⚠ Cloud Abstracts the ChoiceModerate threat

The default-assumption moat lives in the minds of the developers who choose hardware, and the danger is that a growing share of them are ceasing to choose hardware at all. As AI is increasingly consumed as a service — a model called through an API, a capability rented from a cloud — the developer never touches the chip, never installs CUDA, and never forms a preference, because the hardware decision has moved up to the handful of cloud and model providers who run the service.

This is dangerous because it relocates the choice to exactly the parties most motivated to diversify away from Nvidia. When millions of developers each chose their own hardware, Nvidia's default swept them all in; when a few cloud providers choose on everyone's behalf, those providers — who build their own chips and negotiate hardest — decide, and they have every incentive to route workloads onto whatever is cheapest for them, Nvidia or not.

For now, those cloud providers still overwhelmingly run Nvidia today, because that is what their customers' models are built and optimized for, and because Nvidia's performance keeps it the best economic choice even for a buyer who would love an alternative. The abstraction hides the chip, but underneath the API the workload still, for now, lands mostly on Nvidia hardware.

Put it at moderate. The shift to AI-as-a-service genuinely removes the hardware choice from the many and concentrates it among the few most eager to find alternatives, which is a real long-run pressure on the default — but those few still run mostly Nvidia — even as they build TPUs, Trainium, and Maia in parallel¹ — because it remains the best-performing and best-supported option, and the models they serve are still built in Nvidia's world.