Gavin Baker on AI Infrastructure, Power Constraints, and Semiconductor Investing

Investor Gavin Baker reads the AI build-out through its two hard inputs — watts and wafers — and argues that a single supply-constrained foundry in Taiwan may be all that stands between the current boom and a classic technology bubble.

Key ideas

1. The spring 2026 drawdown was the buying opportunity, not the warning. Anthropic added roughly $11bn of annualised revenue in a single month — more than Palantir, Snowflake, and Databricks built in a decade combined — while tech traded at its cheapest relative multiple in ten years. Baker calls it ‘the most extraordinary moment in the history of capitalism.’
2. Watts get solved by capitalism; wafers are the true bottleneck. The power shortage eases through 2027–28 as turbine capacity ramps and, later, as orbital compute (‘racks in space’) comes online. But leading-edge wafer supply sits with TSMC, and its capacity decisions are the single variable to watch for whether the cycle overbuilds.
3. TSMC’s discipline is, paradoxically, what prevents a bubble. If TSMC built everything Jensen wanted, Nvidia could sell $2–3tn of GPUs — enough to overshoot demand and trigger a crash. The wafer constraint is a governor on the whole system; the risk is that Intel or Samsung ‘breaks’ and floods capacity.
4. All the economic value at the model layer accrues to the frontier. Frontier tokens capture the overwhelming majority of returns, and access is being rationed through usage-based enterprise pricing — ‘AI is just shifting from all you can eat to pay by the drink.’ A $250/month plan now buys a lobotomised model.
5. The biggest tail risk to the entire AI trade is a violation of Richard Sutton’s Bitter Lesson — human algorithmic ingenuity (a memory trick, a distillation) displacing raw compute. Baker is less worried than most because he thinks proximity to ASI may temporarily bend that rule.

Summary

The most extraordinary moment in capitalism

Baker frames March–April 2026 as a drawdown of the second kind — not a repudiation of his thesis but a mispricing he could lean into. The NASDAQ sold off while AI fundamentals inflected: Anthropic added ~$11bn of ARR in a month, a pace with no precedent in business history. He contrasts Anthropic’s capital efficiency with OpenAI’s, estimating Anthropic burned perhaps 80% less to reach comparable scale, and argues that a compute-unconstrained Anthropic would already be at $100–200bn of run-rate revenue. The macro backdrop helped: with the US now the world’s largest oil and gas exporter, an energy shock that once meant 1970s-style shortages instead improved American relative manufacturing competitiveness, keeping him focused on AI at historically attractive valuations.

Watts: solved by capitalism, then by space

Baker separates the two inputs. Power is a solvable problem — the binding constraint has shifted from energy and chips to ‘zoning and approval’, and capitalism reliably grinds down supply shortages of turbines and generation over a few years. He expects the watts shortage to begin easing in 2027–28. The longer-run answer is orbital compute, which he insists on reframing: not Pentagon-sized data centres in space but individual Blackwell-class racks, cooled by radiators in sun-synchronous orbit and linked by lasers through vacuum — the same laser-through-vacuum links SpaceX already runs across the Starlink fleet. Training stays on Earth; inference is the natural fit for orbit. He is bullish that Starship and SpaceX’s engineering depth make this real sooner than sceptics think.

Wafers: the constraint that governs the cycle

Wafers are different — leading-edge supply is controlled by ‘plenty older humans in Taiwan’ at TSMC, guardians of Morris Chang’s legacy. Drawing on Carlota Perez’s history of technology bubbles, Baker notes that every foundational technology has produced one; a bubble funds the build-out, then supply overshoots demand and crashes (worst when debt-fuelled, as in 2000). The current build-out is healthier — funded from operating cash flow, every GPU at 100% utilisation, unlike the dark fibre of 2000. His hope is that TSMC’s supply discipline prevents an overbuild outright: a ‘Goldilocks zone’ where it expands enough to keep Intel and Samsung from emerging as a scaled second source, yet not so much that wafers stop being scarce. The new SpaceX/Tesla ‘TerraFab’ joint venture — with Intel’s institutional knowledge and the A-teams from ASML, KLA, Lam, and Applied Materials — is his candidate for the largest US fab.

The frontier captures the value

The economic surprise is that returns at the model layer concentrate almost entirely at the frontier: last year’s mind-blowing model is this year’s ‘intolerable’ one. The Pareto frontier of intelligence-per-cost, dominated by Google nine months ago, is now held by Anthropic, OpenAI, and Grok, with Gemini clinging on. Access is the new moat — frontier labs have shifted to usage-based pricing, so flat monthly plans deliver a rate-limited, token-starved model. Baker draws the telecom analogy: growth persisted while cellular and long-distance charged by usage and collapsed once they went all-you-can-eat; AI is moving the other way, from all-you-can-eat to ‘pay by the drink’, which is why he expects OpenAI and Anthropic to clear $200bn+ ARR.

Different, hard, and in the token path

On chip start-ups, Baker invokes the tank designer’s ‘iron triangle’ — every design trades attack, defence, and mobility. Trying to build a better GPU is futile against Nvidia’s TSMC pricing and model-co-design advantages; the winning move is something ‘different and hard to do’, like Cerebras’s wafer-scale computing or aggressive prefill/decode disaggregation. His rule of thumb: 1% share is worth ~$100bn, a fine venture outcome, but anything merely obvious will be crushed before it reaches scale. The same test applies to application founders — you must be in Jamin Ball’s ‘token path’ or in a niche too small for the model labs to bother with. Disaggregating prefill from decode also extends GPU useful lives to 10–15 years, which Baker thinks could lower financing costs and ‘save private credit’ (see Private Credit).

Risks: the bitter lesson, and personal safety

The biggest risk to the whole trade is a violation of the Bitter Lesson — human ingenuity beating brute compute, as a viral ‘turbo quad’ memory optimisation briefly threatened DRAM demand. Baker is less worried because he thinks the field is close to ASI, and a super-intelligent system’s first act — making itself more efficient — might itself be a temporary bitter-lesson violation. He closes on darker knock-ons: AI has net destroyed trillions in application-layer value even counting Cursor and Cognition; deepfake-enabled cybercrime makes a family ‘safe word’ prudent; and he worries genuinely about rising political violence directed at AI figures. Yet he remains an optimist, citing AI-assisted drug discovery for a child’s rare disease, and hopes AI dominance yields a new Pax Americana rather than instability.

Speakers

• Gavin Baker — technology and AI investor; managing partner and CIO of Atreides Management; a recurring Invest Like the Best guest known for reading the AI cycle through its hardware and energy constraints.

See also

• Dylan Patel on the Token Economy, AI Supply-Demand, and the Permanent Underclass — the prior Invest Like the Best episode on the same supply-demand stack and the token economy.
• Bitter Lesson — the compute-beats-ingenuity principle Baker names as the single biggest risk to the AI trade.
• Token Economics — frontier-token value capture, usage-based pricing, and the token path.
• Five-Layer AI Cake — Jensen’s energy → data centres → chips → models → applications stack, the frame Baker uses to explain where profits accrue.
• Private Credit — GPU financing and why longer useful lives may reprice the build-out’s debt.
• Scaling Laws — why more compute continuing to win underpins Baker’s frontier thesis.