AMD Unveils Helios AI Rack With 72 MI455X GPUs and 2.9 Exaflops, Directly Challenging Nvidia

My take: This one genuinely excited me. Today at AMD's San Francisco keynote, the company unveiled its Helios AI rack: 72 MI455X accelerators, 2.9 exaflops of FP4 inference, 31 TB of pooled HBM4 memory, at $5.25 million per unit. Nvidia has dominated this market with very little real competition, and Helios arrives with an open software stack (ROCm) that shifts that dynamic.

For anyone working with AI infrastructure or making hardware purchasing decisions, this matters in practical terms: when a real competitor enters this market with comparable specs, pricing across the industry tends to follow. GPU cloud customers should see better options over the next 12 to 18 months.

If your company or team relies on GPU clouds to run large models, direct AMD-vs-Nvidia competition at the rack level is already a variable worth tracking. Have you evaluated what it would cost to move part of your workload to infrastructure with an open stack?