Microsoft Names AMD's Helios Its First Rack-Scale AI Bet on Azure
TL;DR
- Oracle committed 50,000 MI450 GPUs for Q3 2026 deployment, adding volume specificity to hyperscaler commitments the launch announcement left unnamed.
- AMD frames Helios against Nvidia Vera Rubin NVL72 at equal rack size, with 31.1 TB of HBM4 memory and 2.9 exaFLOPS FP4 compute per rack.
- Microsoft adds three Azure VM families with Helios: ND MI455X v7 for inference, HDv2 for data prep with 32 TB NVMe, and HXv2 for EDA and HPC.
The interesting thing about AMD's announcement with Microsoft on July 20, 2026 is not that another hyperscaler is buying AMD silicon; it is that Microsoft agreed to be the named lighthouse customer for Helios, AMD's first rack-scale system, and to run it on Azure for frontier-model inference. According to AMD's newsroom, the deployment pairs Instinct MI455X GPUs with 6th Gen EPYC Venice CPUs, Pensando networking, and the ROCm software stack, sold as an integrated rack rather than as loose parts.
The rack itself is the pitch. Reporting from Tom's Hardware and other outlets peg Helios at up to 72 MI455X accelerators per liquid-cooled, double-wide rack, with a total of 31TB of HBM4 memory. That is AMD's answer to Nvidia's NVL72 form factor, and having Microsoft say publicly it will run these racks on Azure, surfaced to customers as the new ND MI455X v7 VM, is the endorsement AMD has been trying to buy with roadmap slides.
Alongside the GPU story, Azure is adding two EPYC Venice-only VM families: HDv2 for agentic AI and data-pipeline work, and HXv2 for semiconductor design. Those are less headline-grabbing than the Helios racks, but they matter for AMD's CPU business and they extend a pattern in which Microsoft is standardizing on AMD across more of the fleet, not just as a spot buy.
The honest caveat is that shipments do not begin until the second half of 2026, so this is a forward commitment rather than production capacity you can rent today. Lisa Su, chair and CEO of AMD, and Satya Nadella, chairman and CEO of Microsoft, both put their names to the announcement, but the reporting does not give the number of racks, the contract value, which Azure regions get the hardware first, or which Microsoft workloads such as Copilot or OpenAI serving will actually land on Helios versus staying on Nvidia. Take the 'at scale' phrasing as intent, not a number.
Where this goes is worth watching. If AMD executes the H2 2026 ship and ROCm holds up under real Azure inference load, the pricing conversation for the next inference generation changes for every hyperscaler still on a single-vendor GPU diet, and the second-source narrative for frontier inference stops being theoretical.
What others are reporting
-
CNBC Read →
Leads with the Nvidia rivalry frame; positions Microsoft as the latest addition to a four-hyperscaler launch cohort and tracks AMD stock reaction on the day of announcement.
-
Tom's Hardware Read →
Technical breakdown of MI455X paired with EPYC Venice across three new Azure VM families; notes liquid cooling constraints limit customer availability to 2026-early 2027.
-
Tom's Hardware Read →
Only source that quantifies Oracle's specific volume: 50,000 MI450 GPUs in Helios-architecture superclusters, with Q3 2026 deployment start and further expansion in 2027.
Through our decade-long collaboration with AMD, we're continuing to deliver the best price-performance
-
Tom's Hardware Read →
Supply-chain credibility context: AMD pushed back on delay reports while Nvidia VR200 may ship ahead of schedule, raising the stakes of H2 2026 customer delivery decisions.
Helios systems on target for 2H 2026
-
Windows News Read →
Flags that ND MI455X v7 is inference-targeted specifically; advises Azure customers to validate ROCm compatibility on existing MI300X instances now ahead of Helios availability.
ND MI455X v7 is optimized for inference, placing it right where the industry is shifting: away from one-off training jobs toward persistent, always-on AI services
-
Shacknews Read →
Frames Microsoft as the capstone validation of Helios alongside Meta, OpenAI, and Oracle; situates the deal within AMD's broader stock momentum leading into the launch.
Microsoft's new AMD deployments mark an important milestone as we deliver leadership compute solutions to Azure customers
-
Crypto Briefing Read →
Emphasizes AMD's open-ecosystem and rack-level integration strategy as the competitive lever against Nvidia, rather than individual GPU benchmark comparisons.
A single Helios rack can push up to 1.4 exaFLOPS of FP8 compute and 2.9 exaFLOPS of FP4 compute.
-
Stocktwits Read →
Carries Lisa Su's direct CEO quote from the announcement event; frames the 5% AMD stock gain as market pricing of Nvidia competition risk becoming real.
AMD and Microsoft have spent years building high-performance infrastructure together, and today we're extending that partnership — Lisa Su, AMD CEO
Originally reported by newsroom.amd.com
Read the original article →Original headline: Microsoft Names AMD Helios First-Customer Deal, Deploying 72-GPU MI455X Racks at Scale on Azure in H2 2026