Via nvidia.com
AMD, Google, and Cerebras are all making aggressive moves to chip away at Nvidia's roughly 81-90% share of the AI accelerator market.
Nvidia commands an estimated 81% to 90% of the AI data center accelerator market, a dominance built on years of GPU innovation and the stickiness of its CUDA software ecosystem.
AMD, Google, Cerebras, Amazon, Meta, and Microsoft are all investing heavily in alternatives, whether through merchant GPUs, custom silicon, or entirely new architectures.
The challengers are getting specific #
AMD secured deployment commitments of 6 GW each from OpenAI and Meta, plus 2 GW from Anthropic, all for its MI450 series and Helios rack-scale platforms.
Cerebras launched its CS-4 rack-scale inference system, claiming 750 PFLOPS of performance, roughly double what its previous generation delivered.
Google signed a commercial deal with Marvell Technology on July 29 to expand their custom-silicon partnership. The arrangement included a warrant for 58.97 million Marvell shares. Google is targeting general availability of its TPU v8 by late 2026.
Why customers are building their own chips #
Google has been doing this longest with its Tensor Processing Units. Amazon has its Trainium and Inferentia chips. Meta has been developing custom silicon internally. Microsoft has its Maia AI accelerator.
Training runs are massive, unpredictable, and favor Nvidia’s brute-force GPU clusters. Inference is more predictable, more cost-sensitive, and more amenable to optimization through custom silicon. That’s exactly where Cerebras, Google’s TPUs, and AMD’s new platforms are focusing their efforts.
What this means for the competitive landscape #
Nvidia’s roadmap includes the Vera Rubin platform following its Blackwell generation GPUs.
AMD can point to 14 GW of deployment commitments across OpenAI, Meta, and Anthropic, and Google is backing its custom silicon strategy with equity-linked deals with Marvell. AMD has been investing in its ROCm software stack, and hyperscalers can build custom software for their own chips, but replicating the breadth of CUDA’s ecosystem remains a multi-year project.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our