Korva is a scroll of the internet’s best ideas.
Renting compute beats building models
Why rent compute
A company that builds data centers can rent compute to customers who are already lined up, and demand for it is intense. Tax rules add to the appeal: capital spending on AI infrastructure can be written off in the year it is made, so a 26% corporate tax rate gives back about 26 cents of every dollar spent. A Morgan Stanley report puts Microsoft's return on invested capital in tokens as a service above 30%.
Why models are riskier
Frontier models take tens of billions of dollars to build, and the frontier labs and open-weight models keep catching up to each other, so the profit from a model is far less certain than rent on a data center. Google has the largest base of enterprise and consumer customers and stakes in both SpaceX and Anthropic, so it can host any model without having the best one.
The channel conflict
Companies that do both face a conflict. Google Cloud wants every chip to rent out to Anthropic, while the teams building Google's own models want the same compute to compete with Anthropic. SpaceX rents compute to Anthropic while building Grok and Cursor. Anthropic and OpenAI have no such conflict, since neither is in the infrastructure business.
Why do rival AI models keep catching up?
Research ideas spread quickly through published papers, conference talks, and researchers who move between labs. Open-weight releases let anyone study and build on a finished model. A weaker model can also be trained on the outputs of a stronger one, a technique called distillation that Geoffrey Hinton, Oriol Vinyals, and Jeff Dean described in a 2015 paper.
Why does selling infrastructure beat betting on winners?
During the California Gold Rush that began in 1848, many miners went broke while merchants who sold supplies did well. A cloud provider's rental income comes from usage no matter which model customers choose. Sam Brannan, a San Francisco merchant, is often cited as California's first millionaire, made by selling picks, pans, and shovels.
How does channel conflict arise inside one company?
Channel conflict appears when a seller competes with its own partners or its own internal units for the same customers or supply. Compute is the shared supply here: a chip rented to an outside customer is a chip the in-house model team cannot use. Each unit is measured on its own results, so each wants the best allocation. Airlines and hotels face the same tension when they sell directly and also through travel agents and booking sites.
More from All-In with Chamath, Jason, Sacks & Friedberg
