{
 "episode_id": "invest-like-the-best-2026-05-20-gavin-baker-watts-and-wafers",
 "title": "Gavin Baker - Watts and Wafers",
 "podcast": "invest-like-the-best",
 "published_at": "2026-05-20",
 "speakers": [
  "Patrick O'Shaughnessy",
  "Gavin Baker"
 ],
 "summary": "In his sixth conversation with Patrick O'Shaughnessy, Atreides Management CIO Gavin Baker frames the next phase of AI around two physical constraints — Watts (power) and Wafers (chip fabrication) — arguing the near-term power shortage will begin to ease in 2027-2028 and that SpaceX's 'orbital compute' (Blackwell-sized racks in sun-synchronous orbit, linked by lasers) will solve it long-term. He contends the current AI buildout differs fundamentally from the 2000 dot-com bubble because it is funded out of operating cash flow with GPUs at 100% utilization, and that TSMC's wafer-capacity discipline is the single most important variable for whether a bubble forms. Baker also discusses Elon Musk's Terrafab fab joint venture, the disaggregation of GPUs (prefill vs. decode) extending GPU useful lives to 10-15 years and rescuing private credit, why new chip startups must do something 'different and hard' rather than build a better GPU, and why economic value keeps accruing to frontier model tokens while the application layer has net destroyed trillions in value. He closes worried about a 'diversity breakdown' (everyone bullish), cross-sectional valuation dislocations (SemiCap vs. DRAM multiples), rising personal-safety and geopolitical risk from AI, and yet remains an 'AI optimist maximalist' citing a rare-disease cure enabled by agents.",
 "topics": [
  {
   "name": "Drawdowns, March/April Selloff, and the Strait of Hormuz",
   "timestamp": "3:45",
   "summary": "Baker distinguishes drawdowns where you're wrong from periods of underperformance where you can lean in, calling March 2026 the latter; he argues the AI inflection (Anthropic adding $11B ARR in a month) was the most extraordinary moment in capitalist history, and that a closed Strait of Hormuz was paradoxically good for America's relative manufacturing competitiveness."
  },
  {
   "name": "Anthropic vs. OpenAI Capital Efficiency and Valuations",
   "timestamp": "9:47",
   "summary": "Baker compares OpenAI and Anthropic as different animals on capital efficiency, argues Anthropic is compute-constrained (could be at $100-200B ARR unconstrained), and explains why not raising at maximum valuation is wise, using Elon's decade of underpricing SpaceX as the model."
  },
  {
   "name": "Watts: Power Shortage, Turbines, and Orbital Compute",
   "timestamp": "14:24",
   "summary": "Baker argues capitalism will solve the Watts shortage by 2027-2028 (with zoning/approval now the bigger gate than energy or chips), and reframes 'orbital compute' as Blackwell-sized racks in sun-synchronous orbit cooled by radiators and linked by lasers, which he believes SpaceX has solved and which will dominate inference long-term."
  },
  {
   "name": "Wafers, TSMC's Capacity Discipline, and the Bubble Question",
   "timestamp": "24:00",
   "summary": "Baker frames TSMC as the gatekeeper preventing an AI bubble through wafer scarcity, citing Carlota Perez and historical bubble precedent; he argues this buildout differs from 2000 (operating-cash-funded, 100% GPU utilization) and names TSMC's capacity decisions the single thing to watch."
  },
  {
   "name": "Elon's Terrafab and Bringing Fab Capacity to America",
   "timestamp": "29:00",
   "summary": "Baker describes the SpaceX/Tesla Terrafab joint venture (partnered with Intel) to build the world's largest fab in America, predicting success via Elon's hardware reputation, the semicap 'A-teams,' and aggressive talent recruitment ('Taiwan town' next to the fab)."
  },
  {
   "name": "Returns to the Frontier, the Bitter Lesson, and Usage-Based Pricing",
   "timestamp": "31:17",
   "summary": "Baker explores why economic returns accrue overwhelmingly to frontier tokens, names a violation of Sutton's bitter lesson the biggest risk to the AI trade, dismisses TurboQuant fears, and explains the shift from all-you-can-eat to usage-based pricing as deeply bullish for lab revenue."
  },
  {
   "name": "Continual Learning and Harnesses",
   "timestamp": "35:45",
   "summary": "Baker discusses harness engineering as a meaningful (if secondary) lever co-developed with models, and frames continual learning — a model dynamically updating its weights in real time — as a third pivotal open question that, if solved, produces a fast takeoff."
  },
  {
   "name": "New Chip Companies: Different and Hard (Cerebras, Trainium, AMD)",
   "timestamp": "39:59",
   "summary": "Baker uses the tank-design 'iron triangle' analogy to argue chip startups must do something different AND hard rather than build a better GPU; he praises Cerebras's wafer-scale bet, ranks Trainium ahead of AMD's MI450 among GPU challengers, and explains prefill/decode disaggregation as a richer design canvas."
  },
  {
   "name": "GPU Useful Lives, Private Credit, and Sellers of Shortage",
   "timestamp": "46:34",
   "summary": "Baker argues prefill/decode disaggregation extends GPU useful lives to 10-15 years, which lowers financing costs and may 'single-handedly save private credit,' and notes that owners of installed bases (sellers of shortage) and CPU fleets are advantaged in an agentic world."
  },
  {
   "name": "AI-Native Founders, the Application Layer, and the Token Path",
   "timestamp": "49:03",
   "summary": "Baker argues value is accruing to energy/data centers/chips/models but not applications, that coding-focused companies (Cursor, Cognition) won by getting to scale, and that founders must be 'in the token path' or in a defensible niche to survive the model companies."
  },
  {
   "name": "Open-Source Frontier, Anti-Distillation, and Game Theory",
   "timestamp": "52:42",
   "summary": "Baker argues Jensen could approach the frontier with his own model whenever he wants, that Chinese open-source models rely on distilled American tokens, and lays out a new prisoner's-dilemma game theory over whether frontier labs release models via API."
  },
  {
   "name": "Cross-Sectional Valuation Dislocations and the Diversity Breakdown",
   "timestamp": "1:00:06",
   "summary": "Baker says he's worried about a 'diversity breakdown' (everyone bullish) and that valuations cannot all be true — SemiCap at 40x vs. DRAM at mid-single digits — with the lowest-quality commodity suppliers outperforming during the shortage, while he wishes for more AI and memory bears."
  },
  {
   "name": "Big-Cap Riffs and AI Price Efficiency",
   "timestamp": "1:05:52",
   "summary": "Baker riffs on the megacaps: Google strong on installed base despite losing its TPU edge, Meta credited for becoming AI-first, Amazon strong via Trainium, and Microsoft's Satya making courageous but punished bets to use compute internally rather than serve OpenAI."
  },
  {
   "name": "Knock-On Effects: Personal Safety, Geopolitics, Pax Americana, and AI Optimism",
   "timestamp": "1:12:02",
   "summary": "Baker reflects on AI destroying application-layer value, rising personal-safety and political-violence risk, Ukraine winning via battlefield AI, a possible AI-informed Pax Americana, and his maximalist optimism citing a rare-disease cure enabled by agents."
  }
 ],
 "entities": [
  {
   "name": "Gavin Baker",
   "type": "person",
   "salience": "central",
   "role_in_episode": "Founding partner and CIO of Atreides Management and the guest; holder of nearly every substantive claim and prediction in the episode."
  },
  {
   "name": "Patrick O'Shaughnessy",
   "type": "person",
   "salience": "central",
   "role_in_episode": "Host of Invest Like the Best; frames the Watts-and-Wafers theme and elicits Baker's views via questions."
  },
  {
   "name": "TSMC",
   "type": "company",
   "salience": "central",
   "role_in_episode": "Taiwan Semiconductor, framed as the wafer gatekeeper whose capacity discipline is the single most important variable determining whether an AI bubble forms; works with Jensen on a handshake, no contract."
  },
  {
   "name": "NVIDIA",
   "type": "company",
   "salience": "central",
   "role_in_episode": "The dominant GPU maker; Baker argues it gets the best TSMC prices, would foreclose any easy chip trade-off, and could sell $2-3T of GPUs if TSMC fully supplied it."
  },
  {
   "name": "SpaceX",
   "type": "company",
   "salience": "central",
   "role_in_episode": "Central to the orbital-compute thesis (racks in space, world's largest satellite fleet) and a Terrafab joint-venture party; Baker calls its engineers the most talented on Earth and it the most important defense contractor in America."
  },
  {
   "name": "Anthropic",
   "type": "company",
   "salience": "central",
   "role_in_episode": "Cited as the proof of AI's extraordinary inflection (added $11B ARR in a month), as compute-constrained relative to its potential, and as more capital-efficient than OpenAI."
  },
  {
   "name": "Elon Musk",
   "type": "person",
   "salience": "discussed",
   "role_in_episode": "Held up as the model fundraiser (never greedy on valuation), the driver behind Terrafab and orbital compute, and, in Baker's view, doing more for America than any other American."
  },
  {
   "name": "OpenAI",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Compared unfavorably to Anthropic on capital efficiency/cost-per-token; cited as a reference frontier asset and as a party to the open-source game theory."
  },
  {
   "name": "Terrafab",
   "type": "organization",
   "salience": "discussed",
   "role_in_episode": "SpaceX/Tesla joint venture (partnered with Intel) to build the world's largest fab in America; Baker predicts it will succeed and attract the semicap A-teams and top talent."
  },
  {
   "name": "Cerebras",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Atreides venture investment; held up as the model of a chip startup doing something 'different and hard' (wafer-scale computing) that took three chip generations to get right."
  },
  {
   "name": "Trainium",
   "type": "product",
   "salience": "discussed",
   "role_in_episode": "Amazon's AI chip; ranked by Baker as the best-performing GPU challenger today, with Trainium 3's switch scale-up network needed to economically inference MoE models."
  },
  {
   "name": "Google",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Lost its per-token cost leadership via conservative TPUv8 design decisions but remains strong on the largest installed base of compute and YouTube data."
  },
  {
   "name": "Microsoft",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Satya credited for courageous (market-punished) bets to use compute internally rather than serve OpenAI; Baker is skeptical Microsoft has the team to train frontier models."
  },
  {
   "name": "Amazon",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Positioned strongly via Trainium and robotics P&L efficiencies; named one of the two companies (with NVIDIA) most deeply engaged with startups."
  },
  {
   "name": "Meta",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Credited by Baker, via Zuckerberg, for becoming a genuinely AI-first company internally; its MSL model 'Muse' was a positive surprise though not on the Pareto frontier."
  },
  {
   "name": "orbital compute",
   "type": "concept",
   "salience": "central",
   "role_in_episode": "Baker's reframing of data centers in space as Blackwell-sized racks in sun-synchronous orbit linked by lasers, which he argues will solve the long-term power constraint and dominate inference."
  },
  {
   "name": "the bitter lesson",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Richard Sutton's principle that compute beats human algorithmic ingenuity; Baker names a violation of it the single biggest risk to the entire AI trade."
  },
  {
   "name": "frontier tokens",
   "type": "concept",
   "salience": "central",
   "role_in_episode": "The locus of the central open question: whether frontier model tokens will keep capturing the overwhelming majority of economic value at the model layer."
  },
  {
   "name": "prefill/decode disaggregation",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The separation of memory-capacity-bound prefill from bandwidth-bound decode; framed as both a richer chip-design canvas and the reason GPU useful lives extend to 10-15 years."
  },
  {
   "name": "continual learning",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "A model dynamically updating its weights in real time; framed as the third pivotal open question that, if solved, produces a fast AI takeoff."
  },
  {
   "name": "private credit",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "In pain from SaaS and GPU loans; Baker argues extended GPU useful lives could 'single-handedly save' it by lowering financing costs."
  },
  {
   "name": "Grok 4.3",
   "type": "product",
   "salience": "mentioned",
   "role_in_episode": "xAI model Baker uses heavily; described as on the Pareto frontier and the best low-cost 500-billion-parameter model."
  },
  {
   "name": "Astera",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "A long-held Atreides position (first bought in the Series C); held up as a 'miscategorized' switch company wrongly placed in copper-loser baskets."
  }
 ],
 "claims": [
  {
   "holder": "Gavin Baker",
   "claim": "The AI inflection of early 2026 — Anthropic adding $11B of ARR in a single month — is the most extraordinary moment in the history of capitalism, with no prior precedent.",
   "stance": "opinion",
   "conviction": 0.9,
   "subjects": [
    "Anthropic"
   ],
   "timestamp": "3:45",
   "quote": "what was happening in AI was, I think, the most extraordinary moment in the history of capitalism, the history of American business."
  },
  {
   "holder": "Gavin Baker",
   "claim": "The closing of the Strait of Hormuz was relatively good for America because it improved US relative manufacturing competitiveness (cheap domestic natural gas vs. doubled/tripled prices abroad), aligning with the Trump administration's focus on relative position.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "Strait of Hormuz",
    "US manufacturing competitiveness"
   ],
   "timestamp": "5:55",
   "quote": "the Strate of Formos being closed is actually relatively awesome for America. Why? Particularly for the goals of the current administration."
  },
  {
   "holder": "Gavin Baker",
   "claim": "AI/tech became as cheap relative to the rest of the market as at any point in the last decade precisely when the AI fundamentals were most bullish, a striking market inefficiency.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "NVIDIA"
   ],
   "timestamp": "5:55",
   "quote": "tech essentially got as cheap as it's been versus the rest of the market, has at any point over the last 10 years."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Anthropic is materially compute-constrained: if it could get all the compute it wanted, it would be doing roughly $100-200B in ARR today rather than ~$50B, so it is effectively trading at ~5x unconstrained revenue.",
   "stance": "opinion",
   "conviction": 0.6,
   "subjects": [
    "Anthropic"
   ],
   "timestamp": "10:41",
   "quote": "if Anthropic could just wave a magic wand and get all the compute they wanted, they'd probably be doing well north of $100 billion today, maybe $150 billion."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Never being greedy on valuation is a durable 'superpower' for a founder: by consistently making investors money (Elon for 20 years), you can raise as much capital as you want whenever you want, a benefit that lasts decades.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "Elon Musk",
    "SpaceX"
   ],
   "timestamp": "12:07",
   "quote": "because he's made people money for now 20 years, he has a superpower. That is, he can essentially raise as much capital as he wants whenever he wants."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Capitalism will solve the Watts (power) shortage, with the bigger gating factor now being zoning and approval rather than energy or chips, absent regulatory/political blowback.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "orbital compute"
   ],
   "timestamp": "14:24",
   "quote": "I think capitalism is going to solve the Watts shortage, absent big regulatory political blowback, which I think is a real possibility."
  },
  {
   "holder": "Gavin Baker",
   "claim": "The near-term power shortage will begin to alleviate in 2027-2028 as new energy sources come online.",
   "stance": "prediction",
   "conviction": 0.6,
   "subjects": [
    "orbital compute"
   ],
   "timestamp": "14:24",
   "quote": "I think the Watts shortage will probably begin to alleviate 27, 28"
  },
  {
   "holder": "Gavin Baker",
   "claim": "Orbital compute is real and viable — SpaceX engineers, the most talented on Earth, are confident they have solved the radiator/cooling and maintenance problems — and it will ultimately solve the terrestrial power constraint.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "SpaceX",
    "orbital compute"
   ],
   "timestamp": "16:39",
   "quote": "I do think it is the most talented group of engineers on planet Earth. And they are very confident they have solved this."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Inference is well suited to orbital compute, but training will remain on Earth for a long time, and terrestrial data centers will stay valuable for decades — so orbital compute is not bearish for them.",
   "stance": "prediction",
   "conviction": 0.7,
   "subjects": [
    "orbital compute"
   ],
   "timestamp": "21:01",
   "quote": "Inference, I think, is very sensible for orbo compute. Training will be done on Earth for a long time. I don't think that this is super bearish for terrestrial data centers."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Companies in the power-production and cooling ecosystem that are aggressively ramping capacity should think hard about the risk that their ramps hit just as orbital compute becomes real.",
   "stance": "opinion",
   "conviction": 0.5,
   "subjects": [
    "orbital compute"
   ],
   "timestamp": "21:01",
   "quote": "a lot of these capacity ramps are going to be hitting. Just as I think all of the silly skeptics start to understand that orbital compute is very real."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Based on 200-300 years of foundational-technology precedent (railroads, canals, South Sea, internet), an AI bubble should be expected — though this buildout is fundamentally different because it is funded out of operating cash flow with GPUs at full utilization.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "TSMC"
   ],
   "timestamp": "24:00",
   "quote": "based on the last 200 or 300 years, forget the internet bubble, we had a railroad bubble, canal bubble, every kind of bubble, South Sea bubble, we should expect a bubble. That's terrifying."
  },
  {
   "holder": "Gavin Baker",
   "claim": "TSMC's wafer scarcity is what has prevented an AI bubble so far; if it fully supplied NVIDIA, demand could overbuild, so a Goldilocks zone of capacity discipline is needed to avoid a crash.",
   "stance": "assertion",
   "conviction": 0.75,
   "subjects": [
    "TSMC",
    "NVIDIA"
   ],
   "timestamp": "24:00",
   "quote": "If Taiwan Semi did what Jensen wanted, I think Nvidia could sell $2 trillion of GPUs in 26 or 27, maybe 2.5 trillion, maybe 3 trillion. But there is a limit where consumers would consume so much, they probably would be in an overbuild."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Intel and Samsung will eventually break ramp discipline, which would force everyone else to break it too — so the key variable to watch for a bubble is TSMC's ability to maintain its leading-node lead and its capacity decisions.",
   "stance": "prediction",
   "conviction": 0.6,
   "subjects": [
    "TSMC"
   ],
   "timestamp": "26:52",
   "quote": "If I were to watch one thing to understand where there is a bubble, it's Taiwan CME's capacity decisions."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Elon's Terrafab will be successful, aided by its Intel partnership (50 years of institutional knowledge), the semicap 'A-teams' who want a second source, and aggressive talent recruitment, and is different enough that it won't alienate TSMC.",
   "stance": "prediction",
   "conviction": 0.7,
   "subjects": [
    "Terrafab",
    "Elon Musk",
    "Intel"
   ],
   "timestamp": "26:52",
   "quote": "It's a SpaceX, I believe Tesla's involved as well, joint venture to build the world's largest fab here in America. I think they're going to be successful."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Elon Musk is probably doing more for America than any other living American — reshoring manufacturing, reviving defense tech, decarbonizing — but his foray into politics makes it hard for some to see this clearly.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "Elon Musk"
   ],
   "timestamp": "26:52",
   "quote": "which is sad because I do think he's probably doing more for America than any other American."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Whether frontier tokens will keep capturing the overwhelming majority of economic value at the model layer is one of the most important open questions for an AI investor, and Baker is now more open-minded that they may not.",
   "stance": "opinion",
   "conviction": 0.6,
   "subjects": [
    "frontier tokens"
   ],
   "timestamp": "31:17",
   "quote": "Are frontier tokens going to continue capturing the overwhelming majority of economic value created at the model layer?"
  },
  {
   "holder": "Gavin Baker",
   "claim": "A violation of Richard Sutton's bitter lesson (human algorithmic ingenuity beating raw compute) is the single biggest risk to the entire AI trade; people closest to AI are the most skeptical it will happen.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "the bitter lesson"
   ],
   "timestamp": "31:17",
   "quote": "a violation of Richard Sutton's bitter lesson is for sure the biggest risk to this trade, to all of AI."
  },
  {
   "holder": "Gavin Baker",
   "claim": "We are very close to ASI, and it is uncertain whether the bitter lesson holds for 300-600 IQ models — a temporary violation could occur as an ASI makes itself more efficient.",
   "stance": "opinion",
   "conviction": 0.5,
   "subjects": [
    "the bitter lesson"
   ],
   "timestamp": "31:17",
   "quote": "The reason I am a little less skeptical is I think we're very close to ASI. And who knows if the bitter lesson holds for 400 IQ models?"
  },
  {
   "holder": "Gavin Baker",
   "claim": "Google's TurboQuant memory optimization, hyped on X as a DRAM-demand killer (a 'more stupid DeepSeek'), will have no real impact on DRAM demand — no AI engineer Baker found believed it would.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "Google"
   ],
   "timestamp": "31:17",
   "quote": "I was unable to find a single AI engineer on planet Earth who believed that TurboQuant would have any impact on DRAM demand."
  },
  {
   "holder": "Gavin Baker",
   "claim": "The shift from all-you-can-eat to usage-based ('pay by the drink') pricing is deeply bullish for AI labs and is why OpenAI and Anthropic will exceed well over $200B in ARR this year — mirroring how cellular/long-distance grew under metered pricing.",
   "stance": "prediction",
   "conviction": 0.7,
   "subjects": [
    "OpenAI",
    "Anthropic",
    "frontier tokens"
   ],
   "timestamp": "36:20",
   "quote": "I think the shift to usage based pricing is probably why you will see OpenAI and Anthropic exceed well over $200 billion in ARR this year."
  },
  {
   "holder": "Gavin Baker",
   "claim": "To understand frontier AI's true capability today you must be on an enterprise usage-based plan with Claude Code or Codex; fixed-price plans give a rate-limited, 'lobotomized' version that will mislead investors' intuition.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "frontier tokens"
   ],
   "timestamp": "36:20",
   "quote": "To understand what Frontier AI is capable of today, even for a non-coding use case, you need to have Cloud Code or Codex."
  },
  {
   "holder": "Gavin Baker",
   "claim": "If continual learning (a model dynamically updating its weights in real time) is solved, it produces a really fast AI takeoff; people seem confident it is just around the corner.",
   "stance": "prediction",
   "conviction": 0.5,
   "subjects": [
    "continual learning"
   ],
   "timestamp": "38:37",
   "quote": "if we get that, then we have a really fast takeoff. And people seem confident that continual learning is kind of just around the corner."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Chip startups should not try to build a better GPU; the only viable path is to do something different AND hard, because NVIDIA gets the best TSMC prices and will fast-follow any easy trade-off.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "NVIDIA",
    "TSMC"
   ],
   "timestamp": "40:21",
   "quote": "if you're trying to make a better GPU, good luck. If you are doing something different, it also needs to be hard to do."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Among GPU challengers, Trainium is doing the best today (Trainium 3 has the switch scale-up network needed to economically inference MoE models), while AMD's MI450 is unproven.",
   "stance": "assertion",
   "conviction": 0.65,
   "subjects": [
    "Trainium"
   ],
   "timestamp": "40:21",
   "quote": "And today I think probably Tranium is doing the best. Nobody's a better GPU."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Cerebras did something genuinely different and hard — wafer-scale computing — and that architectural decision lets it do things no one else can; it can theoretically run any size model, though it took three chip generations to get right.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "Cerebras"
   ],
   "timestamp": "42:51",
   "quote": "What Cerebris has done is something hard and fundamentally different. Wafer scale computing, it comes with a set of trade-offs, but that architectural decision they made was hard, and let's them do something that no one else can do."
  },
  {
   "holder": "Gavin Baker",
   "claim": "The disaggregation of prefill and inference will extend GPU useful lives to 10-15 years (not the 1-4 years skeptics claim), since older GPUs can be repurposed for prefill behind newer decode systems.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "prefill/decode disaggregation",
    "NVIDIA"
   ],
   "timestamp": "46:36",
   "quote": "The disaggregation of inference means that I think these GPUs are going to have 10 or 15 year lives."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Extended GPU useful lives may single-handedly save private credit, because longer-lived GPUs and lower financing rates (5-6% vs. low-7s) materially change the cost of financing the AI buildout.",
   "stance": "prediction",
   "conviction": 0.6,
   "subjects": [
    "private credit",
    "prefill/decode disaggregation"
   ],
   "timestamp": "45:00",
   "quote": "this is going to be amazing for the useful lives of GPUs and may single-handedly save private credit."
  },
  {
   "holder": "Gavin Baker",
   "claim": "The right venture test is whether an idea is non-obvious, different, AND hard to execute — if it becomes obvious to the world before you reach scale, you're in trouble; the model companies are capturing value at energy/data centers/chips/models, not applications.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "frontier tokens"
   ],
   "timestamp": "49:03",
   "quote": "The profits are accruing to energy, accruing to data centers, accruing to chips, they're accruing to models, not really accruing to the applications"
  },
  {
   "holder": "Gavin Baker",
   "claim": "A software/AI company must be 'in the token path' (or in a truly niche, defensible area), otherwise life may be hard — and even vertical-niche data moats are dubious because the data comes from humans and frontier labs may undercut them.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "frontier tokens"
   ],
   "timestamp": "51:29",
   "quote": "if you're a software company or an AI company of any kind, you have to be in the token path."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Jensen/NVIDIA could probably get pretty close to the frontier with its own model whenever it wanted; it is the logical counter-move to what OpenAI and Anthropic are trying to do to NVIDIA.",
   "stance": "opinion",
   "conviction": 0.6,
   "subjects": [
    "NVIDIA",
    "OpenAI",
    "Anthropic"
   ],
   "timestamp": "52:41",
   "quote": "I have a belief that whenever he wants, Jensen can probably get pretty close to the Frontier."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Today's open-source frontier consists of Chinese models trained on distilled/stolen American tokens; American labs' anti-distillation work and a new prisoner's-dilemma game theory will keep open source a fixed timeframe behind the frontier.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "NVIDIA"
   ],
   "timestamp": "52:42",
   "quote": "You will see that open source Frontier, which today consists of Chinese models with stolen American tokens."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Cross-sectional AI valuations cannot all be true: SemiCap equipment companies trade at ~40x next-quarter annualized earnings while DRAM companies trade at mid-single digits, a multiple gap not justified by their relative business models.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "NVIDIA"
   ],
   "timestamp": "1:00:06",
   "quote": "the valuations do not make sense. They just flat out do not make sense. They cannot all be true."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Because we are in shortages, the lowest-quality companies (high-cost, unreliable commodity suppliers) are doing the best and getting bid up by retail on X — a classic commodity-cycle dynamic that worries Baker because those gains reverse on the downturn.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "commodity cycle"
   ],
   "timestamp": "1:00:06",
   "quote": "Because we are in shortages, the lowest quality companies are doing the best."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Baker is worried about a 'diversity breakdown' — everyone has become bullish on AI / DRAM and there are too few bears — which historically precedes a bubble.",
   "stance": "opinion",
   "conviction": 0.65,
   "subjects": [
    "AI bubble",
    "DRAM"
   ],
   "timestamp": "1:00:06",
   "quote": "I don't know anyone like me who's not really bullish on DRAM."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Google remains in a great position despite losing its TPU cost advantage because it has the largest installed base of compute plus uniquely valuable YouTube/search data.",
   "stance": "opinion",
   "conviction": 0.75,
   "subjects": [
    "Google"
   ],
   "timestamp": "1:05:52",
   "quote": "Google was incredible last year because they had that TPU advantage, which is now gone. The reason I think they're still in a great position is just they have the most compute of everyone."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Satya/Microsoft is making the right but market-punished, courageous bet to use its compute internally to improve its own products rather than just selling GPUs to OpenAI; Baker is skeptical Microsoft has the team to train frontier models.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "Microsoft"
   ],
   "timestamp": "1:05:52",
   "quote": "the decision Satya is making, which the market has punished him for, but I think is the right decision."
  },
  {
   "holder": "Gavin Baker",
   "claim": "AMD, Microsoft, and Meta show essentially zero engagement with chip startups (unlike NVIDIA, Amazon, Google, Broadcom), which Baker thinks will become a meaningful disadvantage because the best teams are now at startups.",
   "stance": "opinion",
   "conviction": 0.65,
   "subjects": [
    "NVIDIA",
    "Amazon",
    "Google"
   ],
   "timestamp": "1:05:52",
   "quote": "you see essentially zero engagement with startups from AMD, Microsoft, and Meta."
  },
  {
   "holder": "Gavin Baker",
   "claim": "AI has net destroyed trillions of dollars of value at the application layer — even counting Cursor and Cognition — and the companies winning are those with the highest ratio of utilized GPUs per human.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "frontier tokens"
   ],
   "timestamp": "1:12:02",
   "quote": "AI has net destroyed, even if you count Cursor, Cognition, the most successful AI natives, trillions of dollars of value has been destroyed by AI at the application layer."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Baker is increasingly worried about personal safety, especially for high-profile people associated with AI, as political violence rises and AI becomes more political.",
   "stance": "opinion",
   "conviction": 0.7,
   "subjects": [
    "personal safety",
    "political violence"
   ],
   "timestamp": "1:12:02",
   "quote": "I am more and more worried about personal safety. And I worry about this a lot more for people who have a much bigger public presence and are much more associated with AI."
  },
  {
   "holder": "Gavin Baker",
   "claim": "Ukraine is really starting to win the war primarily because it has the best battlefield AI outside America and Israel, and how adversaries respond to America's AI edge is a destabilizing geopolitical question.",
   "stance": "opinion",
   "conviction": 0.6,
   "subjects": [
    "Ukraine war",
    "battlefield AI"
   ],
   "timestamp": "1:12:02",
   "quote": "I think the reason Ukraine is really winning is they have the best battlefield AI outside of probably America and Israel."
  },
  {
   "holder": "Gavin Baker",
   "claim": "It is dystopian that the best AI is now only available to those with a lot of money, and society needs to solve that; Baker remains an AI optimist maximalist despite acknowledging AI is a discontinuity to navigate.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "frontier tokens"
   ],
   "timestamp": "1:14:52",
   "quote": "It is a little dystopian that now the best AI is only available to people with a lot of money. We need to solve that."
  }
 ],
 "predictions": [
  {
   "holder": "Gavin Baker",
   "prediction": "The near-term AI power (Watts) shortage will begin to ease in 2027-2028 as new energy sources come online.",
   "timeframe": "2027-2028",
   "subjects": [
    "orbital compute"
   ],
   "timestamp": "14:24",
   "quote": "I think the Watts shortage will probably begin to alleviate 27, 28",
   "due": "2028-12-31"
  },
  {
   "holder": "Gavin Baker",
   "prediction": "Orbital compute (Blackwell-sized racks in sun-synchronous orbit linked by lasers) will become real and ultimately solve the long-term power constraint, dominating inference.",
   "timeframe": "long-term",
   "subjects": [
    "SpaceX",
    "orbital compute"
   ],
   "timestamp": "14:24",
   "quote": "And then I think orbital compute will really solve that."
  },
  {
   "holder": "Gavin Baker",
   "prediction": "If TSMC fully supplied NVIDIA's demand, NVIDIA could sell $2-3 trillion of GPUs in 2026 or 2027.",
   "timeframe": "2026-2027",
   "subjects": [
    "NVIDIA",
    "TSMC"
   ],
   "timestamp": "24:00",
   "quote": "If Taiwan Semi did what Jensen wanted, I think Nvidia could sell $2 trillion of GPUs in 26 or 27, maybe 2.5 trillion, maybe 3 trillion.",
   "due": "2027-12-31"
  },
  {
   "holder": "Gavin Baker",
   "prediction": "One of Intel or Samsung will eventually break fab capacity/pricing discipline, forcing everyone else to break it too.",
   "timeframe": "unspecified",
   "subjects": [
    "TSMC"
   ],
   "timestamp": "26:27",
   "quote": "one of Intel and Samsung, they are not going to stay disciplined. They will break and then at some level that will force everyone else to break."
  },
  {
   "holder": "Gavin Baker",
   "prediction": "Elon's Terrafab will succeed in becoming the world's largest fab in America.",
   "timeframe": "unspecified",
   "subjects": [
    "Terrafab",
    "Elon Musk"
   ],
   "timestamp": "26:52",
   "quote": "joint venture to build the world's largest fab here in America. I think they're going to be successful."
  },
  {
   "holder": "Gavin Baker",
   "prediction": "OpenAI and Anthropic will exceed well over $200 billion in ARR this year, driven by usage-based pricing plus more compute coming online.",
   "timeframe": "by end of 2026",
   "subjects": [
    "OpenAI",
    "Anthropic",
    "frontier tokens"
   ],
   "timestamp": "36:20",
   "quote": "you will see OpenAI and Anthropic exceed well over $200 billion in ARR this year.",
   "due": "2026-12-31"
  },
  {
   "holder": "Gavin Baker",
   "prediction": "The disaggregation of prefill and decode will extend the useful life of GPUs to 10-15 years, lowering financing costs and helping rescue private credit.",
   "timeframe": "unspecified",
   "subjects": [
    "prefill/decode disaggregation",
    "NVIDIA",
    "private credit"
   ],
   "timestamp": "46:36",
   "quote": "The disaggregation of inference means that I think these GPUs are going to have 10 or 15 year lives."
  },
  {
   "holder": "Gavin Baker",
   "prediction": "If continual learning (real-time weight updating) is solved, AI will experience a really fast takeoff; it appears to be near.",
   "timeframe": "near-term / unspecified",
   "subjects": [
    "continual learning"
   ],
   "timestamp": "38:37",
   "quote": "if we get that, then we have a really fast takeoff. And people seem confident that continual learning is kind of just around the corner."
  },
  {
   "holder": "Gavin Baker",
   "prediction": "Frontier labs will increasingly stop releasing their best models via API (anti-distillation game theory), keeping open source a fixed timeframe behind the frontier.",
   "timeframe": "unspecified",
   "subjects": [
    "NVIDIA",
    "OpenAI",
    "Anthropic"
   ],
   "timestamp": "55:03",
   "quote": "I do think Jensen is going to keep open source a certain timeframe behind the frontier."
  },
  {
   "holder": "Gavin Baker",
   "prediction": "If the returns to frontier tokens relative to other tokens come down, there will be an explosion in value creation at the application layer.",
   "timeframe": "unspecified",
   "subjects": [
    "frontier tokens"
   ],
   "timestamp": "51:29",
   "quote": "if the returns to these Frontier tokens relative to other tokens come down, there's going to be an explosion in value creation at the application layer."
  }
 ],
 "facts": [
  {
   "statement": "Anthropic added $11 billion of ARR in a single month in early 2026 — roughly the combined revenue of Palantir, Snowflake, and Databricks built over ~10 years each.",
   "subjects": [
    "Anthropic"
   ],
   "value": "$11B ARR added in one month",
   "timestamp": "3:45",
   "quote": "Anthropic, they added $11 billion of ARR.",
   "source": "Gavin Baker"
  },
  {
   "statement": "The SaaS and Cloud revolution created roughly $5-10 trillion of value in total.",
   "subjects": [
    "SaaS"
   ],
   "value": "$5-10 trillion",
   "timestamp": "3:45",
   "quote": "the SaaS and Cloud Revolution, it created, we'll call it between $5 and $10 trillion of value.",
   "source": "Gavin Baker"
  },
  {
   "statement": "Anthropic burned roughly 80% less capital than OpenAI to reach a similar revenue scale, reflecting a materially lower cost per token and different structural ROIC.",
   "subjects": [
    "Anthropic",
    "OpenAI"
   ],
   "value": "~80% less capital burned",
   "timestamp": "10:07",
   "quote": "they burned maybe 80% less than OpenAI.",
   "source": "Gavin Baker"
  },
  {
   "statement": "Anthropic's rumored latest mark is about $900 billion valuation against roughly $50 billion in ARR.",
   "subjects": [
    "Anthropic"
   ],
   "value": "$900B valuation / $50B ARR",
   "timestamp": "10:41",
   "quote": "being aggressive really paid Anthropic at $900 billion for $50 billion in ARR, growing at ridiculous rates.",
   "source": "Gavin Baker"
  },
  {
   "statement": "Anthropic's Opus is generating roughly 70% fewer tokens for the same question, an effective deprecation of intelligence density tied to compute constraints and the shift to usage-based pricing.",
   "subjects": [
    "Anthropic"
   ],
   "value": "~70% fewer tokens",
   "timestamp": "10:41",
   "quote": "Claude is even on Opus is generating 70% less tokens for the exact same question.",
   "source": "Gavin Baker"
  },
  {
   "statement": "SpaceX compounded its valuation at roughly low-30% per year for a decade by never pushing valuation.",
   "subjects": [
    "SpaceX",
    "Elon Musk"
   ],
   "value": "~low 30%/yr for a decade",
   "timestamp": "13:27",
   "quote": "My friend Antonio pointed out SpaceX compounded it low 30% per year for a decade.",
   "source": "Gavin Baker"
  },
  {
   "statement": "For data-center infrastructure, zoning and approval have overtaken energy and chips as the biggest gating factor, per the head of data-center infra investing at a major PE firm.",
   "subjects": [
    "data centers"
   ],
   "timestamp": "14:24",
   "quote": "it used to be energy and chips were our biggest gating factors. Now it's zoning and approval, much more important.",
   "source": "Gavin Baker"
  },
  {
   "statement": "A Blackwell rack weighs ~3,000 pounds and measures eight feet high, four feet deep, three feet wide; an orbital-compute satellite is about that size with ~500-foot solar wings and a long radiator.",
   "subjects": [
    "orbital compute"
   ],
   "value": "3,000 lb; 8 ft x 4 ft x 3 ft",
   "timestamp": "14:24",
   "quote": "It's eight feet high. It's four feet deep, three feet wide.",
   "source": "Gavin Baker"
  },
  {
   "statement": "SpaceX operates the world's largest satellite fleet — about 98-99% of all satellites in orbit — and cools every Starlink today.",
   "subjects": [
    "SpaceX"
   ],
   "value": "~98-99% of satellites in orbit",
   "timestamp": "16:39",
   "quote": "SpaceX operates the world's largest satellite fleet, which is 98 or 99 percent of all satellites in orbit.",
   "source": "Gavin Baker"
  },
  {
   "statement": "Starlink V3 is expected to operate at 20 kilowatts, versus a Blackwell rack at 100 kilowatts, informing the orbital-compute power budget.",
   "subjects": [
    "SpaceX",
    "orbital compute"
   ],
   "value": "Starlink V3 ~20 kW; Blackwell rack ~100 kW",
   "timestamp": "16:39",
   "quote": "I think Starlink V3 is going to operate at 20 kilowatts.",
   "source": "Gavin Baker"
  },
  {
   "statement": "The current AI buildout is overwhelmingly funded out of operating cash flow, with every GPU running at ~100% utilization — unlike the year-2000 fiber buildout where ~99% of fiber was unutilized.",
   "subjects": [
    "TSMC",
    "NVIDIA"
   ],
   "value": "100% GPU utilization vs. 99% unutilized fiber",
   "timestamp": "24:00",
   "quote": "every GPU is running at 100% utilization when 99% of fiber was unutilized.",
   "source": "Gavin Baker"
  },
  {
   "statement": "TSMC's leading-node lead over Intel and Samsung is roughly 9-15 months.",
   "subjects": [
    "TSMC"
   ],
   "value": "~9-15 months",
   "timestamp": "26:27",
   "quote": "You got to remember, it's whatever it is, it's 9, 12, 15 months.",
   "source": "Gavin Baker"
  },
  {
   "statement": "Elon built a data center in 122 days, versus the ~3 years everyone else takes.",
   "subjects": [
    "Elon Musk"
   ],
   "value": "122 days vs. ~3 years",
   "timestamp": "30:33",
   "quote": "Everybody else has taken three years to build a data center. He built one in 122 days.",
   "source": "Gavin Baker"
  },
  {
   "statement": "The current Pareto frontier (intelligence vs. cost) is dominated by Anthropic and OpenAI, with Grok 4.3 on it and Gemini 3.1 hanging on — a reversal from nine months earlier when Google dominated it.",
   "subjects": [
    "Anthropic",
    "OpenAI",
    "Grok 4.3",
    "Google"
   ],
   "timestamp": "31:17",
   "quote": "Now, the Pareto frontier is dominated by Anthropic, OpenAI, and then Grok 4.3 is on the Pareto frontier.",
   "source": "Gavin Baker"
  },
  {
   "statement": "Google lost its per-token cost leadership as a result of making conservative design decisions with TPUv8.",
   "subjects": [
    "Google"
   ],
   "timestamp": "31:17",
   "quote": "Google losing their per cost token leadership as a result of making very conservative design decisions with TPUV8",
   "source": "Gavin Baker"
  },
  {
   "statement": "Baker's venture rule of thumb is that 1% market share in AI chips is worth about $100 billion, a good venture outcome.",
   "subjects": [
    "NVIDIA"
   ],
   "value": "1% share ≈ $100B",
   "timestamp": "40:21",
   "quote": "my rule of thumb is 1% market share is going to be worth 100 billion. 100 billion is a pretty good venture outcome.",
   "source": "Gavin Baker"
  },
  {
   "statement": "It took Cerebras three generations of chips to get wafer-scale computing right; Atreides was a venture investor in the company.",
   "subjects": [
    "Cerebras"
   ],
   "value": "3 chip generations",
   "timestamp": "42:51",
   "quote": "it took Cerebris three generations of chips to get it right.",
   "source": "Gavin Baker"
  },
  {
   "statement": "GPU financing is being underwritten on roughly 3-4 year lives, and CoreWeave's lowest financing was around the low 7% range — versus a potential 5-6% if useful lives extend.",
   "subjects": [
    "private credit",
    "NVIDIA"
   ],
   "value": "3-4 yr underwriting; ~low-7% financing",
   "timestamp": "46:36",
   "quote": "they were underwriting that to, I think, three or four years.",
   "source": "Gavin Baker"
  },
  {
   "statement": "Cross-sectionally, SemiCap equipment companies trade at ~40x next-quarter annualized earnings while DRAM companies trade at mid-single-digit multiples — a gap far wider than the 5-vs-12 (and once 3-vs-45) seen at prior cycle peaks.",
   "subjects": [
    "NVIDIA"
   ],
   "value": "~40x vs. mid-single-digit P/E",
   "timestamp": "1:00:06",
   "quote": "you have SemiCAP equipment companies trading at 40 times next quarter's annualized earnings and DRAM companies trading at mid-single digit.",
   "source": "Gavin Baker"
  }
 ],
 "notable_quotes": [
  {
   "speaker": "Gavin Baker",
   "timestamp": "3:45",
   "quote": "what was happening in AI was, I think, the most extraordinary moment in the history of capitalism, the history of American business.",
   "why": "The thesis-setting frame for the entire conversation — that early 2026 AI is unprecedented in capitalist history."
  },
  {
   "speaker": "Gavin Baker",
   "timestamp": "24:00",
   "quote": "So Taiwan Semi, if we don't get a bubble, we need to throw a party for them because they will have single-handedly prevented a bubble.",
   "why": "Crystallizes Baker's core Wafers thesis: TSMC's capacity discipline is the swing factor between healthy buildout and a crash."
  },
  {
   "speaker": "Gavin Baker",
   "timestamp": "14:39",
   "quote": "I would like to redefine orbital compute has racks in space, not giant, floating pentagon-sized data centers in space. That's silly. What makes a data center is you're connecting these racks with lasers.",
   "why": "His memorable reframing of the orbital-compute idea from sci-fi caricature to concrete engineering."
  },
  {
   "speaker": "Gavin Baker",
   "timestamp": "31:17",
   "quote": "a violation of Richard Sutton's bitter lesson is for sure the biggest risk to this trade, to all of AI.",
   "why": "Names the single biggest tail risk to the AI investment thesis in one line."
  },
  {
   "speaker": "Gavin Baker",
   "timestamp": "57:05",
   "quote": "If we do not all become masters of the machine gun, we're getting mastered.",
   "why": "The Last Samurai metaphor that captures his stance on adopting AI as an investor — a line a listener remembers."
  },
  {
   "speaker": "Gavin Baker",
   "timestamp": "1:12:02",
   "quote": "trillions of dollars of value has been destroyed by AI at the application layer.",
   "why": "A striking, contrarian framing of where AI value is and is not accruing — the application-layer destruction thesis."
  }
 ],
 "extracted_by": "agent",
 "prompt_version": "episode-extract@3",
 "_file": "invest-like-the-best-2026-05-20-gavin-baker-watts-and-wafers.json"
}