Watts and Wafers - Gavin Baker on AI's Power and Silicon Bottlenecks
Gavin Baker, founding partner of Atreides Management, joins Patrick O'Shaughnessy for a sweeping discussion of the physical constraints shaping AI's next phase. He argues that the near-term power shortage will begin easing in 2027-2028 as new energy sources come online, with orbital compute — racks of accelerators linked by lasers and cooled by sun-synchronous radiators — eventually solving the long-term problem via SpaceX's Starship. On wafers, he frames TSMC's capacity decisions as the single most important variable for whether the AI buildout produces a bubble; disciplined TSMC expansion, in his view, is what prevents a 2000-style overbuild. Baker offers pointed views on the frontier labs. He values Anthropic at roughly $900B on $50B ARR and argues it could be doing $100-200B if unconstrained, crediting its dramatically lower cost-per-token versus OpenAI. He maps the current Pareto frontier as dominated by Anthropic and OpenAI, with xAI's Grok 4.3 as the best low-cost 500B-parameter model and Google slipping after losing per-token cost leadership with conservative TPUv8 design choices. He praises Meta's Muse as a surprise, credits Zuckerberg for making Meta AI-first, and says Satya Nadella is making the right but costly call to repurpose Microsoft compute toward internal models rather than just reselling it to OpenAI. He also sizes the new economics: disaggregation of pre-fill and decode extends GPU useful life to 10-15 years, easing private credit stress, while the shift to usage-based pricing could push OpenAI and Anthropic past $200B in ARR this year. On chips, he argues startups must be 'different and hard' — citing Atreides-backed Cerebras and its wafer-scale approach — and warns that a temporary violation of the bitter lesson by ASI, plus questions about continual learning, are the biggest open risks to the trade.
Original sources
Podcast Invest Like the Best
Episode:Gavin Baker - Watts and Wafers - [Invest Like the Best, EP.473] ↗
YouTube Invest Like the Best
Video:Watts, Wafers, and the Future of AI Infra | Gavin Baker ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.