⚠ The Buyers Are Building Their OwnHigh threat

A threat that hangs over both AMD and Nvidia, but that is especially acute for the challenger, is that AMD's biggest customers are also becoming its competitors. The hyperscalers that buy AI accelerators — the same handful of giants who dominate demand — are designing their own custom AI chips: Google has its TPUs, Amazon its Trainium and Inferentia, and other cloud and AI leaders are building bespoke silicon of their own. Every workload they move onto their own chips is compute they buy from neither Nvidia nor AMD, and it shrinks the merchant market both depend on.

This is more dangerous for AMD than for Nvidia for a simple reason: AMD's whole AI thesis rests on being the second source, and the customers' in-house chips are the ultimate second source — the fullest expression of the very desire to escape single-vendor dependence that AMD is trying to monetize. A hyperscaler that has built a capable internal chip has less need for a merchant alternative to Nvidia, because it has made its own. So the same strategic impulse that pulls AMD into the market can, as it matures, push AMD back out of it, leaving the merchant vendors to fight over the workloads the giants choose not to serve themselves. AMD's counter is that custom chips are hard, narrow, and slow to develop, that general-purpose GPUs remain essential for the fast-changing frontier of AI, and that the total market is growing fast enough to support everyone. That is largely true today. But the long-run risk is real and structural: AMD is selling to customers who would, given the choice, rather not need it, and who are spending billions to make sure that one day they don't — Google's TPUs, Amazon's Trainium, and Microsoft's Maia are that spending made silicon¹.