Gavin Baker argues AI fundamentals are improving even as AI stocks fall roughly 40 to 60 percent, citing rising compute rental prices up 50 to 60 percent in six or seven months.
Baker says hyperscaler gigawatts are modeled to monetize at Ampere-era rates, which he considers unlikely, so a stronger monetization outlook would cut the roughly 700bn dollars of credit demand he sees in the models.
Baker calls Nvidia's lowest forward P/E in ten years and its credit-wrapper model clever and misunderstood, while warning that regulation is the biggest risk to AI.
AI stocks have fallen while fundamentals keep improving, Baker says
Baker argues the AI sell-off ignores improving fundamentals. Cash flow is accelerating and compute prices are climbing, while the scare stories rest on misread signals about excess capacity and open-source competition. The real exposure is credit, since debt-funded buildouts depend on gigawatts monetizing well, and regulation is the largest outside threat.
Takeaways
Fundamentals improve even as AI stocks fall. July resembled 2022, with AI names down roughly 40 to 60 percent in a straight line. Baker searched all month for one negative quantitative metric and found none. Operating cash flow at major hyperscalers such as Microsoft and Amazon accelerated from 28 to 32 this quarter.
Rising compute prices are bullish. GPU rental prices are up 50 to 60 percent in six or seven months, and older contracts trade far below spot. One renter of a Blackwell cluster pays about two dollars an hour and hopes to pay near four later.
The sell-off came from misread signals. Meta renting out compute looked like excess capacity. Then Muse 1.1 and open-source releases from Kimi and GLM shifted token mix. A token needs the same compute on any model, so cheaper routing moves margin from frontier labs without cutting compute demand.
Credit is the real risk. Real yields have risen and credit default swaps have widened. Models assume hyperscaler gigawatts monetize at Ampere-era rates, two generations behind. Better monetization would cut the roughly 700bn dollars of credit demand, and debt-funded buildouts can unwind quickly, as in the internet bubble.
Nvidia is cheap and misunderstood. Nvidia trades at its lowest forward P/E in ten years. Baker calls its credit-wrapper model, with revenue share on GPU purchases, clever. Long-term supply agreements now matter more than price, and breaking one could cost a supplier its franchise.
Regulation is the biggest risk to AI. Baker faults the industry's public messaging. He cites New York's data center moratorium and a debunked book that overstated data center water use by ten thousand times. He argues data centers lower power prices and raise blue-collar wages.
Open source is dark matter in the data. Third-party data suggests Anthropic's growth flattened slightly. Baker replies that OpenAI and open-source providers are accelerating. Open source is hard for public markets to measure but shows up in inference clouds, which grow almost as fast as early frontier labs while burning little cash.
Tap a name to go there · tap a line to see why they are linked