AMD unveils Helios rack system, MI455X GPU, and Anthropic partnership

AMD kicked off its Advancing AI conference with a series of announcements designed to challenge Nvidia's dominance in AI infrastructure. The centerpiece is Helios, a rack-scale system packing 72 Instinct MI455X GPUs, 31 terabytes of HBM4 memory, and 2.9 exaflops of FP4 inference compute — AMD's direct answer to Nvidia's NVL72. Source: The Next Web The MI455X GPU, built on the new CDNA 5 architecture, features 320 billion transistors and delivers theoretical peak performance up to 40.26 PFLOPS in MXFP4 format. Source: Tom's Hardware AMD also introduced a Venice-X CPU for data centers, expected in 2027. Source: TechCrunch

Helios and open standards

Helios uses 18 compute trays, each with four MI455X GPUs and an EPYC Venice CPU. The system relies on open standards: UALink for scale-up interconnect and Ultra Ethernet for scale-out, aiming to attract data center operators wary of vendor lock-in. Source: The Next Web Engineering samples ship in H2 2026, with mass production in Q2 2027. Source: The Next Web AMD claims Helios beats Nvidia's Vera Rubin on several metrics. Source: TechCrunch

Partnerships and investments

AMD committed up to $5 billion to Anthropic, which will deploy 2 gigawatts of Helios GPUs to run Claude. The first gigawatt arrives in H1 2027. Source: CNBC, The Verge The deal includes a multi-year engineering collaboration: Anthropic's Claude will optimize workloads for AMD Instinct GPUs and accelerate ROCm software development — addressing AMD's key software gap versus CUDA. AMD also partnered with Cerebras, whose wafer-scale chips will be integrated into Helios systems, claiming 5x higher tokens per second per watt.