AMD’s Helios Rack System Takes Aim at Nvidia – But Don’t Count Microsoft as a Customer Yet

For years, AMD has chipped away at Nvidia’s dominance in AI hardware, winning customers with its Instinct GPUs and EPYC CPUs. But the battle has always been at the component level. Now, AMD is taking the fight to the rack. This week, the company publicly detailed Helios, its first purpose-built rack-scale AI system, designed to compete head‑to‑head with Nvidia’s GB300 NVL72 and upcoming Vera Rubin platforms. Helios is not just another GPU server — it’s a complete, optimized rack solution that AMD says can deliver up to 3 AI exaflops (FP4) per rack and offers 50% more memory capacity than Nvidia’s next‑generation system. Yet amid the hype, one key piece of the story remains unconfirmed: Microsoft is not yet a named Helios customer, despite reports suggesting otherwise.

AMD first previewed Helios at the OCP Global Summit in San Jose, showcasing a static rack aligned with Meta’s Open Rack Wide (ORW) standard. The system is built on a next‑generation architecture that includes AMD Instinct MI455X / MI400‑series GPUs, EPYC “Venice” CPUs (Zen 6‑based), and Pensando “Vulcano” NICs for scale‑out networking — all unified under the ROCm open‑source software stack. AMD describes Helios as the “blueprint for yotta‑scale AI infrastructure,” targeting trillion‑parameter model training and high‑volume inference.

Performance figures are staggering. HPE, which will be one of the first to offer a turnkey Helios rack, specifies 72 MI455X GPUs per rack, delivering up to 2.9 AI exaflops FP4, 260 TB/s aggregated scale‑up bandwidth, 31 TB of HBM4, and 1.4 PB/s of HBM bandwidth. For context, each MI400‑class GPU is said to pack about 40 petaflops FP4, 432 GB of HBM, and nearly 20 TB/s of memory bandwidth. These numbers are clearly aimed at hyperscalers and cloud giants building frontier AI clusters.

For years, AMD has sold GPUs and CPUs that compete with Nvidia’s chips, but the real battle in AI infrastructure has shifted to the rack level. Nvidia’s GB300 NVL72 and Vera Rubin systems integrate Grace CPUs, Blackwell GPUs, and NVLink into a tightly optimized package that hyperscalers can deploy at scale. Helios is AMD’s attempt to match that integration — but with a twist: it’s built on open standards like the OCP Open Rack Wide spec, originally contributed by Meta, and uses the ROCm software ecosystem. AMD’s messaging emphasizes openness, serviceability, and interoperability, in contrast to Nvidia’s more proprietary approach.

The confirmed ecosystem is already impressive. Meta worked with AMD to define the ORW base. HPE has announced its own Helios‑based turnkey rack, available worldwide in 2026. Celestica is designing and manufacturing the scale‑up networking switches. Oracle will be among the first adopters of MI355X‑powered racks along the Helios roadmap. And Supermicro unveiled a Helios rack platform at Computex 2026. These are serious partners, and AMD’s list of early adopters reads like a who’s who of cloud and infrastructure.

But here’s the catch: Microsoft and OpenAI are not listed in any of AMD’s official press releases, partner announcements, or third‑party coverage as confirmed Helios customers. A recent CNBC report claimed that “Microsoft is the newest buyer, joining Meta, OpenAI and Oracle,” but that claim is not corroborated by available primary sources. It’s possible that Microsoft is testing Helios internally or has a non‑public agreement — but as of now, the public record shows only the five partners above. Given the market’s appetite for any sign of Microsoft diversifying away from Nvidia, this distinction matters.

For enterprises and cloud providers planning AI infrastructure, Helios represents a genuine alternative to Nvidia’s rack systems — but only if they are willing to bet on AMD’s roadmap and open‑standard approach. The promise of 50% more memory than Vera Rubin and up to 36× higher performance over previous‑gen AMD racks is compelling, especially for customers running trillion‑parameter models that are memory‑bandwidth bound. The partnership with HPE means that a turnkey Helios rack will be available for purchase, not just a reference design, and the involvement of Celestica and Supermicro ensures multiple supply sources.

However, the ecosystem around AMD’s software (ROCm) is still maturing compared to Nvidia’s CUDA. While AMD has made strides, many AI workloads are still optimized for Nvidia’s stack. Enterprises should carefully evaluate whether their training and inference frameworks run well on ROCm before committing to a Helios deployment. The open standards angle is attractive for hyperscalers who want to avoid vendor lock‑in, but it may add integration complexity for smaller players.

The real test for Helios will come in late 2026, when the first commercial racks ship. Watch for announcements from Microsoft, Google, and AWS — if any of them publicly commit to Helios, it will be a major signal that AMD has cracked the hyperscale market. Also keep an eye on performance benchmarks: AMD’s claims of 50% more memory than Vera Rubin are impressive, but real‑world training throughput and power efficiency will determine whether Helios can truly rival Nvidia’s rack systems. Until then, take the headlines with a grain of salt — especially the ones that name customers not yet confirmed.