Invest Like The Best

Gavin Baker on Orbital Compute, TSMC, and Frontier Models

Key Themes

AI demand inflectioncompute constraintsorbital computeTSMC and wafersfrontier model economicschip tradeoffsvaluation dispersionAI geopolitics

1h 22mMay 20, 2026

Summary

A wide-ranging case for AI’s accelerating compute demand, looming infrastructure bottlenecks, and the long-term winners across chips, cloud, and frontier models.

This conversation argues that AI demand is still early and powerful enough to justify aggressive capital formation, even as valuations swing and infrastructure bottlenecks move from power to permits to wafer supply. Gavin Baker makes the case for orbital compute as a long-run answer to energy constraints, highlights TSMC’s centrality to the AI supply chain, and emphasizes that frontier-model economics are increasingly shaped by usage-based pricing, memory, and continual learning. He also explores how AI is changing investment baskets, chip design tradeoffs, hyperscaler strategy, and the geopolitical and safety implications of broader AI adoption.

Monitor AI demand signals, not just headlines, because the episode argues that underlying fundamentals can keep improving even during selloffs.

The speaker repeatedly frames the March/April drawdown as a buying opportunity because compute demand, ARR growth, and infrastructure bottlenecks were strengthening.

Watch TSMC wafer capacity and discipline as a key leading indicator for AI bubble risk.

The discussion argues that wafer supply, not only demand, may determine whether AI infrastructure spending becomes an overbuild or remains constrained.

Treat compute access as a strategic moat, especially for frontier model providers and hyperscalers.

The conversation suggests that where compute is allocated and who can secure it will materially shape long-term winners in AI.

Separate AI infrastructure exposure into different workload and memory regimes rather than treating the sector as one trade.

The speaker distinguishes prefill from decode, memory capacity from bandwidth, and semicap from memory, implying more selective positioning is needed.

Look for genuinely differentiated chip architectures instead of incremental GPU clones.

The episode argues that new chip entrants need hard tradeoffs and real novelty to matter, because incumbents and foundry partners can copy obvious ideas.

Expect AI monetization to keep shifting toward usage-based pricing, which could expand spend at the frontier.

The speaker describes a transition from all-you-can-eat access to pay-by-the-drink pricing that can materially lift ARR for leading labs.

Be wary of broad AI baskets; correlations and leadership can change quickly across subsectors.

The episode says cross-sectional AI trade baskets broke down and that low-quality suppliers can outperform during shortages, making factor bets less reliable.

Consider AI application businesses with a clear token path and defensible data moat, but remain skeptical of easy moats.

The speaker argues that businesses outside the token path may struggle, while vertical data advantages only matter if frontier labs do not quickly move into the niche.

Factor in safety, cybersecurity, and personal-risk planning as AI becomes more politically and socially visible.

The closing segment explicitly raises concerns about violence, political instability, and the need to overinvest in cybersecurity and contingency planning.

Select any chapter text to Deep Dive with AI

01AI’s March Inflection and the Case for Aggressive Capital

The chapter argues that March and early April marked an extraordinary AI inflection point: Anthropic’s rapid ARR growth, rising compute demand, and falling valuations for AI equities created a rare opportunity to buy into the trend. It contrasts Anthropic and OpenAI on capital efficiency and compute access, and suggests frontier model companies may want to raise large amounts of capital given strategic uncertainty and geopolitical risk.

March’s selloff is framed as a drawdown that did not change the AI thesis.

Anthropic’s ARR growth is described as unprecedented in scale.

Reasoning models are portrayed as a major driver of future compute demand.

AI and tech valuations are presented as unusually attractive relative to the broader market.

Anthropic and OpenAI are contrasted on efficiency, burn, and compute access.

Large capital raises are framed as rational for frontier AI firms facing uncertainty.

Energy and geopolitics are treated as part of the AI investment backdrop.

02Orbital Compute, Power Constraints, and TSMC’s Role in AI Buildout

The conversation explores orbital compute as a practical long-term response to power and zoning constraints, with racks in space linked by lasers rather than giant floating data centers. It also emphasizes the strategic centrality of SpaceX/Starlink-style engineering and the continuing importance of TSMC and Taiwan’s semiconductor ecosystem to AI supply chains.

Power and zoning are becoming important data-center bottlenecks.

Orbital compute is framed as solar-powered racks in space, not a floating cloud abstraction.

SpaceX and Starlink are presented as the engineering proof points for the concept.

Inference is viewed as a better near-term fit for space-based compute than training.

TSMC is described as central to Taiwan and to the AI supply chain.

AI is treated as a foundational technology that can still have a bubble without losing long-term value.

03TSMC, Capacity Constraints, and Bubble Risk in AI

This chapter weighs whether AI infrastructure spending can become a bubble and argues that current supply constraints, especially at TSMC, may delay that outcome. It contrasts today’s cash-flow-funded buildout with the debt-heavy dot-com era, then pivots to the strategic importance of Intel, Samsung, and the proposed Terafab effort before ending on frontier-model pricing and performance competition.

A bubble can emerge when enthusiasm and supply expand too far.

Today’s AI buildout is more cash-flow funded than the 2000 era was.

GPU utilization is portrayed as very high, unlike dot-com-era excess capacity.

TSMC wafer capacity is seen as a key brake on overbuild.

Intel and Samsung are possible sources of future capacity loosening.

Terafab is framed as an important U.S. fab initiative.

Frontier-model value capture remains concentrated at the top of the stack.

04Continual Learning, Frontier Tokens, and Chip Design Tradeoffs

The discussion examines whether AI may temporarily violate the ‘bitter lesson’ if systems approach ASI, how memory and continual learning could reshape model behavior, and why usage-based pricing is making frontier AI more economically powerful. It then turns to chip design, arguing that new entrants need genuinely hard tradeoffs rather than incremental GPU imitation.

The speaker thinks ASI could alter familiar AI principles like the bitter lesson.

Harnesses matter, but model quality matters more.

Usage-based pricing can expand frontier-model monetization.

Continual learning could be a major leap if solved.

Chip design is an iron-triangle tradeoff problem.

Prefill and decode represent distinct hardware bottlenecks.

Startups cannot rely on special access to TSMC roadmaps as a moat.

05Cerebras, GPU Lifetimes, and AI Application Moats

The chapter highlights Cerebras as an example of hard architectural differentiation and argues that AI infrastructure changes could extend GPU useful lives and improve financing economics. It then moves to venture strategy, data moats, token-path businesses, and the competitive pressure open-source frontier models create for closed providers.

Cerebras is presented as a rare hard-and-different infrastructure bet.

Wafer-scale computing can create unique scaling advantages.

GPU useful lives may be longer than many assume.

Lower financing rates can materially change AI buildout economics.

Founders need to know whether a business is obvious too early.

AI startups need defensible data advantages, but they are fragile.

Open source frontier competition creates strategic game theory.

06AI as the New Machine Gun: Models, Memory, and Market Baskets

The conversation turns to AI as a tool to be mastered, then moves into market dislocations across semicap, memory, and AI baskets. It closes with a bullish view of Google’s strategic assets and a nod to Meta’s AI-first execution.

AI should be mastered as a productivity tool.

Agents are already useful for research and summaries.

Semicap and memory valuations look disconnected.

Shortages can let low-quality suppliers outperform.

AI baskets no longer trade as one unit.

Misclassified names may create opportunity.

Google’s data, compute, YouTube, and cloud are strategic strengths.

07AI Power, Safety, and Geopolitical Consequences

The final chapter compares hyperscalers, emphasizes startup engagement as a strategic edge, and then widens into political risk, battlefield AI, medical breakthroughs, and broader distribution questions around access to the best models. It ends with cautious optimism about AI’s benefits and the need for humility and broader access.

Hyperscaler compute allocation is a strategic decision.

Startup engagement can strengthen chip and platform ecosystems.

AI may destroy value at the application layer even as it creates value in infrastructure.

Political and personal safety risks are rising around AI.

AI advantage may matter on the battlefield and in geopolitics.

AI can already deliver meaningful medical breakthroughs.

Broader access to top AI systems is becoming a policy concern.