AMD's Helios rack AI system debuts as Nvidia challenger with Microsoft as a customer

AMD is shipping its first rack-scale AI system, Helios, to a growing list of customers that now includes Microsoft. It is the first direct rival to Nvidia's Grace Blackwell and Vera Rubin systems for frontier model inference and data-center AI workloads.

What happened

AMD announced that Helios will begin shipping later this year to customers including Microsoft, Meta, OpenAI, Oracle, and Tata Consultancy Services. Each Helios rack contains 18 compute trays, with each tray housing up to four Instinct GPUs powered by a single EPYC data-center CPU. The system also includes Pensando networking chips, acquired by AMD in 2022, for integrated high-performance networking.

Microsoft CEO Satya Nadella said Azure will expand with two new Venice CPU-based instances for agentic AI and data pipelines, and that Helios will power Azure AI services. Microsoft plans to deploy Helios alongside its in-house Maia chips and existing MI300X deployments, continuing a partnership that began with the MI300X in 2023.

According to the Futurum Group, Nvidia controls over 95% of the data-center GPU market, while AMD holds roughly 4.5%. Futurum analyst Daniel Newman suggests Helios could help AMD capture 20-25% of the market, representing hundreds of billions in potential revenue starting in 2027.

Why AI builders should care

AMD is positioning Helios around total cost of ownership and lowest cost per token. For teams deploying frontier models at scale, any shift in GPU economics directly affects inference budgets and infrastructure choices.

The key variable is software. AMD's ROCm stack is the open-source alternative to Nvidia's CUDA ecosystem. Counterpoint Research analyst Neil Shah told CNBC that Helios chips are "on par" with Nvidia GPUs, but the "secret sauce is in the software and optimization." Builders evaluating Helios should assess ROCm compatibility with their existing model pipelines and tooling.

Practical implications

Helios brings together four in-house AMD components: GPUs, CPUs, networking, and software. Each of the 18 trays integrates four Instinct GPUs with one EPYC CPU, plus up to 12 Pensando networking chips per tray. The dense rack-scale design targets inference for frontier models rather than training, where AMD sees memory bandwidth and memory capabilities as advantages.

Microsoft's adoption adds a major cloud distribution channel. Developers using Azure AI services should expect Helios-backed instances in the future, alongside existing Maia and MI300X options. AMD says eight of the top 10 AI companies already run workloads on its Instinct GPUs, including OpenAI, Cohere, and SpaceXAI.

Caveats

Pricing and exact compute capacity were not disclosed. The Futurum Group estimates Helios cost between $5 million and $5.5 million, compared with $3.5 million to $4 million for Nvidia Vera Rubin. However, these are analyst estimates, not official figures.

AMD's claims about total cost of ownership depend on real-world deployment results and software parity with CUDA. Nvidia's ecosystem remains dominant in the near term, and Helios success will hinge on whether early deployments deliver the promised cost-per-token advantages.