Watts, Wafers, and the Future of AI Infra | Gavin Baker

Key points

• **'Anthropic added $11B of ARR in ONE month — the combined size of Palantir + Snowflake + Databricks. Nothing like this has ever happened in the history of capitalism.'** Gavin's headline framing this week sets the new ceiling. Cross-references [Brad's $14B-in-April figure from Issue 05 Altimeter](/issues/2026-05-17) and [Krishna Rao's $30B end-Q1 / 500% NDR disclosure on ILTB last week](/issues/2026-05-17). **The triangulation across 3 vantage points (CFO, Brad on CNBC, Gavin on ILTB) has now hardened to: Anthropic exited April at $44-50B run-rate, growing at a 10x/year exponential, in single-quarter ARR adds bigger than the lifetime of any prior SaaS company.**
• **Anthropic Opus is producing 70% FEWER tokens than before for the same question — 'they have clearly deprecated the intelligence of Claude.'** Gavin's most underdiscussed disclosure. **Implication: unconstrained, Anthropic would be at $100-150-200B ARR today** because the compute floor is forcing token rationing per request. *'You might be buying it at more like 5x unconstrained run-rate revenue.'* **Direct continuation of [Krishna Rao's 'pricing-stability + Jevons-paradox Opus 4.5 cut' from Issue 05 ILTB](/issues/2026-05-17)** — the price cut was paired with token rationing to ration compute. The clean inference: Anthropic's actual revenue ceiling is *compute*, not demand. Cross-reference to [Issue 05 Brad Gerstner's 30-40%-compute-delays observation](/issues/2026-05-17) for why this matters now.
• **March 2026 sell-off framing: 'pent-up alpha, not wrong-thesis pain.'** *'Two kinds of drawdowns — wrong-thesis where you have to take your medicine and crystallise the loss; pent-up-alpha where you can lean in.'* Gavin's takedown of the March bears: 'tech got as cheap as it's been versus the rest of the market in 10 years, at the most extraordinary AI inflection in capitalism's history.' **Concrete capital-positioning lesson**: when AI fundamentals decouple from price action, the strait-of-Hormuz scare was *bullish* for US (natural gas $2.50 vs Europe/Asia 2-3x higher = relative manufacturing advantage), exactly what Trump cares about. **Anti-doom data point: the most-extraordinary-moment-in-capitalism happened in the same window as the bears were loudest.**
• **Orbital compute as the actual Watts solution — 'racks in space, not pentagon-sized buildings.'** Gavin's most concrete reframe: **'A Blackwell rack weighs 3,000 lbs, 8ft × 4ft × 3ft. SpaceX has shown an illustration — it's a rack with 500ft solar wings on each side, kept in a sun-synchronous orbit so the panels are always in the sun and the radiator (hundreds of feet) is always in shadow.'** **Why it works**: SpaceX operates 98-99% of all satellites in orbit; Starlink V3 = 20 kW (cooled today); scaling to 60-100 kW = a Blackwell rack equivalent; **racks connected by lasers through vacuum** (already on every Starlink). The cooling/repair scepticism is addressed: 'until you have floating Optimus, you don't repair — but Starship will change the space economy in ways we can't imagine.' **Direct response to [Issue 05 Brad's 'Tom Brady launches a data centre' sell-signal](/issues/2026-05-17)** — orbital compute extends the cycle by years before that signal fires.
• **TSMC is the bubble-prevention mechanism. 'If TSMC did what Jensen wanted, Nvidia could sell $2-3T of GPUs in 2026-27' — overbuild territory.** Gavin's framing: 'If we don't get a bubble, we need to throw a party for TSMC.' **Watch for**: TSMC capacity decisions as the leading indicator. **The break risk**: if Intel or Samsung emerges as a credible 30%+ market share second source, the discipline breaks, capacity floods, and bubble dynamics begin. Currently TSMC leads Intel/Samsung by 9-15 months on the leading node — **the Goldilocks zone is when they stay just-far-enough-ahead to keep them in check but expand enough to keep TSMC at constraint.**
• **TerraFab disclosure: SpaceX + Tesla joint venture for world's largest fab in America, partnership with Intel for 50 years of institutional knowledge.** *'Elon's reputation in hardware engineering means the A-teams at ASML, KLA Tencor, Lam Research, Applied Materials will be here. Next to TerraFab they'll build Taiwan Town — 'I'm going to move your favourite restaurants and their whole staff from Taiwan to Texas.' Then Japan Town, Korea Town. That's not the way Intel or Samsung think.'* **Direct read-across to [Jacob Helberg's Pax Silica forward-deployed industrial base from Issue 05](/issues/2026-05-17)** — TerraFab is the domestic complement to the Philippines forward-deployed base. **'The single largest defence-and-AI integrated infrastructure move in any administration's lifetime.'**
• **Bitter-lesson violation is the biggest risk to the entire AI trade.** Richard Sutton's bitter lesson = compute beats human algorithmic ingenuity. **The March 2026 'Turbo Quant' scare**: Google released a memory-optimization paper during the middle of Micron/Samsung/SK Hynix LTA negotiations. 'Goes viral on X — DRAM is cooked.' Gavin couldn't find a single AI engineer who believed Turbo Quant would impact DRAM demand. **But the bitter-lesson risk is structural**: 'if we get to ASI, the first thing it will want is to be smarter and have more resources — it'll make itself more efficient. The bitter lesson literally includes humans in it. We'll find out if it applies to 300/400/500/600-IQ AGIs.' **The cleanest hedge thesis for the entire AI infra cohort gets articulated here for the first time on the tape.**
• **GPU useful life extending from 3-4 years to 10-15 years via disaggregation of prefill/decode. 'This may single-handedly save private credit.'** Why: disaggregation means Hopper and Ampere can stay useful for prefill (memory-capacity bound) while newer chips do decode (memory-bandwidth bound). 'Until they melt — though they do melt.' **Financing implication**: dropping the depreciation curve from 3-4 yrs to 10-15 yrs changes financing cost from low-7% (CoreWeave low) to 5-6%. **The most important under-discussed mechanism that supports the [Issue 04 'GPU debt isn't the next great financial crisis' thesis](/issues/2026-05-10).** Also flips the 'SPV off-book hyperscaler debt' tail risk Brad acknowledged in Issue 05 — longer useful life means more cash flow to service the debt.
• **Pareto frontier disclosure**: 'Google dominated the Pareto frontier (intelligence vs cost) 9 months ago at every point. Now dominated by Anthropic + OpenAI. Grok 4.3 holds the ~500B-parameter spot. Gemini 3.1 is hanging on — and if I were to bet, they're subsidising it out of pride.' **Why Google lost the cost edge**: 'TPU V8 made very conservative design decisions to try to take it away partially from Broadcom and Nvidia. Continuing to make aggressive choices, they lost.' **Direct cross-reference to [Reiner Pope on TPU vs GPU architectural tradeoffs in this week's Dwarkesh](/issues/2026-05-24)** — Google's TPU coarse-grained design lost flexibility, and the flexibility now matters more than pure cost.
• **Application-layer value: 'net DESTROYED at trillions of dollars' even counting Cursor + Cognition.** Gavin's blunt take: 'in Jensen's five-layer cake (energy → data centres → chips → models → applications), profits aren't really accruing to applications.' **The 'token path' framing (Jamon Ball, Altimeter)**: 'If you're a software/AI company of any kind, you have to be in the token path. Databricks is in it. Comparable companies are in it. If you're not in the token path and not in some really niche thing, life may be hard.' **Direct counter-data to [Issue 05 Brad's 'too hard basket' framing on software](/issues/2026-05-17)** — same conclusion, different vantage (Brad: market multiple reversion; Gavin: net value destruction at the layer). **The exception window: 'if frontier-token returns come down relative to other tokens, there's going to be an explosion in value creation at the application layer.'**
• **The 'all you can eat → pay by the drink' pricing shift is structurally bullish for the labs.** Gavin's telecom analogy: 'Cellular was a great growth industry until everyone went to all-you-can-eat. Long distance the same. AI is going the OPPOSITE direction — flat-rate $250/mo subscription getting rate-limited; usage-based enterprise plans getting the un-deprecated tokens.' **The labbotomy disclosure: you cannot use $200-300 frontier model subscription to understand what frontier AI is capable of today — you need Cloud Code or Codex on an enterprise plan.** **'OpenAI and Anthropic will exceed well over $200B in ARR this year — not just from more compute, but from being able to push frontier-token pricing via usage-based enterprise.'**

Notable quotes

Themes

• Anthropic +$11B ARR in one month = combined SaaS lifetimes of Palantir+Snowflake+Databricks
• Anthropic Opus deprecating intelligence (70% fewer tokens) — unconstrained = $100-200B ARR
• Orbital compute as the Watts solution — Starlink V3 racks-in-space + laser interconnects
• TSMC capacity decisions as the singular bubble-prevention mechanism
• GPU useful life 3-4yrs → 10-15yrs (saves private credit, lowers financing 7% → 5-6%)