AMD launches Helios rack to rival NVIDIA
AMD's Helios rack-scale system packs 72 MI455X GPUs into one $5-5.5M unit, claims 30% more inference tokens per dollar, and lands Microsoft as launch customer.
AMD formally launched its Helios rack-scale system on July 23 at its Advancing AI event in San Francisco. One Helios rack combines 72 Instinct MI455X GPUs, sixth-generation EPYC “Venice” server CPUs, Pensando networking, and AMD’s ROCm software stack into a single unit rated at 2.9 exaflops of FP4 inference performance, priced between $5 million and $5.5 million. AMD says the design produces 30% more inference tokens for every dollar spent compared with rival racks. Microsoft is a launch customer and will deploy Helios on Azure for frontier-model inference, with shipments to customers beginning in the second half of 2026.
Helios is the same rack architecture Anthropic committed to days earlier when AMD agreed to deploy up to 2 gigawatts of MI450-series GPUs for Anthropic in exchange for as much as $5 billion in AMD equity investment; this launch is the product Anthropic’s capacity will actually run on. Pairing a named enterprise customer (Microsoft) with a named AI-lab customer (Anthropic) in the same week is AMD’s clearest signal yet that it can supply frontier-scale training and inference at a scale buyers will commit real money to, not just a roadmap slide.
For anyone tracking GPU supply as a cost or availability risk, Helios matters less for its exact benchmark numbers, which are AMD’s own, than for the fact that it is shipping with paying customers attached. NVIDIA’s competing Vera Rubin platform is making a similar performance-per-dollar pitch to its own partners, and the practical read for builders is that 2026’s capacity crunch now has two credible rack-scale suppliers instead of one, which is the first real structural pressure on GPU pricing power in over a year.