Software King of the World Microsoft has turned AMD’s Helios rack from slideware into something Azure plans to run at serious scale.
AMD and Vole announced an expanded AI infrastructure pact on 20 July 2026, with Azure lined up to deploy Helios for frontier model inference.
The timing is handy for AMD, whose Advancing AI 2026 conference opens at San Francisco’s Moscone Center this week.
AMD chief executive Lisa Su is due to deliver the keynote on 23 July 2026, which gives her something better to wave about than another roadmap.
The interesting bit is not just that Microsoft is buying AMD kit. It is what sort of plumbing Azure has decided to validate. AMD and Microsoft helped found the UALink Consortium, the industry group behind the open interconnect specification that Helios uses.
That means Azure is backing an architecture both firms helped write, while Nvidia’s proprietary NVLink still dominates rack-scale AI.
Helios is AMD’s first rack-scale AI system. It packs 72 Instinct MI455X GPUs, sixth-generation EPYC Venice CPUs, Pensando networking and the ROCm software stack.
The whole lot sits in an Open Compute Project-compliant chassis with liquid cooling, because air cooling that much silicon would be optimistic to the point of comedy.
Each MI455X carries 432GB of HBM4 memory, giving one Helios rack 31TB of aggregate high-bandwidth memory.
AMD claims that adds up to 1.4PB/s of combined memory bandwidth, enough to hold giant model weights and a hefty inference cache in one rack.
The company rates the rack at 2.9 exaFLOPS of FP4 inference compute and 1.4 exaFLOPS at FP8 precision.
The selling point is the interconnect. Helios uses UALink through Ethernet as AMD’s open-standard answer to Nvidia’s closed NVLink fabric. For initial second-half 2026 deployments, AMD uses a Broadcom co-designed Ethernet switch fabric rather than native UALink switching silicon.
Purpose-built UALink switches from outfits such as Astera Labs are not expected to be production-ready until 2027.
For Microsoft, the longer game is supplier control. Nvidia’s NVLink switches are proprietary and lock customers into Nvidia’s rack-scale world.
Helios has a genuine memory advantage. Its 31TB of HBM4 per rack beats Nvidia’s Vera Rubin NVL72 figure of 20.7TB by about 50 per cent.
Nvidia still leads on some headline compute. Vera Rubin NVL72 is rated at 3.6 exaFLOPS of FP4, compared with Helios at 2.9 exaFLOPS.
Training makes the gap nastier for AMD, with Vera Rubin at 2.5 exaFLOPS FP8 against Helios at 1.4 exaFLOPS.
That is why AMD’s pitch focuses on inference growth, hardware cost advantages, and customers who would rather not rent Nvidia’s entire cathedral.
The uploaded figures say AMD hardware can carry a 15 to 30 per cent cost advantage against comparable Nvidia hardware in current cloud markets.
Microsoft joins a customer list that already includes Meta, OpenAI, Oracle and Tata Consultancy Services. Meta has apparently committed to as much as 6GW of AMD GPUs, with 1GW of Helios racks targeted this year.
