AMD unveils MI455X and Helios; Anthropic commits to 2GW of it
A second source of AI compute gets real: 432GB HBM4 chips, 72-GPU Azure inference racks landing late 2026, and a claimed 15-30% edge on inference cost.
Copy markdown
The MI455X and the Helios rack, unveiled
AMD's new Instinct MI455X carries 432GB of HBM4 and 19.6 TB/s of bandwidth per chip; 72 of them form a Helios rack with 31TB HBM4, 2.9 exaFLOPS of FP4 inference and 1.4 exaFLOPS of FP8 training. It's AMD's most serious swing yet at Nvidia's rack-scale lock-in.
Anthropic commits up to 2GW — Claude's compute diversifies
Anthropic will deploy up to 2 gigawatts of MI450-series GPUs, with the first gigawatt in H1 2027, while AMD takes an equity stake of up to $5B in Anthropic. If you live in Claude Code, this is a second major supplier feeding the model's capacity, and AMD is using Claude to tune ROCm in return.
Azure will rent you AMD inference in late 2026
Microsoft announced ND MI455X v7 VMs built for production inference — 72-GPU Helios racks with 31TB HBM4 — arriving in H2 2026, plus a CPU-only HDv2 family (~500 EPYC cores, 4TB RAM) for AI data pipelines. That's a real cloud alternative to Nvidia for serving open weights.
The pitch: cheaper inference, and ROCm that finally works
AMD claims a 15-30% cost edge versus comparable Nvidia hardware on inference, and ROCm now runs as a first-class PyTorch, vLLM and SGLang backend at a reported ~90-95% of H100 throughput. Running your open models on AMD is a real option now, not a weekend science project.
Reality check: you'll reach it through the cloud first
Helios mass production is pegged at Q2 2027, with hyperscalers (Meta, Microsoft, Oracle, OpenAI) getting H2 2026 allocation and Oracle planning an MI450 supercluster in Q3 2026. Individual builders will reach these chips through rented instances long before a box ships to them.