AMD ships Helios rack-scale AI system; Microsoft joins Meta, OpenAI as first customers

AMD announced its first rack-scale AI system, Helios, with Microsoft, Meta, OpenAI, and Oracle as anchor customers deploying it starting in the second half of 2026. Helios integrates AMD's MI455X GPUs (72 per rack), 6th-gen EPYC Venice CPUs (256 cores each), Pensando DPUs, and ROCm software into a liquid-cooled, double-wide reference design, delivering 2.9 exaflops at FP4 precision with 31TB of combined HBM4 memory across the rack.

Helios is priced between $5-5.5 million per rack (vs. Nvidia's Vera Rubin at estimated $3.5-4M), making it heavier and wider but offering 19.6 terabits/second of bandwidth per GPU and stronger memory capabilities. The system is optimized for AI inference workloads and agent-driven tasks. AMD data center revenue grew 57% year-over-year in Q1 2026; the company expects tens of billions in annual AI data center revenue starting 2027, with Helios as the primary driver.

For architects: Helios represents AMD's first credible challenge to Nvidia's 95% GPU market dominance (AMD holds ~4.5%). By packaging GPUs, CPUs, networking, and software as one integrated rack—like Nvidia's Vera Rubin—AMD shifts competition from components to systems. Microsoft's scale commitment to deploy Helios on Azure signals confidence in ROCm's maturity for production frontier workloads. Industry analyst Daniel Newman estimates AMD could reach 20-25% market share if Helios succeeds; the determining factor will be whether ROCm can reliably host demanding AI workloads at the scale Nvidia's CUDA handles.