AMD Unveils Helios with World's First 2nm GPU MI455X: OpenAI, Meta, and Microsoft Among Customers

On July 23, 2026, AMD held the Advancing AI 2026 event in San Francisco, where CEO Lisa Su announced the new Instinct MI455X accelerator and the Helios rack-scale system. Helios is now in full production and will start shipping in late Q3 2026. OpenAI, Meta, Microsoft, and Oracle have signed on as customers, with OpenAI planning large-scale deployment from late 2026 and expansion in 2027. The MI455X features 320 billion transistors using a chiplet design: four XCD compute chiplets on TSMC N2 (2nm) process, two FCD cache/interconnect chiplets and I/O on N3P (3nm), and 12 HBM4 memory stacks totaling 432GB per GPU. Helios integrates 72 MI455X GPUs and 18 EPYC Venice CPUs (Zen 6, 256 cores, 2nm), delivering 2.9 ExaFLOPS FP4 and 1.4 ExaFLOPS FP8. The system uses OCP open rack standards, UALink for internal scaling, and Ultra Ethernet for external networking, with AMD Pensando networking chips. AMD emphasized an open ecosystem to attract cloud providers and OEMs. The MI455X supports CDNA 5 architecture, 256 WGP compute units, and 32-thread wavefront (down from 64), with peak MXFP4 performance of 40.26 PFLOPS per GPU. Key customers include OpenAI (up to 6GW multi-generation deal, first 1GW in H2 2026), Meta (up to 6GW, custom MI450 series), Oracle Cloud (50,000 MI450 GPUs from Q3 2026), and Anthropic (up to 2GW strategic partnership from 2027). AMD also granted warrants for up to 160 million AMD shares to OpenAI and Meta, and plans up to $5 billion investment in Anthropic. AMD's VP Andrew Dieckman claimed CUDA is becoming a "non-event" as developers shift to higher-level frameworks like PyTorch, vLLM, and Triton, though the long-term software ecosystem challenge remains significant.