The Customers Who Design Their OwnNarrow moat

Every hyperscaler large enough to be a meaningful AMD customer is also designing its own accelerators, generally with Broadcom or Marvell. That is not a hypothetical competitive threat; it is the current arrangement, and AMD sells into it.

The logic from the customer's side is consistent. Custom silicon handles the workloads a hyperscaler runs at enormous scale and understands completely, where the design cost amortises over millions of units. Merchant accelerators handle everything else — new model architectures, external cloud customers, workloads that change faster than a chip programme can. AMD competes for the second category, and so does Nvidia.

This bounds the addressable market in a way share statistics obscure. If the largest buyers take an increasing proportion of their compute in-house, the merchant market grows more slowly than total AI spending — and AMD is competing for second place within it. The offsetting fact is that AMD's own customers include the labs and neoclouds who have no ability to design silicon and every incentive to avoid depending on one supplier.

Watch the split between hyperscaler custom silicon and merchant accelerators in industry shipment estimates. That ratio, more than AMD's share of the merchant portion, determines how large this business can become — against an accelerator position estimated at 5-7%¹ today.