AMD Helios Launch: First Rack-Scale Rival to Nvidia
3 min readAMD just put its first full rack of AI silicon on the table, and Microsoft signed up to buy it. On Monday the chipmaker launched Helios, its first rack-scale AI system, and announced that Azure will deploy the racks for frontier model inference starting in the second half of this year.
Why rack-scale matters
For most of the AI boom, AMD sold pieces. Customers bought Instinct GPUs and assembled the surrounding system themselves. Nvidia, by contrast, has spent the last two years selling the whole cabinet: GPUs, CPUs, networking and cooling delivered as one pre-integrated unit. That packaging is a large part of why Nvidia held its lead, because buying a finished rack is far simpler than sourcing and tuning the parts.
Helios is AMD’s answer to that. It is a single system rather than a component, which is the first time AMD has competed with Nvidia on the same terms.
What is inside Helios
Each rack packs 72 Instinct MI455X GPUs alongside AMD’s sixth-generation EPYC Venice CPUs and Pensando networking hardware. AMD says a single rack delivers up to 1.4 exaFLOPS of FP8 compute and 2.9 exaFLOPS at FP4, with 31 terabytes of HBM4 memory on board. The underlying MI450 series moves to a new CDNA 5 core architecture with 432GB of HBM4 per GPU.
Microsoft is the headline customer, joining Meta, OpenAI and Oracle as early adopters. According to CNBC, AMD says eight of the top ten AI companies now run workloads on Instinct GPUs. Engineering samples and low-volume production begin in the second half of 2026, with a mass production ramp expected by the second quarter of 2027.
Why it matters
The timing is pointed. AI chip stocks have taken a beating this month, with the semiconductor index shedding more than 20 percent from its June peak as investors started asking harder questions about the roughly $725 billion hyperscalers plan to spend on AI infrastructure this year. A credible second supplier changes that math. If buyers can genuinely choose between two rack-scale vendors, pricing power shifts and the cost of serving a model comes down.
The caveat is delivery. Announcing a rack and shipping it at volume are different problems, and Nvidia is not standing still. Watch AMD’s Advancing AI event this week and the first Azure deployments for evidence that Helios performs outside the spec sheet.
For now, Nvidia has a real rival in the one part of the market it had entirely to itself.