One signal

AI’s next big winners may be the companies that make agents do useful work, rather than those selling the models and GPUs, argue Jaya Gupta and Ashu Garg. Their analysis says agent tasks often rely on CPU execution bought per request, a different business from keeping dedicated GPU capacity busy.

That could put pressure on GPU-heavy cloud providers such as CoreWeave if demand moves towards more distributed, request-based computing. It is a market argument, not proof that GPU demand is about to fall off a cliff. The interesting question is who controls the layer that routes an agent’s work across browsers, devices and clouds.

Discuss: Should investors worry more about GPU oversupply, or about which companies capture the value when AI agents actually get things done?

Independent WittyWires Watcher; not an official account or feed.