Anthropic is spending big on AMD hardware. Up to 2 gigawatts of Instinct MI455X GPUs, deployed inside AMD's new Helios racks. But the more interesting part of the deal AMD announced at its Advancing AI 2026 keynote in San Francisco isn't the order size. It's what Anthropic is giving AMD in return: access to Claude.
Under a multiyear agreement between the two companies, AMD's own engineering teams will use Claude to optimize workloads for Instinct GPUs and speed up development of ROCm, the open software stack AMD needs working well if its chips are going to compete at scale. AMD says it will also roll Claude out broadly across its engineering and product teams, not just the group building ROCm.
That's a specific kind of vote of confidence. AMD's silicon has rarely been the argument against it. The software layer around that silicon, the compilers, libraries and tooling that make a GPU actually usable at scale, has historically been the harder problem, and it's the layer where Nvidia's CUDA ecosystem has had years of a head start. Pointing an AI coding assistant at that gap is a bet that it can close faster with AI help than with headcount alone.
Helios: 90 chips a rack, and a client list to match
The headline hardware is Helios, AMD's rack-scale AI platform, now in production for what AMD calls multi-gigawatt deployment. Each rack pairs 72 Instinct MI455X GPUs with 18 sixth-generation EPYC "Venice" CPUs, tied together with AMD's Pensando networking and run on ROCm. AMD claims up to 30% more tokens per dollar than "the leading competitive solution," a phrase that avoids naming Nvidia directly while clearly pointing at it.
Anthropic isn't the only major name lined up. OpenAI plans to bring Helios online starting in the fourth quarter of 2026, pairing its own Triton framework with ROCm to optimize GPT-class workloads. Meta is validating sixth-gen EPYC platforms in its labs and testing Helios racks ahead of a larger rollout. Cerebras is pairing its low-latency inference chips with Helios for large-scale serving. Oracle, Microsoft, HUMAIN, Tensorwave, Vultr and Cirrascale round out the announced customer list, with systems shipping through OEMs including HPE, Lenovo, Supermicro and Bull.
Everything else from the keynote: robotics chips, ROCm.ai and Cisco
AMD used the event to push into adjacent territory too. It launched the Ryzen AI Embedded X100 series and new Kria AI system-on-modules aimed at robotics and other physical AI applications. It introduced ROCm.ai, a new AI-assisted layer meant to fold development, deployment and optimization into one workflow. And on the client side, AMD announced a partnership with Cisco to build Ryzen AI into enterprise collaboration devices and smart workplace hardware.
None of that changes the more immediate story. AMD frames all of this as chasing an AI computing market it expects to reach about $2 trillion by 2030, and it's convinced that closing the software distance to its rival is now partly a job for AI itself. Whether that actually narrows the gap will show up in how fast ROCm improves over the next year, not in what got said on stage in San Francisco.
