AMD used its sold-out Advancing AI conference in San Francisco on Thursday to formally position Helios, its new rack-scale AI system, as a direct competitor to Nvidia's Vera Rubin and Grace Blackwell platforms. Chair and CEO Lisa Su said Helios will ship later this year and named Microsoft, OpenAI, Meta, Oracle, and Anthropic as launch customers, each planning gigawatt-scale deployments. Su called it the tech industry's "highest-performance AI rack," a claim The Register's benchmarking backs on several metrics against Vera Rubin.
Rack-scale systems bundle hundreds of GPUs, CPUs, networking, and cooling into a single dense unit tuned for training and serving frontier models. AMD first revealed Helios in 2025 and put it onstage at CES 2026 in January. The commercial ramp starts now.
“built to train and run the most demanding frontier models in the world at massive scale”— Lisa Su, AMD Chair and CEO
The customer roster is the substance of the announcement. Microsoft CEO Satya Nadella said Monday that Azure will expand its AI infrastructure with Helios. Two days later, Anthropic and AMD confirmed a strategic partnership to deploy up to 2 gigawatts of GPUs on the platform — a scale that puts AMD in the same conversation as Nvidia for the largest AI training clusters being built. OpenAI, Meta, and Oracle round out the initial commitments.
Key facts
- 01AMD unveiled Helios at its Advancing AI conference in San Francisco on Thursday, with shipments starting later this year.
- 02Microsoft, OpenAI, Meta, Oracle, and Anthropic have all signed on to deploy Helios at gigawatt scale.
- 03Anthropic and AMD announced a strategic partnership Wednesday to deploy up to 2 gigawatts of GPUs on the system.
- 04Lisa Su projected the AI accelerator market will reach $1.4 trillion by 2030, approaching the size of today's entire semiconductor market.
- 05AMD's Venice-X CPU, also announced Thursday, is targeted for a 2027 launch.
AMD also introduced Venice-X, a data-center CPU designed to pair with Helios and other high-compute workloads. Venice-X is scheduled to launch in 2027, giving AMD a longer roadmap that stretches past this generation of accelerators. The company is trying to sell customers on a full-stack alternative — CPU, GPU, rack — rather than picking off individual sockets.
Su framed the demand story around agents. Reasoning models and tool-using agents burn far more compute per query than a single-shot chatbot because they iterate. "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.
That framing led to AMD's most aggressive forecast of the day. Su told the audience the AI accelerator market will reach roughly $1.4 trillion by 2030, adding that by the end of the decade it will "approach the size of the entire semiconductor market today." She argued GPUs will hold the vast majority of that share because AI algorithms and workloads keep changing, which favors the programmability of general-purpose accelerators over more rigid custom silicon.
The competitive read is straightforward. Nvidia has effectively been the sole vendor selling rack-scale AI systems at frontier-lab scale, with Grace Blackwell and now Vera Rubin. AMD's MI300 and MI325 accelerators sold well but did not force a second-source conversation at the largest hyperscalers. Helios, backed by named commitments from Microsoft, OpenAI, Meta, Oracle, and Anthropic, is the first time AMD has landed all five of those customers on a single rack platform at the same time.
The Anthropic deal is the standout. A 2-gigawatt GPU commitment is one of the largest single-customer compute deals disclosed to date, and it hands AMD credibility with the frontier-lab buyer segment that has been the hardest to crack. It also follows Anthropic's broader push to diversify supply beyond a single vendor, which AI Chat Daily has tracked across its Google TPU and Amazon Trainium arrangements over the past year.
The caveats are real. Shipping a rack at scale is a different problem than winning the design. Nvidia's lead in software — CUDA, NCCL, the TensorRT stack — remains the reason most training clusters default to its hardware, and AMD's ROCm ecosystem, while much improved, still lags in developer familiarity. Su did not provide unit-level shipment guidance or a revenue target for Helios in 2026, and the gigawatt figures cited by customers are deployment intentions rather than delivered capacity. The Register's performance comparisons are favorable but not independently benchmarked at scale.
For AMD, the significance of Thursday is that the story has moved from "can they compete" to "they have the customers." If Helios ships on schedule and the Anthropic and Microsoft deployments hit their gigawatt targets, AMD's data-center segment gets a step-change in revenue mix that Wall Street has been waiting on for two years. If Su's $1.4 trillion 2030 figure is even directionally right, a credible second supplier is worth tens of billions annually — and the hyperscalers have every incentive to make sure one exists.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
