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

Overview of Invest Like the Best — Gavin Baker, “Watts and Wafers” (EP. 473)

In this wide-ranging conversation, Patrick O’Shaughnessy and Gavin Baker dig into the two physical constraints that Baker thinks will shape the next phase of AI: watts (energy) and wafers (chip supply). Baker argues that AI demand has been far stronger than most people appreciate, that frontier models are still capturing most of the economic value, and that the biggest investor mistake today is underestimating the scale and speed of the buildout. He also lays out a provocative thesis on orbital compute, the strategic importance of TSMC, the rise of custom chips, and why AI may be entering a bubble-like phase even as the fundamentals remain extraordinarily strong.

Main Themes

AI demand is still exploding

• Baker believes recent market volatility has obscured a historic acceleration in AI demand.
• He points to Anthropic’s rapid revenue growth as evidence that AI is creating value at a speed unlike anything in prior software cycles.
• His broader view: this is not a normal tech adoption curve; it may be the most extraordinary moment in the history of capitalism.

Watts: energy is a real constraint, but likely temporary

• Baker thinks the current shortage of power for AI data centers is real but not permanent.
• He expects the near-term energy bottleneck to ease around 2027–2028 as new supply comes online.
• Longer term, he sees orbital compute as a plausible solution if terrestrial energy, zoning, and permitting become binding constraints.

Wafers: chip supply is the more important bottleneck

• On chips, Baker argues the key variable is not just demand, but TSMC’s capacity decisions.
• He believes TSMC’s discipline may be the single most important factor in preventing a classic overbuild/bubble in AI hardware.
• He contrasts the current buildout with the dot-com era, emphasizing that today’s AI infrastructure is still mostly funded by operating cash flow, not debt.

AI Economics and Frontier Models

Frontier models still seem to capture most of the value

• Baker is surprised by how much economic value remains concentrated in frontier models like OpenAI, Anthropic, and xAI/Grok.
• He argues that, for now, the frontier labs are still on the Pareto frontier of intelligence vs. cost.
• A major open question for investors: Will model-layer economics stay concentrated at the frontier, or will value shift toward applications?

Usage pricing is making AI more monetizable

• Baker says the industry is shifting from “all-you-can-eat” to usage-based pricing, which should improve economics for model providers.
• He thinks this is one reason frontier labs can keep scaling revenue rapidly.
• He also warns that many users are not seeing the true capabilities of frontier AI because cheaper tiers are heavily rate-limited.

The “bitter lesson” remains a key risk

• Baker repeatedly returns to Richard Sutton’s idea that brute-force scale often beats human ingenuity.
• His concern is that some optimization tricks or efficiency improvements could temporarily slow AI demand, but he thinks more compute usually wins in the end.
• He’s especially interested in whether this principle still holds as models approach ASI-level capability.

Chips, Foundries, and the Supply Chain

TSMC is the central strategic actor

• Baker sees TSMC as the dominant gatekeeper in the AI hardware era.
• If TSMC expands too much, it risks enabling a bubble; too little, and supply stays artificially constrained.
• He thinks maintaining a moderate lead over Intel and Samsung while avoiding a runaway buildout is the “Goldilocks” scenario.

Elon’s “TerraFab” could matter

• Baker discusses Elon Musk’s proposed U.S. fab effort as a potentially important long-term variable.
• He believes Elon can attract top engineering talent and equipment vendors because of his track record in hardware.
• Still, he views it as a long-lead project rather than an immediate market factor.

Custom silicon and new chip companies are part of the answer

• Baker is bullish on experimentation in chip design, but only if it is both different and hard to copy.
• He thinks there is room for players like:
  • Google TPU
  • Amazon Trainium
  • Cerebras
  • other specialized architectures for prefill, decode, or other AI workloads
• His rule of thumb: even a 1% share in a huge market can be a venture-scale outcome.

Disaggregation of AI workloads is opening new design space

• He explains the split between:
  • Prefill: understanding the prompt/context
  • Decode: generating tokens
• Because these are different bottlenecks, chip designers can make more specialized trade-offs.
• This helps explain why non-GPU architectures may succeed in targeted niches.

Orbital Compute: Baker’s Most Futuristic Thesis

Data centers in space are not fantasy, in his view

• Baker pushes back on the idea that orbital compute means giant floating buildings.
• He describes it instead as racks in space, connected by lasers, powered by solar arrays, and cooled through radiators.
• He argues SpaceX is uniquely positioned to solve the engineering problems because of its reusable rockets, satellite fleet, and hardware talent.

Why it matters

• If terrestrial data centers run into zoning, permitting, or energy limits, orbital compute could become an important pressure valve.
• He thinks inference workloads are especially well suited for space, while training will remain on Earth for a long time.
• Even if orbital compute never dominates, he expects it to force a rethinking of the entire compute stack.

Public Market Implications

Baker’s watchlist: valuations and shortages

• He is increasingly concerned about a diversity breakdown in market thinking, where too many investors become uniformly bullish.
• At the same time, he notes that valuations across AI subsectors are wildly inconsistent:
  • some high-quality names remain reasonably priced
  • lower-quality shortage beneficiaries can be massively overbid
• He sees the biggest opportunity in distinguishing high-quality compounding businesses from speculative “shortage” names.

The current cycle looks like a commodity boom in some ways

• Baker compares the AI hardware cycle to commodity cycles:
  • the hardest-to-love, highest-cost suppliers can rise the most during shortages
  • but that can be a sign of a late-cycle dynamic
• He thinks investors need to be extremely careful about baskets and factor exposure, especially as correlations within AI have started to break down.

Company-by-Company Takeaways

Google

• Still strong because of its enormous installed base of compute, data, and YouTube/search assets.
• Baker was impressed by Google’s earlier TPU advantage, though he thinks that edge has narrowed.
• He sees Google as strategically resilient even if it doesn’t always lead the model race.

Meta

• He gives Mark Zuckerberg credit for making Meta an AI-first company internally.
• Meta’s pace of change matters more than its absolute position, at least over shorter time horizons.
• He was pleasantly surprised by the quality of Meta’s newer models.

Amazon

• Baker is bullish on Amazon because of Trainium and its ability to apply AI internally.
• He expects meaningful P&L gains from robotics and automation in retail.
• Amazon is also one of the most engaged big companies with startups and AI infrastructure.

Microsoft

• He respects Satya Nadella but thinks Microsoft has made some difficult trade-offs.
• Baker argues Microsoft is now more focused on building its own AI products than simply serving OpenAI’s infrastructure needs.
• He views this as risky, but strategically sensible.

NVIDIA

• Still the central force in AI infrastructure.
• Baker thinks NVIDIA could respond quickly if a competitor gets traction, but its lead comes from deep ecosystem control, supply-chain relationships, and constant optimization.
• He believes NVIDIA’s relationships across the ecosystem make it extremely hard to displace.

Broader Risks and Societal Effects

AI may be raising geopolitical and personal risk

• Baker is increasingly worried about political violence and public backlash aimed at visible AI leaders.
• He thinks AI is already affecting warfare, pointing to Ukraine’s battlefield AI as an example of strategic advantage.
• He sees AI as potentially destabilizing for rivals, even while being highly beneficial for the U.S.

He remains strongly pro-AI overall

• Baker is not a skeptic; he believes AI will produce enormous benefits in medicine, productivity, and scientific discovery.
• He tells a moving example of AI helping a family work on a rare disease case.
• Still, he believes society needs to manage the transition carefully, especially because access to the best AI is increasingly concentrated among wealthy users and large institutions.

Key Takeaways

• AI demand is real and still accelerating.
• Power and wafers are the two constraints that matter most.
• TSMC’s capacity decisions may shape whether the cycle becomes a bubble.
• Frontier labs are still capturing most of the value—for now.
• Custom silicon and specialized architectures have real opportunity if they’re truly hard and different.
• Orbital compute is no longer a joke in Baker’s framework; it’s a credible long-term solution.
• Investors should watch for bubble dynamics without underestimating the strength of the fundamentals.

Notable Perspective

“History doesn’t repeat, but it rhymes.”

That quote captures Baker’s stance throughout the episode: this looks like a foundational technology cycle with real shortages, real capital formation, and real bubble risk—yet also with extraordinary long-term upside if the infrastructure and economics keep improving.