AMD has launched Helios, a 72-GPU rack-scale AI system directly challenging Nvidia’s data center dominance with claims of up to 30% more inference tokens per dollar than the Vera Rubin NVL72.
Unveiled at Advancing AI 2026, each Helios rack integrates 72 Instinct MI455X accelerators alongside 18 EPYC Venice CPUs and Pensando networking, pooling 31 TB of HBM4 memory. The MI455X uses TSMC 2nm and 3nm processes with 12 chiplets, offering 432 GB of HBM4 per accelerator and 4x the token throughput of the prior MI355. AMD claims 18x more tokens per dollar versus MI355.
Helios uses an all-Ethernet, open-standards networking backbone rather than Nvidia’s proprietary NVLink. Schneider Electric released a 246 kW reference design for the racks.
Three hyperscale customers emerged on stage. Anthropic committed to up to 2 GW of Helios deployment alongside a $5 billion AMD equity investment. OpenAI, holding warrants for up to 10% of AMD stock under a 6 GW agreement, has been running GPT workloads on Helios for three months. Meta is validating Helios under a separate 6 GW deal.
The 6th Gen EPYC Venice is the first x86 server CPU on TSMC 2nm, scaling to 256 cores and targeting agent sandbox workloads. AMD projects the server CPU market exceeding $200 billion by 2030.
For enterprise, AMD launched the air-cooled MI350P delivering up to 4.2x more tokens per second per dollar than competitors. ROCm.Ai connects AMD’s software platform to coding agents including Claude, Codex and Cursor for AI-assisted kernel optimization. AMD also partnered with Cerebras for disaggregated inference, pairing Helios with wafer-scale decode engines for up to 5x throughput on speed-critical tokens.
Helios shipments begin in the second half of 2026. Helios 500 with MI500 GPUs follows in 2027, and Helios 600 in 2028.
