Summary

Baker treats March’s AI selloff as “pent-up alpha,” because prices fell while Anthropic added $11 billion of ARR in one month and a 500% NDR statistic was shared on the show. He argues that this one-month addition rivaled the combined businesses of Palantir, Snowflake, and Databricks, while DeepSeek had already demonstrated that reasoning models increase inference demand: Asian AWS availability-zone prices doubled, GPU availability fell, and DRAM went vertical. “I’ve just never seen an exponential like this.”
Reported revenue may radically understate Anthropic’s compute-constrained economics. Against roughly $50 billion of ARR and a $900 billion valuation, Baker estimates unlimited compute might support $100 billion to $200 billion of revenue, or roughly five times his invented “URR, unconstrained run-rate revenue”; he also believes Anthropic burned perhaps 80% less capital than OpenAI at comparable scale. His expectation that Anthropic could generate cash this year remains a forecast, not a certainty.
The terrestrial watts shortage should begin easing in 2027 or 2028, but zoning and approvals may become more binding than turbines or fuel. Longer term, Baker’s answer is “racks in space”: roughly rack-sized satellites with immense solar wings, radiators, and laser links forming virtual data centers, with inference moving orbital while training stays on Earth. Cooling appears solvable to SpaceX engineers; repair remains the honest weakness “until you have probably floating Optimuses.”
TSMC’s capacity decisions are Baker’s single best indicator of whether AI becomes a classic infrastructure bubble. Unlike 2000, today’s build-out is funded overwhelmingly from operating cash flow and GPUs run at 100% utilization, versus 99% of fiber sitting unused; yet if TSMC supplied everything Jensen Huang wanted, Baker thinks NVIDIA might sell $2 trillion to $3 trillion of GPUs in 2026 or 2027 and eventually overbuild. The “Goldilocks zone” is enough expansion to keep Intel or Samsung below roughly 30% share, but not enough to remove wafer scarcity.
Terafab could challenge normal fab timelines without alienating TSMC. Baker describes a SpaceX joint venture, with Tesla possibly involved, that would use Intel’s institutional knowledge and recruit the equipment companies’ A-teams. He says the relevant process gap may be roughly 9 to 15 months, or three to five quarters, behind the frontier, but presents that timing approximately.
Frontier-token economics remain unusually durable, but three uncertainties could overturn the trade: whether frontier tokens retain their premium, whether ASI violates the “bitter lesson,” and when continual learning arrives. Baker says Gemini 3.1 Pro went from mind-blowing to intolerable as the frontier advanced, while Anthropic, OpenAI, and Grok 4.3 now dominate the intelligence-versus-cost Pareto frontier. A model that updates from one experience rather than “put its hand in the fire a million times” could produce a very fast takeoff—and perhaps optimize away some compute demand.
Usage-based pricing could steepen AI revenue as capped subscriptions obscure what frontier systems can do. Baker says capped $250-to-$300 plans increasingly deliver rate-limited, “lobotomized” models, whereas enterprise and usage plans expose the token budget and agent harness; he compares the shift to telecom moving from all-you-can-eat to “pay by the drink.” His aggressive call is that OpenAI and Anthropic will exceed well over $200 billion in ARR this year as compute availability, frontier pricing, and fleets of agents expand usage.
New chip companies need to be both different and hard, because “a better GPU” invites NVIDIA to copy the trade-off with superior economics and customer knowledge. Disaggregating prefill from decode creates openings for specialized memory-capacity and memory-bandwidth architectures; Baker’s venture rule is that 1% market share could be worth $100 billion, with Cerebras’s wafer-scale design as the exemplar. The same disaggregation could extend Hopper and Ampere lives to 10 or 15 years, potentially lowering GPU financing from the low sevens toward 5% or 6%.
Application investors need exposure to the “token path,” defensible scale, or a niche the model companies will not absorb. Baker credits Cursor and Cognition’s coding focus and cites Replit founder Amjad Masad’s view that coding may be the shortest path to useful AI or ASI. He says AI has still destroyed trillions of dollars of application-layer value and that economic gains currently favor businesses with the highest ratio of utilized GPUs per employee. Meanwhile, public-market selection has become finer-grained: semiconductor-equipment companies at 40 times annualized next-quarter earnings and DRAM companies at mid-single-digit multiples “cannot all be true.”

Deep dive

1. March’s drawdown accumulated alpha rather than invalidating the thesis

Baker separates losses where “your hypothesis was invalidated” from underperformance caused by price action that contradicts well-understood fundamentals. March belonged to the second category: investors could lean in and build “pent-up alpha, pent-up future performance.”
Anthropic’s $11 billion of ARR added in one month was, in Baker’s framing, equivalent to the combined businesses Palantir, Snowflake, and Databricks spent roughly a decade building. He also cites a 500% net dollar retention statistic shared on the show and calls this “the most extraordinary moment in the history of capitalism.”
DeepSeek Monday provided the earlier template. The paper appeared a week before the selloff, but by the panic, Asian AWS availability-zone prices had doubled, GPU availability had fallen, and DRAM was surging—the observable evidence that reasoning inference was dramatically more compute-hungry.
The Strait of Hormuz complicated the setup, yet Baker saw a relative U.S. advantage: the Bloomberg natural-gas quote he called “GY” fell 20% while Asian and European gas doubled or tripled. With tech near its cheapest relative valuation in a decade, AI fundamentals outweighed his acknowledged macro uncertainty.

2. Anthropic’s constraint-adjusted multiple looks cheaper than its headline valuation

Baker distinguishes Anthropic sharply from OpenAI on capital efficiency: Anthropic has a “dramatically lower cost per token” and, by his estimate, burned perhaps 80% less money to reach comparable revenue. He thinks Anthropic may already generate cash or may begin doing so this year.
Compute scarcity is suppressing both intelligence and monetization. Baker cites analysis showing Claude, even on Opus, emitting 70% fewer tokens for the same question; because token quantity partly drives answer quality, he estimates unconstrained revenue could be $100 billion, $150 billion, or even $200 billion instead of roughly $50 billion.
O’Shaughnessy asks why Anthropic does not raise $100 billion at a $3 trillion valuation. Baker’s answer is optionality: capital intensity and geopolitical uncertainty reward leaving investors upside, much as Elon Musk preserved a “sacred covenant” by avoiding greedy SpaceX marks and compounding SpaceX at a low-30% annual rate for a decade.

3. Capitalism can manufacture watts, but permissions may become scarcer

Baker expects capitalism to solve the power shortage absent “big regulatory or political blowback,” which he considers a real possibility. One major infrastructure investor told him energy and chips had ceased being the biggest gates: “Now it’s zoning and approval. Much more important.”
The physical supply chain is genuinely difficult—only two machines can cast certain enormous turbine blades, and the West has not made one of those machines in 80 years—but turbine manufacturers are expanding. Repurposed jet engines offer another bridge, supporting Baker’s forecast that the watts shortage begins alleviating in 2027 or 2028.
Orbital compute should be pictured as “racks in space,” not a floating Pentagon. A Blackwell rack weighs 3,000 pounds and consumes roughly 100 kilowatts; Baker envisions rack-sized satellites with solar wings extending perhaps 500 feet per side, sun-synchronous orbits, long radiators, and Starlink-style laser connections.
The strongest pushback is maintenance, not cooling. SpaceX engineers have convinced Baker they can manage thermal design, but failed hardware may remain inaccessible “until you have probably floating Optimuses.” He expects orbital inference, terrestrial training, and valuable Earth data centers to coexist for his lifetime.

4. TSMC is the de facto governor of the AI capital cycle

Baker caricatures wafer supply as controlled by roughly “twenty older humans in Taiwan,” whom he says view themselves as inheritors of Morris Chang’s legacy. Jensen Huang repeatedly asks them for capacity, despite Baker’s claim that NVIDIA and TSMC do business through fairness, partnership, and handshakes rather than a formal contract.
History argues for a bubble whenever markets recognize a foundational technology and suffer a “breakdown in diversity.” Today differs from 2000 because spending comes mainly from operating cash flow and every GPU is fully utilized, versus 99% of fiber going unused—but Baker refuses to assume the last two or three hundred years of canal, railroad, and technology bubbles no longer rhyme.
His cautionary investor is Fidelity’s George Vander Heiden, who fought the 1999 bubble, endured client skepticism, and retired in early 2000 with roughly 40% of his fund in tobacco and 40% in homebuilders. Those positions probably outperformed the Nasdaq by 20 or 30 times over the next three years. The lesson associated with him was brutal: “Being early is the same thing as being wrong.”
If TSMC fully accommodated NVIDIA, Baker believes $2 trillion to $3 trillion of GPU sales in 2026 or 2027 might create overcapacity. The ideal outcome preserves TSMC’s roughly nine-to-15-month process lead, limits Intel or Samsung from gaining well north of 30% share, and prevents either competitor from breaking industry discipline.

5. Terafab could challenge normal fab timelines without alienating TSMC

Baker describes Terafab as a planned SpaceX joint venture that may also involve Tesla, intended to build America’s largest fab. Its Intel partnership would provide 50 years of institutional knowledge; Baker describes the relevant process gap as roughly nine to 15 months, or three to five quarters, behind the frontier, though he presents that timing approximately.
He expects the “A teams” from ASML, KLA-Tencor, Lam Research, and Applied Materials to participate, just as equipment vendors once helped TSMC catch Intel because they disliked dependence on a single source. Elon Musk’s hardware reputation should also recruit engineers conventional semiconductor managers cannot.
O’Shaughnessy’s pushback is the industry’s unavoidable lead time. Baker’s honest response is “We’ll see”: Musk built a data center in 122 days when others took three years, and reportedly secured an office inside Samsung’s Texas fab because its expansion pace frustrated him.

6. Frontier value persists until the bitter lesson or open models break it

Baker is surprised that the overwhelming share of model-layer economics still accrues to frontier tokens. Gemini 3.1 Pro felt “mind-blowing” at launch and “intolerable” later, although he concedes companies may prototype at the frontier before deploying cheaper Vertex or open-source models.
Google dominated the intelligence-versus-cost Pareto frontier nine months earlier; after conservative TPU v8 design decisions cost it per-cost-token leadership, Baker says Anthropic and OpenAI now dominate, with Grok 4.3 the best low-cost 500-billion-parameter model and Gemini 3.1 merely hanging on—possibly through subsidy.
TurboQuant illustrated how markets can overreact to algorithmic efficiency: Baker found no AI engineer who expected the memory optimization to reduce DRAM demand. The genuine risk is an ASI-driven exception to Richard Sutton’s “bitter lesson,” because a 300- or 400-IQ model might use its first resources to make itself dramatically more efficient.

7. Pay-by-the-drink AI and continual learning could compound the curve

Harnesses—the runtime providing tools, prompts, context, memory, and state—matter enormously even if the underlying model matters more. Baker now thinks a $250-to-$300 subscription is insufficient for investors seeking intuition; serious evaluation requires Claude Code or Codex, an enterprise account, and usage pricing.
Capped plans deliver a “lobotomized version of the AI,” while usage plans allow the model to spend the tokens its harness judges necessary. Baker compares this to cellular’s growth era of fixed allowances plus overages: AI is leaving all-you-can-eat for “pay by the drink,” just as one person can have 100 agents.
Continual learning would be the sharper discontinuity. Humans learn after touching fire once; current models may require a million examples and another training or reinforcement-learning cycle. If weights can adjust dynamically in real time, Baker expects “a really fast takeoff” and makes its timing one of the three central investor questions.

8. Chip startups must make trade-offs NVIDIA cannot instantly reproduce

Baker borrows tank design’s “iron triangle”: attack, defense, and mobility cannot all be maximized, just as chip designers must trade within physics and TSMC’s rules. TPU, Trainium, and AMD are essentially trying to be better GPUs, but “nobody’s a better GPU”; Trainium 3 is merely “tugging on Superman’s cape,” while MI450 remains unresolved.
Disaggregation supplies a richer design canvas. Andrew Fox’s analogy is “Prefill is loading the cannon, decode is firing”: understanding context is memory-capacity-bound, while generating tokens is memory-bandwidth-bound, enabling specialized chips to optimize aggressively for one phase.
Baker’s rule of thumb is that 1% share may be worth $100 billion, but the architecture must be “different” and hard. Cerebras qualifies through wafer-scale computing, three difficult chip generations, and efforts to place an optical wafer above compute to escape shoreline-I/O constraints.
Disaggregation also changes collateral economics: Cerebras or Groq systems can sit in front of Hopper or Ampere GPUs, allowing those GPUs to perform prefill “until it melts.” Baker thinks they could have 10-to-15-year useful lives, potentially shifting GPU financing from CoreWeave’s low-sevens floor toward 5% or 6% and helping “single-handedly save private credit.” This matters while private credit is already under pressure from SaaS loans.

9. Applications need token-path leverage before the frontier absorbs them

Baker applies “different and hard” to venture broadly. Retail CEOs could destroy obvious e-commerce startups by driving category margins toward “negative ten thousand percent”; Wayfair survived because its founders solved an operationally difficult problem before the idea became obvious and scale erased the opening.
Baker credits Cursor and Cognition’s scale to their early, intense coding focus and notes that Anthropic was also among the companies focused on coding. Replit founder Amjad Masad’s “bitter lesson adjacent” insight is that coding may be the shortest path to useful AI or ASI, because a sufficiently capable coding system can create tools for everything else.
Yet Baker says AI has net destroyed trillions of dollars at the application layer, even counting Cursor and Cognition. Today’s winners have the highest “effective ratio of utilized GPUs per human”; other software companies need to sit in the “token path,” like Databricks, or build a vertical moat before frontier labs reach the niche.
Open source introduces another prisoner’s dilemma. Baker thinks frontier labs may withhold systems such as Mythos to prevent distillation, but one lab releasing its best model via API could gain revenue and intelligence-enhancing resources, pressuring every rival to follow; NVIDIA can meanwhile keep open source a controlled distance behind. “Open source” still consumes energy and GPUs, and open-source model companies commonly receive a revenue share.

10. Public-market leadership is fragmenting beneath the AI headline

Google’s per-cost-token TPU advantage is gone, but its installed compute, search, YouTube data, and accelerating GCP keep it formidable. Baker treats the Google I/O release that week as a test: failure to leapfrog OpenAI or Claude would strengthen the case that NVIDIA’s architectural advantage is larger than assumed.
Meta earns credit for becoming genuinely AI-first; Muse, MSL’s first model, landed surprisingly close to the Pareto frontier. Amazon combines Trainium with prospective retail-robotics efficiencies over 18 months, while its Nova models are “better than they get credit for.”
Microsoft briefly “flinched” on capital spending in early 2025 and lost allocations, but Baker supports Satya Nadella’s risky decision to reserve compute for Copilot and internal models rather than maximize near-term Azure sales. He estimates Microsoft might otherwise be an $800 stock, while questioning whether its current model team can execute.
Cross-sectional prices now “cannot all be true”: semiconductor-equipment companies trade near 40 times annualized next-quarter earnings while DRAM sits at mid-single-digit multiples. Correlations among GPU compute, networking, optical, DRAM, NAND, and HDD broke in January, creating opportunities in miscategorized names such as Astera, whose biggest product is going to be a switch and which Baker says is wrongly placed in copper-loser baskets.

11. Investors must master the machine while society prepares for its blowback

Baker’s Last Samurai analogy is uncompromising: “The machine gun is here. If we do not all become masters of the machine gun, we’re going to get mastered.” His most valuable agent extracts personally relevant needles from six daily hours of podcasts; others inspect proxies, PSUs, incentives, and compensation changes.
Defensive preparation starts with cybersecurity and an offline family or company safe word that cannot be socially engineered. Baker expects realistic video impersonations to know personal context and request transfers; he also fears rising political violence will increasingly target public AI leaders.
Geopolitically, he attributes Ukraine’s improving battlefield position partly to superior battlefield AI and worries U.S. dominance could destabilize adversaries—though it might instead create another Pax Americana. His counterweight is AI’s medical promise: agents helped find a drug already on the market that could impact one child’s rare disease and helped launch a company seeking a cure.
Baker remains an “AI optimist and maximalist,” but calls the technology an “event horizon” demanding humility. Frontier access becoming contingent on wealth feels “a little dystopian”; the Luddites may be wrong, he argues, while their concerns still require thoughtful answers that make the gains broadly shared.

Watts, Wafers, and the Future of AI Infra | Gavin Baker 73. [520 Special] AI + Love, We Won! | An Off-the-Cuff Conversation with the Hackerthon Champion Couple