Gavin Baker Watts, Wafers, and AI Infra Investment Thesis

Source: Watts, Wafers, and the Future of AI Infra, Invest Like The Best (Patrick O'Shaughnessy), May 20, 2026.

The Framework: Watts and Wafers

Gavin Baker's organizing lens for the AI buildout is two physical constraints — power and silicon — with capitalism racing to solve both while TSMC's discipline may prevent a debt-fueled bubble.

Constraint	Near-Term Dynamic	Who Wins	What to Watch
Watts	Zoning and approvals now gate faster than turbines or chips; capitalism solves by 2027–28	GE Vernova, turbine castings, jet-engine repurposing (Boom Aerospace)	Midterm politics, orbital compute proof points
Wafers	Taiwan Semi supply discipline caps GPU output and may prevent overbuild	TSMC, Nvidia, semi equipment (ASML, KLA, Lam, Applied Materials)	TSMC capacity expansion pace
Orbital compute	Racks in space — not buildings — reframes the watts ceiling long-term	SpaceX, laser-linked satellite fleets	Starlink V3 power scaling, Starship economics
Frontier tokens	Overwhelming model-layer economics still accrue to frontier labs, not apps	Anthropic, OpenAI, Google, xAI	Pareto frontier shifts, usage-based pricing

Investment Thesis #1: TSMC Is the Bubble Brake — and the Most Important Variable in AI

Baker frames Taiwan Semi as the single company that may prevent a Carletta Perez-style infrastructure bubble by withholding wafer capacity even as latent GPU demand explodes.

"If Taiwan Semi did what Jensen wanted, I think Nvidia could sell two trillion dollars of GPUs in 26 or 27, maybe two and a half trillion, maybe three trillion, but there is a limit where consumers would consume so much that you probably would be in an overbuild."

He says Jensen has never had a contract with TSMC — just handshake fairness — and that if TSMC expands in a Goldilocks zone, it keeps Intel and Samsung from becoming a true second source while still constraining wafers enough to avoid a crash. The market signal to watch is TSMC's capacity decisions, not hyperscaler CapEx headlines alone.

Trigger: TSMC announces measured N2/CoWoS expansion that keeps leading-edge share above ~70% without flooding the market; Intel or Samsung breaking price discipline would be the opposite signal.

Names: TSMC (TSM), NVIDIA (NVDA), Intel (INTC), Samsung (foundry alternative if discipline breaks).

Investment Thesis #2: Orbital Compute Reframes the Watts Shortage — SpaceX Owns the Asymmetric Bet

Baker is emphatic that "data centers in space" means racks in space, not Pentagon-sized buildings. A Blackwell rack weighs ~3,000 lbs; SpaceX's illustration is essentially a rack with 500-ft solar wings in sun-synchronous orbit, radiators extending hundreds of feet behind, racks linked by lasers through vacuum — the same laser tech already on every Starlink.

"SpaceX operates the world's largest satellite fleet which is like 98 or 99% of all satellites in orbit... Starlink V3 is going to operate at 20 kW. A Blackwell rack is only 100 kW. And people talk a lot about density... they seem very confident they're going to go right to 100 to 120."

He is not bearish on terrestrial data centers for his lifetime — training stays on Earth, and total compute consumption will rise — but warns that power and cooling companies ramping capacity just as orbital compute becomes credible face a timing risk. Watts shortages probably ease 2027–28 on Earth; orbital compute solves the structural ceiling after that.

Trigger: SpaceX demonstrates economically viable 100+ kW orbital racks; terrestrial power/cooling names re-rate on orbital-credible timelines.

Names: SpaceX (private; IPO watch), GE Vernova (GEV), Howmet Aerospace (HWM) (turbine castings for near-term watts bridge).

Investment Thesis #3: Terafab — Domestic Wafers via Elon's Hardware Playbook

Baker is bullish on Terafab, the SpaceX/Tesla joint venture building America's largest fab, with an Intel partnership supplying ~50 years of institutional knowledge only a few quarters behind the frontier.

"The A teams will be here because of Elon's reputation in hardware engineering... they're going to recruit the best people because the best engineers want to work for Elon... next to Terafab, they'll be a Taiwan town."

Semi equipment leaders — ASML, KLA, Lam Research, Applied Materials — wanted TSMC to catch up in the 1980s–90s and sent their best teams to Taiwan; Baker expects the same dynamic in Texas. Terafab is politically aligned with reshoring goals and, critically, different enough from TSMC's model that it may not alienate the Silicon Shield.

Trigger: Terafab breaks ground with named ASML/KLA/Lam/Applied Materials support; Intel Texas fab pace accelerates under SpaceX pressure (Samsung already gave Elon an office for complaining about build speed).

Names: Intel (INTC), ASML, KLA Corporation (KLAC), Lam Research (LRCX), Applied Materials (AMAT), Tesla (TSLA) (joint venture participant).

Investment Thesis #4: Frontier Tokens and Usage Pricing — The Model Layer Keeps Compounding

Baker says the overwhelming majority of economic returns at the model layer have accrued to frontier tokens, which surprised him — Gemini 3.1 Pro was "mind-blowing" at launch and is now "intolerable." The Pareto frontier (intelligence vs. cost) shifted from Google dominance nine months ago to Anthropic, OpenAI, and xAI Grok 4.3, with Gemini "hanging on."

Anthropic added $11 billion of ARR in one month — more than Palantir, Snowflake, and Databricks combined built over a decade. At a rumored $900 billion valuation on ~$50 billion ARR, Baker argues unconstrained compute could put them at $100–200 billion run rate; they have deprecated Claude intelligence (70% fewer tokens per answer on Opus). AI labs are shifting from all-you-can-eat subscriptions to usage-based pricing, which Baker compares to cellular's fixed-plus-overage era — the growth phase before "all you can eat" killed telecom margins.

"AI is just shifting from all you can eat to pay by the drink... this is probably why you will see OpenAI and Anthropic exceed well over $200 billion in ARR this year."

Trigger: Enterprise usage-based plans from Anthropic/OpenAI show accelerating ARR per seat; Google IO releases something that reclaims Pareto frontier leadership.

Names: Anthropic (private), OpenAI (private), Google (GOOG), xAI (private).

Investment Thesis #5: Trainium, GPU Lifespan Extension, and the Hyperscaler Installed Base

Among GPU alternatives, Baker says Amazon Trainium is "tugging on Superman's cape" — Trainium 3 must ramp with a switch scale-up network for economical inference. Prefill/decode disaggregation means GPUs can live 10–15 years instead of 3–4: Cerebras systems or Groq LPUs (Nvidia-acquired) sit in front of Hopper or even Ampere for prefill while extending legacy fleet life.

"This is going to be really good for the whole private credit industry... if you can start to finance GPUs at more like 5% or 6% instead of... low sevens, that actually mathematically changes the cost to finance this buildout."

Hyperscalers with giant installed CPU fleets — orchestration, tool calls, agents — may "catch up a little bit to the sellers of shortage." Baker rates Amazon strongest on Trainium plus robotics P&L efficiency over the next 18 months; Google on compute scale and data (YouTube valuable for robotics); Meta on AI-first culture and Muse surprise; Microsoft on Satya's courageous choice to use GPUs internally rather than rent to OpenAI (Microsoft would "probably be an $800 stock" if it sold capacity instead).

Trigger: Trainium 3 production ramp; private credit GPU financing terms extend toward 10–15-year useful lives; hyperscaler CPU utilization rises in agentic workloads.

Names: Amazon (AMZN), Google (GOOG), Meta (META), Microsoft (MSFT), Cerebras (private; Atreides venture holding), NVIDIA (NVDA) (Groq LPU integration).

The Ecosystem Map: Where Baker Is Placing Bets

Explicit Positions and Conviction Signals

Anthropic growth inflection — $11B ARR in one month; capital-efficient vs. OpenAI (~80% less burn to similar revenue scale)
TSMC supply discipline — handshake partnership with Nvidia; potential bubble prevention
SpaceX orbital compute — racks-in-space thesis; Starlink fleet as proof of operational scale
Terafab — domestic fab with Intel knowledge + semi equipment A-teams
Cerebras — Atreides venture investment; wafer-scale computing as "different and hard"
Cybersecurity overinvestment — preparing for Mythos 3/4 world; firm-wide "safe word" protocols

Cross-Sectional Valuation Dislocations Baker Flags

Semicap equipment at ~40x next-quarter annualized earnings vs. DRAM at mid-single digits — "those can't both be true"
Nvidia cheap vs. market in early April while GE Vernova embeds "unfathomable" Nvidia share loss
Astera Labs (ALAB) miscategorized in "copper loser" baskets — biggest product is a switch connecting accelerators via copper and optics

Who Captures Value (Baker's Stack View)

Value accrues to companies with the highest ratio of utilized GPUs per human — energy, data centers, chips, models — not the application layer, where AI has "net destroyed" trillions even counting Cursor and Cognition

Key Risks

Richard Sutton's bitter lesson violation — TurboQuant-style algorithmic efficiency scares (Google publicizing DRAM optimizations during Micron/SK Hynix/Samsung LTA negotiations) could temporarily break the "more compute always wins" trade; ASI may eventually self-optimize in ways that violate the lesson
Continual learning breakthrough — sample-efficient real-time weight updates would accelerate takeoff; timing uncertain but "around the corner" per model builders
Diversity breakdown / bubble — universal AI bullishness; low-quality commodity suppliers mooning on shortages then crashing; George Vanderhiden-style underperformance if you fight the bubble too early
TSMC over-expansion or Intel/Samsung discipline break — wafer flood triggers Perez-style crash; "one of Intel and Samsung they're not going to stay disciplined"
Orbital compute timing — terrestrial power/cooling CapEx ramps into obsolescence if SpaceX executes faster than skeptics expect
Frontier token premium collapse — if non-frontier tokens close the gap, application-layer value creation explodes but model-layer economics compress
Application layer destruction — founders outside the "token path" struggle; legacy SaaS and private credit SaaS loans face margin structure mismatch
Geopolitical amplification — Taiwan risk rises with global uncertainty; AI as political target increases personal safety risk for industry leaders
Access inequality — best AI now requires expensive enterprise usage plans; frontier intelligence gated by wallet size

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	NVIDIA (NVDA)	Central to buildout; cheap vs. market in April; Groq LPU extends GPU fleet life	"Nvidia effect" stronger than imagined if Google IO disappoints
1	TSMC (TSM)	Wafer constraint prevents bubble; Jensen's uncodified partner	Watch capacity decisions as bubble indicator
1	Anthropic (private)	$11B ARR/month; capital-efficient frontier lab	Usage pricing + compute constraints = pent-up demand
2	Amazon (AMZN)	Trainium 3 ramp; robotics P&L efficiency next 18 months; Nova underrated	"Trainium is tugging on Superman's cape"
2	Google (GOOG)	Most compute + YouTube data for robotics; GCP accelerating	"Google's never not going to be in a good position"
2	GE Vernova (GEV)	Near-term watts bridge; valuation embeds impossible Nvidia share loss	Cross-sectional disconnect vs. NVDA
2	Micron (MU) / DRAM	Mid-single-digit multiples vs. 40x semicap — mispriced shortage	"I don't know anyone like me who's not really bullish on DRAM"
2	ASML / KLA / Lam / Applied Materials	Terafab A-team deployment + TSMC expansion	Named as Taiwan catch-up enablers
3	SpaceX (private)	Orbital racks-in-space; 98% of satellites already Starlink	Starlink V3 20 kW → 100–120 kW path
3	Intel (INTC)	Terafab partner; second-source if discipline breaks	Samsung gave Elon an office over fab pace
3	Meta (META)	Only internet giant truly AI-first internally; Muse near frontier	Zuckerberg paid up for talent; Muse upside surprise
3	Howmet Aerospace (HWM)	Two Western machines cast turbine blades; watts bottleneck	Turbine capacity announcements accelerating
3	Cerebras (private)	Wafer-scale compute — different, hard, Atreides-backed	Prefill/decode disaggregation extends legacy GPU life
3	Astera Labs (ALAB)	Miscategorized copper loser; switch product connects accelerators	Mis-basketed in factor trades
4	Microsoft (MSFT)	Courageous internal compute bet vs. renting to OpenAI	Would be ~$800 stock on rental model
4	OpenAI (private)	Usage pricing shift; secured more compute than Anthropic	Aggressive compute strategy "paid"

Monitoring Checklist

Bottom Line

TSMC capacity decisions are the bubble indicator. If Taiwan Semi expands enough to let Jensen sell $2T+ of GPUs without demand catching up, the Perez crash playbook activates. Until then, wafer scarcity is a feature — own TSM and watch the one variable Baker says matters most.
Orbital compute is racks in space, and SpaceX is already operating the proof fleet. Starlink's 20 kW satellites scaling toward Blackwell-rack power levels reframes the watts ceiling. This is not bearish for all terrestrial data centers, but it is a timing risk for power and cooling CapEx ramping into 2027–28.
DRAM is the cross-sectional mispricing. Semicap at 40x and memory at mid-single digits cannot both be right. Baker knows no one bearish on DRAM — in a shortage cycle, the highest-cost suppliers moon first, but quality memory names look cheap vs. equipment peers.
Trainium and GPU lifespan extension change the financing math. Prefill/decode disaggregation plus Cerebras/Groq front-ends can extend Hopper/Ampere life to 10–15 years, lowering private credit GPU financing costs and favoring hyperscalers sitting on giant installed CPU/GPU fleets — especially Amazon.
Frontier tokens still eat the model layer. Anthropic's $11B ARR month and the shift to usage-based pricing are the clearest signals that the application layer remains starved while labs compound — until Google IO or a bitter-lesson shock changes the Pareto frontier.