Gavin Baker - Watts and Wafers - [Invest Like the Best, EP.473]
Summary
Gavin Baker, CIO of Atreides Management, analyzes the future of AI through the lens of "watts and wafers," identifying these as the primary physical constraints. He projects the near-term power shortage will ease by 2027-2028 with new energy sources, ultimately solved by orbital compute, envisioning racks in space connected by lasers, with SpaceX's Starship and Starlink V3 playing a crucial role. Regarding wafers, Baker emphasizes TSMC's capacity decisions as the most critical variable to prevent an AI bubble, noting their historical supply discipline. He also highlights Elon Musk's TerraFab, a US-based joint venture with Intel, as a significant initiative to increase domestic chip production. Baker observes that frontier models like Anthropic and OpenAI are currently capturing the overwhelming majority of AI's economic value, with Anthropic achieving \$11 billion ARR in a single month. He notes the shift to usage-based pricing for AI models is a strong revenue driver and discusses how GPU disaggregation extends hardware lifespan.
Key takeaway
For investors evaluating AI infrastructure, recognize that physical constraints in power and chip manufacturing will shape market winners. Your focus should be on TSMC's capacity decisions as a primary indicator of potential market bubbles and the long-term viability of orbital compute solutions like SpaceX's "racks in space." Be aware that frontier AI models currently dominate value capture, and the shift to usage-based pricing signals continued revenue growth, but cross-sectional valuations show potential market inefficiencies.
Key insights
AI's next phase is dictated by physical power and chip supply, with orbital compute and TSMC's discipline as key factors.
Principles
- Watts and wafers are AI's core physical constraints.
- TSMC's capacity decisions prevent AI market bubbles.
- Frontier AI models capture most economic value.
Method
SpaceX plans to deploy "racks in space" as satellites with solar wings and radiators in sun-synchronous orbits, connected by lasers, to create virtual data centers for orbital compute.
In practice
- Monitor TSMC's capacity expansion for market signals.
- Evaluate new chip designs for "different and hard" solutions.
- Consider extended GPU lifespans for financing models.
Topics
- AI Infrastructure
- Semiconductor Manufacturing
- TSMC Capacity
- Orbital Compute
- Frontier AI Models
- GPU Disaggregation
Best for: VP of Engineering/Data, AI Product Manager, Entrepreneur, Investor, Director of AI/ML, CTO
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Invest Like the Best with Patrick O'Shaughnessy.