AMD Unveils 72-GPU AI Rack, Citing Anthropic as a 2-Gigawatt Customer

AMD's first standalone AI conference produced a concrete procurement option rather than another benchmark slide. At Advancing AI 2026, AMD unveiled Helios, a 72-GPU rack built around its new MI455 accelerator on TSMC's 2 nm and 3 nm nodes, with 432 GB of HBM4 memory and 20 PFLOPS of FP8 throughput per rack, per EE Times' on-site report.

Anthropic, one of the AI labs building frontier language models, signed on as lead customer: up to 2 gigawatts of Helios capacity, with 1 GW scheduled for the first half of 2027. AMD is investing $5 billion in Anthropic as part of the deal. OpenAI disclosed 6 gigawatts of AMD hardware in deployment and said MI455 has been running GPT-class workloads in its data centers for three months. It is the first time AMD has shown a rack-scale system and a named frontier-lab reference customer at once.

The competitive case rests on AMD's claim that MI455 delivers up to 18× more tokens per dollar than its MI355 predecessor. Cerebras will supply its wafer-scale chip for inference disaggregation, which AMD says yields 5× throughput on the fastest tokens when paired.

Those are vendor numbers. Independent cost-per-token benchmarks, the 2027 delivery of 2 GW of Helios, and the readiness of AMD's ROCm software, its main competitive gap, will determine whether the option becomes a procurement decision.