What happened
AMD announced the MI400 accelerator family and Helios rack-scale architecture on July 23. The company said Helios was in production and described it as AMD’s first rack-scale AI solution.
Each rack integrates 72 Instinct MI455X GPUs, 18 sixth-generation EPYC CPUs, Pensando networking, and ROCm. AMD said systems would be available through OEMs including Bull, HPE, Lenovo, and Supermicro, with Sanmina and Wiwynn among the infrastructure partners.
A system-level contest
The architecture puts AMD into the same system-level contest as NVIDIA: compute silicon is only one part of a working rack. Networking, host CPUs, memory, software, manufacturing, and serviceability all affect how quickly customers can turn hardware into usable capacity.
AMD claims Helios can deliver up to 30% more inference tokens per dollar than a leading competing solution. That comparison is based on AMD’s own calculations and should not be treated as independent performance evidence.
Production is not deployment
AMD’s production milestone and its customer timelines describe different stages. According to AMD, OpenAI expects to begin bringing Helios online in Q4 2026 and accelerate deployment during 2027. Anthropic has outlined a plan for up to 2 GW of MI455X capacity. Neither plan is current operating capacity.
The commercial test is now execution: whether partners can ship integrated systems on schedule and whether customers can qualify ROCm and their workloads at the planned scale.
What to watch next
- OpenAI’s first stated Q4 2026 Helios deployment milestone.
- OEM delivery schedules and cloud availability.
- Independent workload results and evidence of software readiness at scale.