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

What It Covers

Gavin Baker, CIO of Atreides Management, analyzes the two physical constraints shaping AI's next phase: power (Watts) and semiconductor manufacturing capacity (Wafers). He covers TSMC's role in preventing an AI bubble, orbital compute as a long-term power solution, GPU disaggregation extending hardware lifespans, and where economic value accrues across the AI stack.

Key Questions Answered

• •TSMC as Bubble Prevention: TSMC's deliberate capacity restraint may be the single most important variable preventing an AI infrastructure bubble. If TSMC supplied all latent demand, NVIDIA could potentially sell $2-3 trillion in GPUs in 2026-27, likely triggering oversupply. Investors should monitor TSMC's capacity announcements as the leading indicator of whether the AI buildout stays rational or tips into a debt-fueled collapse similar to 2000.
• •Power Shortage Timeline: The current energy shortage constraining AI data centers begins easing in 2027-28 as new generation capacity comes online. Long-term, orbital compute — racks in space connected by laser links through vacuum, powered by sun-synchronous solar arrays — solves the problem structurally. SpaceX already operates 20-kilowatt satellites; scaling to 100-120 kilowatts per rack is their stated near-term target, making this commercially viable sooner than skeptics assume.
• •GPU Useful Life Extension: The disaggregation of prefill (memory-capacity-bound) and decode (memory-bandwidth-bound) inference workloads extends GPU useful lives from 2-3 years to potentially 10-15 years. Older Hopper and Ampere GPUs handle prefill while newer chips handle decode. This structural shift lowers financing costs from ~7% toward 5-6%, materially reducing AI buildout economics and potentially rescuing private credit funds that underwrote GPU loans at 3-4 year assumptions.
• •Frontier Token Premium Durability: Despite open-source and cheaper models proliferating, the overwhelming majority of economic value at the model layer continues accruing to frontier tokens from Anthropic, OpenAI, and XAI. The shift from flat-rate ($250-300/month) to usage-based enterprise pricing — where rate-limited consumer plans deliver degraded outputs — structurally expands revenue. Baker projects OpenAI and Anthropic could each exceed $200 billion ARR, driven by this pricing model transition rather than purely user growth.
• •Chip Startup Strategy Framework: New chip companies should pursue architectures that are both differentiated AND hard to replicate — not incremental GPU improvements. NVIDIA can copy any easy trade-off and undercut on Taiwan Semi pricing. Cerebras demonstrates the right approach: wafer-scale computing is genuinely difficult to replicate. The disaggregation of prefill and decode creates distinct optimization canvases. Baker's rule of thumb: 1% GPU market share equals roughly $100 billion in value, making narrow differentiation viable as a venture outcome.
• •AI Application Layer Value Destruction: Net value at the application layer has been destroyed by AI, even accounting for successes like Cursor and Cognition. Companies capturing value today share one characteristic: the highest ratio of utilized GPUs per human employee. Founders building vertical AI applications face a structural risk — model companies are expanding into niches faster than startups can build data moats. The token path framework (being embedded in inference workflows like Databricks) offers the clearest path to durable application-layer value.