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

What It Covers

Gavin Baker and Patrick O'Shaughnessy analyze July 2025's AI market selloff, arguing that every quantitative demand metric — GPU spot pricing, token growth, hyperscaler operating cash flows — is accelerating, while the narratives driving the 40-60% drawdowns in AI stocks are largely factually incorrect, with credit risk being the sole legitimate concern.

Key Questions Answered

• •GPU Spot Pricing Signal: Blackwell GPU rental prices rose 50-60% in six to seven months, with one startup paying roughly $2 per GPU hour on a contract signed months ago now facing ~$4 per hour for an identical cluster today. This inverted pricing dynamic — where older contracted compute trades at a steep discount to spot — means hyperscalers are systematically under-earning, and operating cash flows will accelerate materially as contracts reprice at current market rates.
• •Operating Cash Flow Acceleration: Microsoft, Meta, and Amazon collectively grew operating cash flow from a 28% to 32% rate quarter-over-quarter, and adjusting for unusual EU regulatory charges, the figure reaches approximately 35%. Consensus models price hyperscaler compute monetization at Ampere-era rates — two chip generations behind current Blackwell — implying roughly $700 billion in unmodeled upside if monetization tracks even a discount to current Blackwell pricing, which would push aggregate hyperscale operating cash flow toward $2 trillion.
• •Open Source Token Economics: When open source models like GLM 5.2 or Kimi K3 take inference share from frontier models, the compute demand does not decline — a token requires identical GPU flops, HBM memory, and watts regardless of which model produces it. The only effect is margin compression at the frontier model layer, with those margin dollars shifting downstream into AI infrastructure. Investors interpreting open source share gains as bearish for compute demand are misreading the unit economics entirely.
• •LTA Game Theory and Supply Chain Lock-In: Long-term supply agreements between hyperscalers and memory/chip suppliers create asymmetric consequences for defection. Any company breaking an LTA during a perceived oversupply risks losing allocation priority when the cycle tightens — a potentially fatal competitive disadvantage. With at least four major compute buyers (Amazon, Google, AMD, NVIDIA) plus numerous startups competing for allocation, the game theory strongly favors honoring agreements, making the contracted supply base far more durable than prior capital cycles.
• •NVIDIA's Revenue-Share Model: NVIDIA is deploying a credit-wrapper structure where it takes equity stakes in compute buyers and participates in revenue-share arrangements, effectively acting as a royalty holder on AI infrastructure buildout rather than a pure hardware vendor. This model increases NVIDIA's revenue per gigawatt, strengthens competitive positioning against alternative chip vendors, and helps bridge the cash flow gap for buyers who are free-cash-flow negative while operating cash flows ramp — all while NVIDIA's forward P/E sits at its lowest level in ten years.
• •Regulatory Risk as Primary Threat: New York's data center moratorium signals a broader political risk that the AI industry has failed to counter with effective public communication. The factual case for data centers — lower local electricity prices via behind-the-meter deals, sustained high-paying blue-collar employment, hospital and school community investments — remains largely untold to ordinary Americans. A coordinated PR effort running during high-viewership events like the NFL or World Series could reframe the narrative before additional state-level restrictions materialize.