AMD ships Helios rack with 2nm MI455X to challenge Nvidia
AMD put its Helios rack into production — 72 2nm MI455X GPUs plus 256-core Venice CPUs per rack — with OpenAI, Microsoft, Meta and Oracle lining up.
AMD used its Advancing AI 2026 keynote in San Francisco to move from selling chips to selling systems. Helios, its first rack-scale AI platform, is now in production: each liquid-cooled, OCP-compliant double-wide rack combines 72 Instinct MI455X accelerators with 18 sixth-generation EPYC "Venice" CPUs and Pensando networking, delivering 31TB of HBM4, 1.4PB/s of aggregate memory bandwidth and roughly 2.9 exaFLOPS of FP4 inference compute. Shipments begin in Q3 2026.
Two 2nm chips
The MI455X is AMD's first GPU on TSMC's 2nm process — 320 billion transistors across chiplets, 432GB of HBM4 and 23.3TB/s of memory bandwidth. AMD claims 50 percent more memory capacity than Nvidia's competing Rubin rack and up to 30 percent more inference tokens per dollar. Venice, the first high-performance server CPU on 2nm, packs up to 256 Zen 6c cores and 203 billion transistors at 5GHz-plus clocks. CEO Lisa Su told the audience AMD now takes 46 percent of server-CPU revenue — near parity with Intel.
The buyers are named
What separates this launch from past AMD accelerator cycles is the customer list. Microsoft will deploy Helios at scale on Azure in the second half of 2026 — the first hyperscaler commitment. OpenAI's first Helios deployment, part of its previously announced 6GW agreement, is slated for Q4. Meta is validating Venice and testing Helios; Oracle and several neoclouds are also in line. The event came a day after AMD and Anthropic unveiled a 2-gigawatt Instinct deal that includes an AMD equity investment of up to $5 billion in the lab.
Why it matters: Nvidia's dominance has rested partly on the fact that nobody else sold a complete rack-scale system buyers could deploy without assembling parts themselves. Helios ends that — same-generation silicon, full-stack integration, and named hyperscaler and frontier-lab customers with committed timelines. Whether AMD's 30-percent-per-dollar claim survives independent benchmarking, the second source the industry has demanded for three years now exists in production form.