Mark RadarMARK RADAR
About
EN
Sign in

AMD Launches Helios AI Rack as Microsoft Azure Signs On

4 reports · First detected 2026-07-20 · Last active 2026-07-24

Nvidia has long dominated AI accelerators, but competition is shifting from standalone chips to fully integrated rack-scale systems combining compute, networking and software. Helios is AMD’s first such platform, pairing Instinct MI455X GPUs with sixth-generation EPYC “Venice” CPUs, Pensando networking and the ROCm software stack. Its open-standard design gives cloud providers an alternative to Nvidia’s vertically integrated Grace Blackwell and Vera Rubin systems, making hyperscaler adoption a crucial test of AMD’s ability to loosen Nvidia’s hold on AI infrastructure.

AMD formally unveiled Helios on July 23, 2026, saying each rack connects 72 MI455X accelerators and provides 31 TB of HBM4 memory and up to 2.9 exaFLOPS of FP4 compute. Shipments are due in the second half of 2026. Microsoft said on July 20 that Azure would deploy Helios at scale for frontier-model inference and Azure AI services. Futurum Group estimates each rack will cost $5 million to $5.5 million, though AMD has not confirmed pricing; the companies also did not disclose deployment volumes or the contract value.

All Coverage

4 original reports

The Backstory

The history behind this event
AMD Unveils Helios Rack-Scale AI System With MI455X Accelerators to Challenge NVIDIA2026-06-05 · 1 reports · similarity 0.87

AMD has long used its Instinct accelerators and ROCm software to challenge NVIDIA’s dominance in AI data centers. As large language model training and inference shift toward full-rack deployments, competition is expanding from individual chips to the integration of compute, memory and networking. Helios is therefore a key product in AMD’s push to win orders from cloud providers and large computing centers.

AMD unveiled Helios, its first rack-scale AI solution, at Computex 2026. Each rack contains 72 Instinct MI455X accelerators and uses HBM4 high-bandwidth memory for demanding LLM training and inference workloads. Its direct rival is NVIDIA’s next-generation Vera Rubin NVL72/VR200 platform, but AMD has yet to announce pricing or a formal shipping date.

Mark Radar|MARK RADAR

If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →

All times are in Taipei time (GMT+8)