawatt has reached as high as $50 million, against a $10-15M cost of compute.",
   "subjects": [
    "Anthropic",
    "AI value capture"
   ],
   "value": "$50M per MW",
   "timestamp": "0:36",
   "quote": "In the case of Anthropic, the revenue has gone as high as 50 million dollars per megawatt.",
   "source": "Dylan Patel"
  },
  {
   "statement": "OpenAI and Anthropic each started the year around 2GW of compute and end it above 5GW.",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute"
   ],
   "value": "~2GW → 5GW+ in 2026",
   "timestamp": "3:50",
   "quote": "Anthropic OpenAI started at 2 for OpenAI and less than 2 for Anthropic. End of this year, they're both above 5",
   "source": "Dylan Patel"
  },
  {
   "statement": "Contracts already signed put Anthropic and OpenAI at 40-50% of new compute next year.",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "AI compute"
   ],
   "value": "40-50% of 2027 new compute",
   "timestamp": "3:50",
   "quote": "You've got Anthropic OpenAI are taking as much as 40-50% of compute next year.",
   "source": "Dylan Patel"
  },
  {
   "statement": "OpenAI is building its own chips while Anthropic buys TPUs from Google and deploys them with FluidStack.",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "Google",
    "FluidStack"
   ],
   "timestamp": "3:50",
   "quote": "OpenAI with their own chips, Anthropic with TPUs that they're purchasing from Google and deploying with FluidStack.",
   "source": "Dylan Patel"
  },
  {
   "statement": "Incremental global compute additions run about 30GW this year, 50GW next year, and 70GW the year after.",
   "subjects": [
    "AI compute"
   ],
   "value": "30 / 50 / 70 GW",
   "timestamp": "5:52",
   "quote": "if the incremental compute adds this year 30 gigawatts, next year 50 gigawatts and the year after that 70, roughly",
   "source": "Dylan Patel"
  },
  {
   "statement": "Current-generation accelerators deliver 3-5x more performance per watt than the prior generation.",
   "subjects": [
    "NVIDIA",
    "Google",
    "AI compute"
   ],
   "value": "3-5x perf/watt",
   "timestamp": "5:52",
   "quote": "which are way, way, way more efficient, 3X, 5X more performance per watt than the prior generation chips",
   "source": "Dylan Patel"
  },
  {
   "statement": "A gigawatt of Vera Rubin compute requires roughly 55,000 N3 wafers, 6,000 N5 wafers and 170,000 DRAM wafers.",
   "subjects": [
    "NVIDIA",
    "TSMC"
   ],
   "value": "55k N3 + 6k N5 + 170k DRAM wafers per GW",
   "timestamp": "7:14",
   "quote": "you said in order to make a gigawatt of, I think Vera Rubin's, you need 55,000 N3 wafers, 6,000 N5 wafers and 170,000 DRAM wafers",
   "source": "Dwarkesh Patel"
  },
  {
   "statement": "The supply chain is currently on track for about 100 ASML EUV tools a year by 2030.",
   "subjects": [
    "ASML",
    "Carl Zeiss"
   ],
   "value": "~100 tools/year by 2030",
   "timestamp": "11:51",
   "quote": "A hundred ASML tools for 2030",
   "source": "Dylan Patel"
  },
  {
   "statement": "Compute has been transacting at $10-15 million per megawatt and that price is starting to inflect upward.",
   "subjects": [
    "AI compute"
   ],
   "value": "$10-15M per MW",
   "timestamp": "13:39",
   "quote": "this is sort of already led to this compute pricing $10 to $15 million per megawatt, start to inflect up.",
   "source": "Dylan Patel"
  },
  {
   "statement": "Meta was at one point rumored to be as much as 10% of Anthropic's business.",
   "subjects": [
    "Meta",
    "Anthropic"
   ],
   "value": "~10% of Anthropic revenue",
   "timestamp": "18:23",
   "quote": "Meta, who at one point was rumored to be as much as 10 percent of Anthropic's business",
   "source": "Dylan Patel"
  },
  {
   "statement": "SpaceX sold a gigawatt of compute to Google for about $40 billion, far above the prevailing rate.",
   "subjects": [
    "SpaceX",
    "Google",
    "AI compute"
   ],
   "value": "$40B per GW",
   "timestamp": "22:18",
   "quote": "the SpaceX sold for $40 billion a gigawatt to Google",
   "source": "Dwarkesh Patel"
  },
  {
   "statement": "About 70% of AI data-center watts are deployed in America, versus under 10% in China.",
   "subjects": [
    "United States",
    "China",
    "AI compute"
   ],
   "value": "US ~70% / China <10%",
   "timestamp": "34:34",
   "quote": "today, 70 percent of watts are being deployed in America",
   "source": "Dylan Patel"
  },
  {
   "statement": "In 2022 the US was adding 45-50% of world compute and China 30-35%, before the export-control regime.",
   "subjects": [
    "United States",
    "China",
    "Export controls"
   ],
   "value": "US 45-50% / China 30-35% (2022)",
   "timestamp": "34:34",
   "quote": "the US was adding about 45-50 percent of the world's compute. China was adding about 30-35 percent of the world's compute",
   "source": "Dylan Patel"
  },
  {
   "statement": "The leading Chinese labs have at most 100-200 megawatts of compute in total, with ByteDance Seed the outlier.",
   "subjects": [
    "China",
    "AI compute"
   ],
   "value": "100-200 MW total",
   "timestamp": "39:25",
   "quote": "The leading Chinese labs have 100, 200 megawatts total of compute at most.",
   "source": "Dylan Patel"
  },
  {
   "statement": "A lab's compute budget has historically split roughly 50% research, 10% development, 40% inference.",
   "subjects": [
    "Anthropic",
    "OpenAI",
    "AI compute"
   ],
   "value": "50% research / 10% dev / 40% inference",
   "timestamp": "39:25",
   "quote": "it's actually like 50 percent of the compute is research, like 10 percent of the compute is development, and then 40 percent is inference.",
   "source": "Dylan Patel"
  },
  {
   "statement": "Anthropic's pre-training run for Mythos used under 200 megawatts for about two months, far less than its total fleet.",
   "subjects": [
    "Anthropic",
    "Mythos"
   ],
   "value": "<200 MW for ~2 months",
   "timestamp": "41:30",
   "quote": "It's sub 200 megawatts for call it two months.",
   "source": "Dylan Patel"
  },
  {
   "statement": "Anthropic is willing to pay around 20% rates on incremental debt to build capacity, because renting from SpaceX is more expensive.",
   "subjects": [
    "Anthropic",
    "SpaceX",
    "Interest rates"
   ],
   "value": "~20% rate",
   "timestamp": "44:48",
   "quote": "Anthropics willing to pay 20 percent rates for the incremental billion dollars to build their capacity",
   "source": "Dylan Patel"
  },
  {
   "statement": "Corporate income tax is under 10% of US federal revenue while payroll and income taxes are over 80%.",
   "subjects": [
    "United States"
   ],
   "value": "<10% corporate / >80% payroll+income",
   "timestamp": "47:27",
   "quote": "corporate income is less than 10 percent of federal revenues, and 80 percent plus is payroll taxes and income taxes",
   "source": "Dwarkesh Patel"
  },
  {
   "statement": "About 20% of US tax revenue currently goes to servicing the national debt.",
   "subjects": [
    "United States",
    "Interest rates"
   ],
   "value": "~20% of tax revenue",
   "timestamp": "47:27",
   "quote": "currently, 20 percent of tax revenue spending goes towards servicing the debt",
   "source": "Dwarkesh Patel"
  },
  {
   "statement": "Meta has recently raised debt at roughly 5-6%.",
   "subjects": [
    "Meta",
    "Interest rates"
   ],
   "value": "5-6%",
   "timestamp": "56:48",
   "quote": "recently meta's raised at like 5 to 6 percent",
   "source": "Dylan Patel"
  },
  {
   "statement": "In the 1980s Paul Volcker pushed real interest rates to around 8%, causing roughly 40 countries — mostly Latin American — to default that decade.",
   "subjects": [
    "Paul Volcker",
    "Sovereign debt crisis"
   ],
   "value": "~8% real rates, ~40 defaults",
   "timestamp": "58:44",
   "quote": "And that caused some 40 different countries, mostly Latin America, to default in that decade.",
   "source": "Dwarkesh Patel"
  },
  {
   "statement": "Meta is valued around $1.5 trillion, which Patel argues undervalues its compute hoard.",
   "subjects": [
    "Meta"
   ],
   "value": "~$1.5T market cap",
   "timestamp": "1:01:28",
   "quote": "they're like $1.5 trillion company",
   "source": "Dylan Patel"
  },
  {
   "statement": "Effective AI population at the frontier grows roughly 10x a year — frontier flops up 4-5x while compute needed per capability falls ~3x.",
   "subjects": [
    "AI compute",
    "AGI"
   ],
   "value": "~10x/year",
   "timestamp": "1:07:36",
   "quote": "the compute at the frontier, basically the effective AI population size at the frontier labs is increasing 10 X year over year",
   "source": "Dwarkesh Patel"
  }
 ],
 "notable_quotes": [
  {
   "speaker": "Dylan Patel",
   "timestamp": "3:50",
   "quote": "it's really by the end of next year, it's already half of the incremental compute is going to Anthropic and OpenAI.",
   "why": "The headline number behind the episode's title — centralization arriving a year earlier than most models assume."
  },
  {
   "speaker": "Dylan Patel",
   "timestamp": "5:52",
   "quote": "you've got them just controlling most of the usable flops in the world on their own.",
   "why": "The endpoint of the compute-concentration thesis, stated plainly."
  },
  {
   "speaker": "Dylan Patel",
   "timestamp": "30:30",
   "quote": "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?",
   "why": "The capital-allocation choice at the heart of his non-consensus call that inference share falls."
  },
  {
   "speaker": "Dylan Patel",
   "timestamp": "1:01:28",
   "quote": "the limiter on AGI is not how fast can the research engineers like our roommate Sholto can crank the gears. It's actually just like, how much does the rest of the world let that happen?",
   "why": "Reframes the takeoff debate from model capability to political economy — the episode's central argument."
  },
  {
   "speaker": "Dwarkesh Patel",
   "timestamp": "51:31",
   "quote": "Those countries like Pakistan or Nigeria or something, I think are just going to be very fucked in this new interest rate regime.",
   "why": "The sovereign-debt spillover of AI's returns, in one line."
  },
  {
   "speaker": "Dylan Patel",
   "timestamp": "1:12:47",
   "quote": "I don't trust the government and I don't trust Dario and I don't trust Sam.",
   "why": "The governance dilemma stated without hedging — no available custodian for concentrated AI power."
  }
 ],
 "extracted_by": "claude-opus-5",
 "prompt_version": "episode-extract-v1",
 "_file": "dwarkesh-podcast-2026-08-25-dylan-patel-anthropic-openai-will-have-most-of-the-world-s-c.json"
}