AMD showed its 72-chip AI rack. Nvidia's rival ships later this year.
AMD published a film on 23 September 2026 showing its engineers building Helios, a complete AI server sold as one rack. It follows the assembly line and a robot arm handling boards before showing the finished racks.
According to AMD's product page, one Helios rack holds 72 Instinct MI455X graphics processors (GPUs, the chips that train and run AI models) and 31 terabytes of HBM4, the fast memory stacked next to each chip. AMD puts a full rack at up to 2.9 exaflops at FP4, a low-precision number format used for AI work. AMD calls that a theoretical peak from its own lab calculations, not a measured benchmark.
AMD expects volume deployments in the second half of 2026. OpenAI and Meta have each announced plans to deploy 6 gigawatts of AMD GPUs. Those are power-capacity plans, not counts of chips.
AMD describes Helios as a double-wide open rack that combines its GPUs, EPYC processors and networking chips.
When AMD racks reach the clouds you buy from, what would make you move a workload off Nvidia?
🎥 Video: AMD, official YouTube channel · narration: AI voice
What the captions say
- AMD engineers assemble Helios, its rack-scale AI server.
- One rack holds 72 Instinct MI455X GPUs.
- Each compute tray holds four GPUs and an EPYC CPU.
- 31 terabytes of HBM4 memory in one rack.
- Up to 2.9 exaflops of FP4 compute, AMD says.
- It is an open, double-wide rack design.
- AMD expects volume deployments in 2026's second half.
- OpenAI and Meta each plan 6 gigawatts of AMD GPUs.
