AMD has officially launched its Helios rack-scale AI system at the Advancing AI 2026 conference in San Francisco. The system combines 72 AMD Instinct MI455X GPUs with 18 sixth-generation EPYC "Venice" CPUs in a single integrated rack. They are interconnected through AMD Pensando networking and accelerated by the open ROCm software stack.

Related: Earlier coverage: AMD Stock Surges After Microsoft Commits to Azure Deployment of Helios Rackscale.

CEO Dr. Lisa Su positioned Helios as the industry's highest-performance AI rack. She claimed up to 30% more inference tokens per dollar than Nvidia's Vera Rubin platform. AMD also states Helios delivers 15% better compute performance and 50% more high-bandwidth memory capacity than the competing Nvidia system. The company expects its total addressable market across rack systems, PC business, and other processing products to reach $2 trillion by 2030.

What's New

  • System architecture: 72 Instinct MI455X GPUs and 18 EPYC Venice CPUs per rack in a co-optimized design
  • Performance claims: 30% more tokens per dollar, 15% better compute performance, 50% more HBM capacity versus Nvidia Vera Rubin
  • Software stack: Open ROCm platform with support for frontier models including DeepSeek-V4-Flash
  • Networking: AMD Pensando Salina and Pollara 400 for scale-up and scale-out fabrics
  • Configurations: Four customizable rack configurations for different workload profiles

Microsoft announced it will deploy Helios racks in Azure data centers. Meta, OpenAI, Oracle, and Anthropic are also early customers. Meta plans to deploy up to 6 gigawatts of AMD GPUs over time, starting with 1 gigawatt on Helios racks. Anthropic and AMD announced a strategic partnership to deploy up to two gigawatts of GPUs via the new system. Tata Consultancy Services also committed to using Helios.

AMD will begin shipping Helios to customers later in 2026. An earlier version of this article stated the system was already in production; AMD has indicated shipping begins in the second half of the year. The Futurum Group estimates Helios will cost between $5 million and $5.5 million per rack, compared with $3.5 million to $4 million for Nvidia's Vera Rubin system. At up to 7,000 pounds, Helios is also wider and heavier than its Nvidia counterpart.

Why It Matters

Nvidia controls more than 95% of the data center GPU market according to Futurum estimates, while AMD holds roughly 4.5%. Helios represents AMD's first true rack-scale competitor to Nvidia's Grace Blackwell and Vera Rubin systems. Its success could shift market dynamics in a segment projected to reach $1.4 trillion by 2030. AMD's open, full-stack approach — combining in-house GPUs, CPUs, networking, and software — differentiates it from Nvidia's more proprietary ecosystem.

Microsoft's adoption is significant given its existing Maia custom silicon and long-standing partnership with AMD. The company will also add two new Azure compute instances running on EPYC Venice CPUs for agentic AI workloads and semiconductor design. AMD plans to book tens of billions in data center AI revenue starting in 2027, with the majority expected from Helios deployments.

Our Take

Helios gives AMD a credible rack-scale product for the first time. The real test is execution: shipping on schedule, scaling manufacturing, and proving the 30% tokens-per-dollar advantage holds across diverse customer workloads — not just cherry-picked benchmarks. Nvidia's CUDA moat remains deep, and AMD's ROCm ecosystem still has ground to cover for broad ISV adoption.

Sources