AMD's Helios Lands 72 GPUs in One Rack, and Four Hyperscalers Say Yes

The shift is small in count and large in kind. Until now, the largest AI training and inference clusters have run almost exclusively on Nvidia. Helios is AMD's attempt to sell hyperscalers something they have been begging for: a credible second source for the racks that train the biggest models, with a software stack they do not have to negotiate their way into⁠4⁠6⁠9.

Helios is not a chip. It is a full stack: AMD Instinct MI455X GPUs for the math, EPYC "Venice" CPUs for the orchestration, Pensando NICs for the fabric, and ROCm for the software that ties it all together⁠7⁠9. The published specifications are aggressive in the way AMD needs them to be aggressive. Per rack, you get seventy-two MI455X accelerators, thirty-one terabytes of HBM4 memory, 2.9 exaflops of FP4 inference performance, and 1.4 exaflops of FP8 training throughput⁠4. One industry write-up pegged the whole machine at roughly ninety racks when AMD talks about it as a single system⁠4⁠7.

The system lands about four months after Nvidia unveiled its own next-generation Vera Rubin rack, the product Helios is built to compete with⁠9. Helios itself was first announced in June 2025, more than a year before Monday's customer reveal⁠7.

The early customer list is the headline. Microsoft said on the same day that it will deploy Helios racks inside Azure, joining Meta, OpenAI, and Oracle as anchor buyers⁠2⁠5⁠6⁠11. That is three of the largest cloud platforms on earth, plus the operator of one of the most compute-hungry model fleets in existence, all committing publicly to a non-Nvidia rack on the same morning.

As for timing, the sources sketch a two-step ramp. Engineering samples are expected in the second half of 2026, with mass production following in the second quarter of 2027 according to one outlet⁠2⁠4. The narrow spread of dates across the reporting reflects how new the product still is rather than a contradiction; readers should treat 2026 as the year of pilots and 2027 as the year of volume.

The pitch to Microsoft, Meta, Oracle, and OpenAI is not just performance. It is structure. AMD is positioning Helios as an open-standards alternative to Nvidia's proprietary NVL72, Grace Blackwell, and Vera Rubin family, meaning the rack uses published interfaces and AMD's ROCm software rather than a closed ecosystem the customer must license into⁠4⁠6⁠9. For a hyperscaler spending tens of billions a year on AI infrastructure, that distinction is not cosmetic; it is the difference between a one-vendor road and a road with exits.

Counter-reading matters here too. Nvidia's moat is software as much as silicon, and CUDA has decades of accumulated model and kernel work that ROCm does not yet match. A rack that ships on open standards still has to run the models customers care about, at the throughput they expect, with the developer tooling they already know.

The launch coverage is confident about the customer list and the on-paper specs. It is silent on the numbers that will decide whether Helios matters. Several widely circulated figures from the launch chatter, including the size of the AMD team assigned to the project, the per-rack price, the per-rack power draw, the CPU count per rack, and AMD's own revenue projections, do not appear in the sourcing behind this story. They are being treated here as unverified.