Gavin Baker - Watts and Wafers - [Invest Like the Best, EP.473]

Episode Summary

Executive Summary: Patrick O’Shaughnessy and Gavin Baker discuss AI as the biggest capital formation event in history, centered on two bottlenecks: watts and wafers. Baker argues power constraints ease by 2027-28 and orbital compute could solve them long term, while TSMC’s wafer capacity decisions may determine whether AI becomes a controlled boom or a full bubble. They also cover frontier-model economics, chip competition, application-layer value, and rising geopolitical and safety risks.

Main Topics: AI as an unprecedented capital and demand shock (Priority: 5/5): Baker argues AI demand is unlike anything in business history, citing Anthropic’s rapid ARR growth as evidence that frontier labs are scaling faster than any prior software category and that March/April drawdowns were buying opportunities rather than thesis breaks. Watts: power constraints and orbital compute (Priority: 5/5): The conversation frames electricity, zoning, and industrial infrastructure as the near-term limits on AI growth. Baker believes capitalism will solve power shortages over time, with easing around 2027-28 and orbital compute as a long-term solution. Wafers: TSMC as the key supply-side variable (Priority: 5/5): Baker says TSMC’s capacity decisions are the single most important variable for whether AI stays constrained or tips into a bubble. He contrasts the current buildout with 2000, emphasizing that supply is still limited by wafer throughput rather than pure capital availability. Frontier model economics and usage-based pricing (Priority: 5/5): Baker argues frontier tokens still capture most economic value, and that usage-based pricing will push OpenAI and Anthropic well above prior ARR expectations. He highlights Claude/OpenAI as distinct businesses with different capital efficiency and notes model quality is tightly linked to token output and harness design. New chip architectures and the disaggregation of inference (Priority: 4/5): He sees room for specialized chips if they are both different and hard to build, especially as pre-fill and decode split apart. He discusses TPU, Trainium, AMD, and Cerebras as examples of architectures making distinct trade-offs rather than trying to be better GPUs. Application-layer uncertainty and the token path (Priority: 4/5): Baker is skeptical that most AI applications can build durable moats unless they are in the token path or truly niche and hard to replicate. He worries a lot of value has already been destroyed at the application layer even while infrastructure winners thrive. Geopolitical, security, and societal consequences (Priority: 4/5): The discussion closes on rising personal-safety and geopolitical risks from AI, especially as the technology becomes more political. Baker is bullish on AI’s benefits but warns of increased volatility, cybercrime, and strategic instability even as AI supports medicine and national power.

Key Arguments: AI demand growth has no historical precedent; Anthropic adding roughly the combined scale of major SaaS leaders in a month signals a new regime. March/April 2025 looked like a mispriced drawdown because underlying AI fundamentals were accelerating even as the NASDAQ sold off. Power shortages are real but manageable; the more binding long-term constraint is wafers, especially TSMC capacity. If TSMC expanded aggressively, AI could overbuild into a bubble; if it stays constrained, it may inadvertently prevent one. Frontier labs still capture most economic value because customers want the best models and usage-based pricing monetizes heavy demand. Anthropic is structurally more capital efficient than OpenAI, implying different ROICs and potentially much higher unconstrained revenue. Orbital compute should be understood as racks in space connected by lasers, not giant floating data centers. The split between pre-fill and decode allows more specialized chips and opens opportunities for non-GPU architectures. Cerebras and similar companies must do something fundamentally different and hard; being a better GPU is unlikely to win. AI is increasing the useful life of GPUs by enabling architectural workarounds and lowering financing costs for installed compute. Application-layer companies need to be in the token path or have a deep niche moat; otherwise model companies can absorb their value. AI improves national competitiveness but may raise political violence, cyber, and geopolitical risks. The biggest strategic advantage for public hyperscalers may be their installed compute and direct engagement with frontier/startup ecosystems.

Data Points: Anthropic ARR growth: $11 billion added in one month - Baker uses this to illustrate unprecedented AI demand and revenue acceleration. Anthropic scale: $50 billion ARR - Current revenue scale discussed for Anthropic. Anthropic valuation: $900 billion - Referenced as the market value implied by recent funding/marks. Anthropic capital efficiency: ~80% less capital burned than OpenAI - Baker says Anthropic reached similar revenue scale with far less burn. Anthropic token output reduction: 70% fewer tokens - He says Claude now produces materially fewer tokens for the same question, reflecting higher token efficiency or “intelligence density.” AI model revenue outlook: >$200 billion ARR this year - Baker’s estimate for OpenAI and Anthropic combined or frontier-model revenue expansion via usage pricing. Power shortage easing: 2027-2028 - He expects the AI electricity constraint to start easing as new sources come online. Blackwell rack weight: 3,000 pounds - Used to explain orbital compute as racks in space rather than large floating buildings. Blackwell rack size: 8 ft x 4 ft x 3 ft - Physical dimensions cited in the orbital compute discussion. Starlink V3 power: 20 kilowatts - Baker compares current satellite power to a Blackwell rack’s ~100 kW demand. Blackwell rack power: 100 kilowatts - Used to explain the scale of compute hardware that might be put in orbit. TSMC market position: Overwhelming fraction of Taiwan’s GDP, water usage, and electricity usage - Describes TSMC’s centrality to Taiwan and AI wafer supply. Cerebras generation count: 3 chip generations - He notes it took three generations to get Cerebras’ architecture right. GPU useful life: 10-15 years - Baker argues disaggregated inference extends GPU lifetimes well beyond skeptics’ assumptions. OpenAI/Anthropic market access: Usage-based plans over $250-$300/month - He says serious frontier usage requires enterprise or usage-based plans, not consumer-tier subscriptions. Private credit financing cost: ~5%-6% - He suggests lower-cost financing for GPUs would materially change AI buildout economics.

Pivotal Quotes: "Nothing like that has ever happened in the history of capitalism, the history of American business." — Gavin Baker: On Anthropic’s rapid ARR addition and the scale of AI demand. "I think capitalism is going to solve the Watts shortage, absent big regulatory or political blocks." — Gavin Baker: On the power bottleneck and the likelihood of infrastructure catching up over time. "If we do not all become masters of the machine gun, we’re going to get mastered." — Gavin Baker: His metaphor for AI adoption: humans and investors must learn to use the new tool or be displaced by it.

Implications: AI is entering a phase where infrastructure, chip supply, and model monetization matter more than hype. Winners will likely be the firms controlling power, wafers, frontier tokens, and hard-to-replicate architectures, while applications face intense commoditization and increased geopolitical and security risk.