AMD used its Advancing AI 2026 event in San Francisco on July 22–23 to lay out a full-stack answer to Nvidia, unveiling new data-center GPUs, server CPUs, a rack-scale system, and a headline commitment from Anthropic to run up to two gigawatts of AMD silicon. As SiliconANGLE put it, the company "is pushing harder than ever to grab even more market share from Nvidia Corp. in the artificial intelligence chip industry."
The Instinct MI400 Series
The centerpiece is the Instinct MI400 line. The flagship MI455X targets large-model training and high-volume inference, and AMD says it delivers 34x higher token throughput than the previous MI355X, with more than 20 PFLOPs of MXFP8 performance that doubles at MXFP4 precision. A second variant, the MI430X, is aimed at high-performance computing and sovereign-AI deployments with up to 288 TFLOPS of hardware FP64 compute, while the MI350P slots into existing infrastructure. The releases extend AMD's roughly annual accelerator cadence as buyers hunt for alternatives to Nvidia's supply-constrained parts — the same search that has driven OpenAI and Broadcom to build a custom chip and pushed Nvidia into new markets of its own.
Helios and EPYC "Venice"
AMD packaged the GPUs into Helios, a rack-scale system that ties 72 MI455X GPUs together as a single unified processor delivering 2.9 EFLOPS of AI compute and 31TB of HBM4 memory, which AMD says produces up to 30% more inference tokens per dollar than competing racks. Alongside it, the company launched its 6th Gen EPYC 9006 "Venice" server CPUs, built on the Zen 6 architecture, plus Ryzen AI Embedded X100 processors and a Kria robotics developer platform aimed at "physical AI" — chips for robots and autonomous machines.
The Anthropic commitment
AMD and Anthropic announced a strategic partnership to deploy up to two gigawatts of AMD Instinct MI450 Series GPUs in Helios rack-scale systems, with the first gigawatt set to begin coming online in the first half of 2027. As part of the deal, AMD agreed to make a strategic equity investment of up to $5 billion in Anthropic, tied to deployment milestones rather than paid upfront. It is a notable win for AMD: Anthropic has leaned heavily on Nvidia hardware and recently rented an entire 300-megawatt, 220,000-GPU data center from SpaceX to feed Claude, and the company's compute appetite has scaled with a valuation that reached $965 billion. Two gigawatts is data-center scale measured in power draw rather than chip count. Anthropic kept adding capacity outside AMD as well, signing a six-year, $10 billion compute deal with Volta in August for 121 IT megawatts of Nvidia hardware in Norway.
Anthropic has since moved on chip design itself, assembling an in-house custom silicon team while keeping its AMD, AWS, Google and Nvidia hardware.
Source: AMD.
Comments