⚠ Hyperscaler Self-SupplyHigh threat
Nvidia (NVDA) — threat to the moat
Every chip the clouds build themselves is a chip they don't buy — and the temptation grows with every point of Nvidia's margin.
The sharpest expression of the concentration risk is that Nvidia's largest customers are actively building the chips they currently buy from it. The biggest cloud companies have each invested heavily in their own AI silicon1 — precisely to cut their bills and their dependence — and they possess the scale, the talent, and the capital to make the effort worthwhile even if their in-house chips serve only a portion of their workloads. Every chip they build themselves is one they do not buy from Nvidia.
The threat is uniquely dangerous because it comes from within the customer base and grows more tempting the more Nvidia charges. The fatter Nvidia's margins, the greater the prize for a customer who can bring even part of the work in-house, so Nvidia's own pricing power funds and motivates the very effort to escape it. And these customers are not marginal — they are the concentrated few on whom a large share of Nvidia's revenue depends.
Nvidia's bulwark is the same thing that protects it against everyone: the software and systems moat. A hyperscaler can design a capable chip, but getting the whole sprawling world of AI software to run on it as well as it runs on Nvidia is far harder, so the in-house chips have tended to handle specific, stable workloads while the general, cutting-edge work stays on Nvidia. The pace of Nvidia's improvement also keeps the home-grown chip chasing a moving target.
This risk runs high — the highest on the data-center front — because this threat combines the concentration risk with the customers-turned-competitors risk in the single most motivated, best-resourced set of buyers. It is unlikely to displace Nvidia at the frontier soon, but it will steadily claim the commoditized, high-volume workloads, capping Nvidia's share and pressuring its margins over time, and it comes from exactly the customers Nvidia can least afford to lose.
- ReportedEach of the biggest clouds has invested heavily in its own AI silicon (TPU, Trainium, Maia).AWS Trainium & Microsoft Maia — hyperscaler custom AI silicon programs (plus specialized inference startups) — Announced/shipping 2023–2026 · publ. 2023–2026 · source ↗