Nvidia held roughly 86% of accelerator revenue in late 2025 on $130.5bn of sales. AMD's MI-series is competitive on raw specifications and has been for two years. It has not meaningfully moved the share.
The reason is CUDA. Two decades of libraries, kernels, framework integrations and trained engineers sit on top of it, and every one of those is a switching cost paid by the customer rather than by Nvidia. Hardware advantage erodes in eighteen months; a toolchain that everyone already knows does not.
The credible threat is not another merchant GPU — it is customers designing their own. Broadcom builds custom AI ASICs for the hyperscalers and reports an AI backlog of $73bn, with Marvell and Alchip taking similar work. That is why the share is forecast to ease to 75–81% in 2026. Note where this lands on the map: the design layer is the one chokepoint the United States holds outright, and it is the one being actively eroded from inside by its own biggest customers.