The Hyperscalers: Half the Data CenterThin moat

Hyperscale customers accounted for 55% of Nvidia's data-center revenue in the second quarter of fiscal 2027, down from 59% a year earlier, with the remainder spread across AI clouds, enterprises, industrial users and sovereign projects¹. Given data centre is around 90% of the company, that makes a handful of American cloud operators the single most important fact about Nvidia's revenue.

They are also the best possible customers in the short run and the most dangerous in the long run. Best, because their capital budgets are enormous, their purchasing is programmatic, and they compete with each other for AI capability in a way that makes under-buying riskier than over-buying. Dangerous, because every one of them — Google with TPUs, Amazon with Trainium, Microsoft with Maia, Meta with MTIA — is funding an alternative to the thing they are buying, which is why this company's root threats include a page devoted to customers turned competitors.

The reassurance so far is that custom silicon has grown alongside Nvidia purchases rather than instead of them: no hyperscaler has yet reduced its Nvidia orders because its own chip worked. The question is whether that reflects a permanent division of labour — custom parts for known internal workloads, Nvidia for everything else — or simply that demand has been growing faster than anyone's ability to substitute. Watch the hyperscale share of data-center revenue: a decline would be the first sign that substitution has started, and it would show up here before it shows up in the headlines.