News

AMD Ships Helios This Year: 72-GPU AI Racks—OpenAI, Meta, Microsoft Line Up

AMD's Helios rack packs 72 MI455X GPUs and 31 TB of HBM4, shipping this year with OpenAI, Meta, Oracle, Anthropic and Microsoft all lined up to deploy it.

AMD used its sold-out Advancing AI conference in San Francisco on Thursday to position Helios, its first rack-scale AI system, as a direct shot at Nvidia’s Vera Rubin and Grace Blackwell racks, TechCrunch reported. The system, which packs 72 Instinct MI455X GPUs per rack, ships later this year.

Chair and CEO Lisa Su called Helios the tech industry’s “highest-performance AI rack,” per TechCrunch, “built to train and run the most demanding frontier models in the world at massive scale.” The customer list already reads like a who’s who of frontier AI: OpenAI, Meta, Oracle, Anthropic, and Microsoft all have deployment plans.

The announcement lands days after two related moves: Microsoft CEO Satya Nadella said Monday that Azure will expand with Helios systems, and AMD’s partnership with Anthropic—announced Wednesday—covers up to 2 gigawatts of GPUs, a deal Frontierbeat covered in its report on AMD’s multibillion-dollar Anthropic bet.

Inside the Helios rack-scale AI system

The MI455X silicon underneath is AMD’s new CDNA5 architecture: 256 workgroup processors across eight compute dies, 432 GB of HBM4 memory per GPU at 23.3 TB/s, and a peak rating of 40.26 petaflops of MXFP4 compute, according to a teardown by Chips and Cheese. That peak figure is a theoretical ceiling, not a measured result—real-world throughput will land lower.

At the rack level, Helios aggregates up to 2.9 exaflops of MXFP4 compute and 31 TB of shared HBM4, wired together with UALink-over-Ethernet delivering 3.6 TB/s of peak bidirectional scale-up bandwidth per GPU. Twelve Broadcom Tomahawk 6 switch ASICs handle the fabric, and each of the 18 compute trays pairs four MI455X GPUs with a 96-core EPYC “Venice” CPU.

The Register’s analysis puts Helios at 50% more HBM4 capacity and scale-out bandwidth than Nvidia’s Vera Rubin, 15-25% higher AI-training performance, and an AMD-estimated 30% performance-per-dollar advantage—at a rack power draw of 225-245 kW. AMD fellow Alan Smith told the outlet the MI455X’s L2 cache bandwidth is “1.5x the aggregate bandwidth of the Infinity cache on MI355X.”

A $1.4 trillion market and the agentic argument

Su’s justification for the buildout is the shift to agentic AI. “When you ask the agent to do something, it actually has dozens of steps, and it has to reason, and it has to call tools, and it has to access data, and it has to keep doing it over and over until it solves the problem, and so you need lots of GPUs to do all that,” she said, per TechCrunch.

She also raised AMD’s market forecast: “We’re now expecting that by 2030, the AI accelerator market is going to reach about $1.4 trillion. What that means is, by the end of the decade, the AI accelerator market is going to approach the size of the entire semiconductor market today,” Su said at the event.

The competitive stakes are visible across the industry. Inference-chip startup Etched just hit a $10.3 billion valuation betting against general-purpose GPUs, while AMD also previewed its Venice-X data center CPU for 2027. Helios is the company’s clearest statement yet that it intends to fight Nvidia at the rack, not just the chip.

FAQ

When does AMD’s Helios rack ship?

AMD says Helios ships later in 2026, with OpenAI, Meta, Oracle, Anthropic and Microsoft among the customers planning deployments.

How does Helios compare with Nvidia’s Vera Rubin?

The Register calculates 50% more HBM4 memory and scale-out bandwidth, 15-25% higher AI-training performance, and an AMD-claimed 30% performance-per-dollar lead, though Nvidia claims a 25% FP4 edge for inference with adaptive compression.

What is inside each Helios rack?

72 Instinct MI455X GPUs on the CDNA5 architecture, 31 TB of shared HBM4 memory, 96-core EPYC Venice CPUs, and Broadcom Tomahawk 6 switching, drawing roughly 225-245 kW per rack.