Gavin Baker, who runs the investment firm Atreides Management, argued in a post on X on 13 October that shrinking profit margins for AI model makers will push the industry to buy more computing power, not less.
Baker is an investor, and the post does not say what his funds hold. He is forecasting a market he has money in, so this is a view from someone with a position. It is one post, and it cites almost no data.
His central expectation is a “hybrid, multi-model future.” Open-weight models, which anyone can download and run, would handle the bulk of the work. Baker calls it “likely” that open-weight tokens (the chunks of text a model reads and writes) become the overwhelming majority of those processed. The leading closed labs still do well in his telling, though perhaps not as well as seemed possible in May. He gives no share, no date and no revenue figure for any of it.
The part that sounds backwards is Baker’s claim that when the companies building models earn less, compute demand “mechanically” goes up. Baker does not walk through the steps, so what follows is our reading, not his. If no single model can charge a premium, the price of each unit of AI work falls. Cheaper work gets used for more jobs, so the total volume rises even as profit per unit shrinks. A discount grocer earns pennies per item and still moves more food than a boutique.
Baker attaches one condition: someone has to pay for the training. He names Nvidia chief executive Jensen Huang as the “trainer of last resort.” That backstop is also a purchase, since Nvidia sells the chips training consumes, and the post offers no second volunteer.
On financing, Baker takes on a weekend dispute over Anthropic’s computing agreements. By his account, the $413 billion of contractual commitments sounds alarming, yet it translates to about $31 billion of cloud spend each year, which he calls easily financed from its reported revenue run rate. A further $161 billion goes to Broadcom, which he expects Broadcom and private credit to fund. The post names no source for those figures and does not say what period the $413 billion covers.
He also cites 3Fourteen Research data showing that clusters of Nvidia’s B200 chips would earn a 37 percent internal rate of return, a standard yearly-return measure, even if rental prices fell 40 percent a year. That is a conditional scenario, and he says prices are still rising. Baker adds that older cloud contracts signed in 2024 and 2025 at below-market rates will reprice higher as they expire, and he thinks the buildout may be near its peak need for outside financing. He hedges that one with “may.”
His stated worry is an actual overbuild, which is why he welcomes shortages of power, chips and capital. He is also wary of concentration, but on values grounds: he does not want a handful of people steering the technology. His nearest-term bear case is a major cyber event.
Whether any of this holds is checkable soon. Baker says cloud rental prices are still climbing and that Anthropic’s commitments are fundable, and both claims will show up in contract renewals and financing announcements over the next ninety days.
Based on a post by Gavin Baker on X, 13 October 2026
