AMD: The Only Other Merchant GPUWide moat

AMD is the only company that offers what Nvidia offers: a general-purpose accelerator sold to anyone who wants one, rather than a custom part designed for a single buyer. For years that mattered less than it sounded, because the software did not exist — ROCm trailed CUDA badly enough that AMD's hardware specifications were close to irrelevant.

That gap has narrowed where it counts most. AMD's rack-scale Helios systems put 72 accelerators in an open rack against Nvidia's own rack machines, and OpenAI committed to deploying six gigawatts of AMD Instinct silicon across multiple generations, paying partly in AMD warrants¹. A buyer of that sophistication signing a multi-generation commitment is the strongest evidence yet that a credible second source exists — which is precisely what every large AI buyer has been trying to create.

What has not changed is the ratio. Nvidia holds roughly 70% of the AI-chip market²; AMD's accelerator share remains under 10%, and the software distance, while shorter, is still real for training workloads. AMD's whole data-center segment, server CPUs included, brought in $16.6 billion in 2025³, about a twelfth of Nvidia's. The honest read is that AMD is now a genuine alternative rather than a theoretical one, and that this is worth more to buyers as negotiating leverage than it is worth to AMD as revenue. Watch AMD's data-center revenue against Nvidia's, and watch whether the OpenAI gigawatts actually energize on schedule. A second source that ships on time changes the price of the first one.