Summary of Gavin Baker - AI Market Jitters - [Invest Like the Best, EP.485]

Overview of Gavin Baker - AI Market Jitters - [Invest Like the Best, EP.485]

This episode is a deep dive into the recent AI selloff and why Gavin Baker believes the market has become too pessimistic relative to the underlying fundamentals. The conversation centers on accelerating AI demand, persistent GPU shortages, rising compute prices, the evolving role of open source models, and the growing importance of financing structures like LTAs and credit wrappers. Despite the market jitters, Baker argues the AI buildout is still constrained by supply, not demand, and that public markets are likely underestimating both infrastructure monetization and the durability of AI usage growth.

Main Themes

AI fundamentals are still accelerating

Baker’s core argument is that the market drawdown in AI stocks has not been driven by weak fundamentals.

• He says he has not found a single meaningful quantitative sign of AI demand deceleration.
• In his view, every major metric is still moving up:
  • GPU availability remains tight
  • GPU rental prices are rising
  • DRAM spot prices are rising
  • token growth is accelerating
• He believes public markets are missing the strength of private AI companies, especially:
  • OpenAI
  • Anthropic
  • open-source inference clouds
  • AI-native startups using a mix of frontier and open-source models

The market selloff was driven by narrative shocks

Baker walks through several catalysts that spooked investors:

• Meta’s compute rental move was misread as a sign of weak CapEx, when he thinks it reflected strong economics for existing compute.
• Open-source model improvements, especially from models like GLM 5.2 and Kimi K3, triggered fears of margin compression.
• China’s progress on DUV equipment added to semiconductor concerns.
• Rising real yields and widening credit spreads created anxiety around how future compute expansion would be financed.

His view: these are real market concerns, but they do not yet outweigh the evidence that AI usage is expanding rapidly.

Compute, Pricing, and the Capital Cycle

Installed compute is being repriced higher

One of Baker’s most important points is that older compute contracts are rolling off into a much higher spot market.

• In 2024 and 2025, many people expected GPU prices to decline gradually.
• Instead, Baker says prices for older GPUs have gone vertical.
• Companies that locked in long-term agreements now have contracted compute far below market-clearing prices.
• As those contracts expire, he expects significant repricing upward.

Operating cash flow may fund most of the buildout

Baker argues that consensus may be underestimating hyperscaler operating cash flow.

• He believes Microsoft, Meta, and Amazon are already showing faster operating cash flow growth.
• If compute is monetized closer to current spot rates, hyperscaler cash generation could rise sharply.
• That would reduce the need for external credit and improve the financing picture.

Credit matters, but maybe less than feared

The biggest macro worry is whether AI infrastructure buildout becomes debt-dependent.

• Real yields are up
• Credit spreads are wider
• CDS on tech names has moved higher

Baker agrees this would be dangerous if the buildout had to be financed heavily with debt. But he thinks the rising value of installed compute and improving cash flows could make the whole system more self-funding over time.

Open Source vs Frontier Models

Open source is not necessarily bearish for AI demand

Baker strongly pushes back on the idea that open source model adoption is negative for infrastructure.

• His view: a token is a token from a compute perspective.
• If cheaper open-source models take share from frontier models, the margin pool changes, but the compute demand does not disappear.
• Lower model prices may actually increase token consumption and overall compute usage.

Frontier models still capture most value, but not all of it

He expects the economic pie to keep growing, while frontier labs may no longer capture nearly all of the value.

• Frontier models are likely to retain premium economics
• Open-source models may process a majority of total tokens
• Inference clouds can customize, fine-tune, and route workloads across models, making AI-native software more defensible

Multi-model architecture is becoming the norm

Baker sees the future as increasingly multi-model:

• Use frontier models for planning and high-value reasoning
• Use open-source or cheaper models for routine work
• Route workloads dynamically based on cost, quality, and task type

This, he argues, is good for customers and still bullish for infrastructure because it increases total token throughput.

Risks and Watchpoints

The main risk is regulatory

If Baker had to name the biggest risk to AI, it would be regulation.

• He cites New York’s data center moratorium as a warning sign
• He thinks public narratives around data centers are distorted:
  • people fear higher electricity bills
  • water usage is often overstated
  • job creation is underappreciated
• He argues data centers often bring:
  • lower local power costs
  • better community investment
  • high-paying blue-collar jobs
  • ongoing maintenance and upgrade work

The real concern would be a sustained demand slowdown

What would change his mind?

• If operating cash flow does not continue to accelerate
• If GPU prices materially and sustainably collapse
• If the major AI labs plateau or decline
• If compute stops being scarce

So far, he says, none of that is happening.

He is watching long-horizon technical breakthroughs

Baker is especially interested in possible advances in:

• continual learning
• sample-efficient learning
• model efficiency and disaggregation of inference workloads

He notes that these innovations would be good for the world, but could eventually alter the shape of compute demand. For now, he thinks they are more likely to increase AI adoption than reduce infrastructure value.

Notable Company and Industry Observations

NVIDIA still looks strategically advantaged

Baker is notably bullish on NVIDIA’s positioning:

• It is central to financing the AI buildout
• It has leverage through equity stakes, partnerships, and revenue-sharing structures
• It can support customers through “credit wrapper”-style arrangements
• He thinks the market misunderstands how strong this business model could become

He also finds NVIDIA’s valuation surprisingly low given its ecosystem power and strategic optionality.

SpaceX may be a major AI infrastructure story

A surprising thread in the conversation is Baker’s belief that SpaceX could become an important compute company.

• He references speculation around large-scale power bring-up
• Starbase and orbital compute are becoming more real
• He thinks public markets may be underpricing SpaceX’s role in future compute infrastructure
• He sees orbital compute and Starlink-related technology as potentially meaningful long-term

Memory and LTAs are becoming a strategic game

Baker discusses how long-term agreements are changing the dynamics of the memory market.

• Companies are trading short-term upside for supply certainty
• Breaking an LTA could damage future allocations
• Memory, HBM, and supply-chain control may shape market share for years
• This is a more interdependent and fragile market than past cycles

Baker’s Bottom Line

His view is still bullish, but with humility

The episode has a strong “pressure test” feel. Baker says he spent the month actively trying to find negative data and came away more constructive than before.

His main conclusions:

• AI demand is still accelerating
• Compute remains scarce
• Market fears are real, but likely overstated
• Open source strengthens the ecosystem rather than killing demand
• Financing risk is the biggest near-term watch item
• Regulation is the largest systemic risk

Takeaways for Investors

• Don’t confuse a selloff with weakening fundamentals.
• Watch compute prices, not just public-market sentiment.
• Follow hyperscaler operating cash flow closely.
• Open source may shift margins, but not necessarily total AI demand.
• Credit markets and LTAs are becoming critical to the AI capital cycle.
• Regulation and public narrative remain major outside risks.

Closing Thought

The episode is essentially a reminder that AI is still in a scarcity regime. Baker’s thesis is that the market is pricing in too much normalization too soon, while the actual evidence points to continuing shortage, accelerating usage, and expanding monetization. In his view, the AI story is not cracking — it is still compounding, just in a more complex and capital-intensive way than many investors expected.