AMD unveils Helios AI rack to challenge Nvidia

AMD Advanced Micro Devices, Inc. ↘ 5.17% unveiled its Helios rack-scale AI system and Instinct MI455X GPUs at its Advancing AI 2026 keynote on Thursday in San Francisco, marking the chipmaker's most direct challenge yet to Nvidia's ↘ 4.99% dominance in AI data center infrastructure. CEO Lisa Su delivered the keynote at the Moscone Center, presenting what the company calls the world's best AI rack.Advanced Micro Devices, Inc.+1

Inside the Helios Rack

Each Helios system packs 72 MI455X GPUs across 18 compute trays, with each tray housing four accelerators and one EPYC Venice CPU. The full rack delivers 2.9 exaflops of FP4 inference performance and 1.4 exaflops of FP8 training capacity, drawing on 31 terabytes of pooled HBM4 memory. Each MI455X chip carries 432GB of HBM4 memory and 320 billion transistors, built on TSMC's Taiwan Semiconductor Manufacturing Company Limited ↘ 1.07% 2nm and 3nm processes.ServeTheHome+3

AMD also formally launched EPYC Venice, the first x86 server CPU manufactured on TSMC's 2nm node. The flagship configuration scales to 256 Zen 6 cores and 512 threads, a 33% increase over the previous-generation Turin processors, with AMD claiming over 70% higher performance and efficiency versus Zen 5. Pensando networking hardware provides scale-out connectivity within each rack.Advanced Micro Devices, Inc.+2

Customers and Competition

Microsoft ↗ 1.94%, Meta ↘ 0.22%, OpenAI, and Oracle ↗ 4.27% are confirmed as anchor customers, with deployments beginning in the second half of 2026. According to CNBC, Microsoft will deploy Helios across its Azure data centers for frontier AI model inference and enterprise applications.youtube+3

The pricing gap with Nvidia remains a challenge. Futurum Group estimates Helios at $5 million to $5.5 million per rack, compared with $3.5 million to $4 million for Nvidia's next-generation Vera Rubin platform. AMD holds roughly 4.5% of the data center GPU market versus Nvidia's 95%, according to CNBC's reporting.TrendForce+2

The Broader AI Chip Battle

AMD's bet centers on memory advantages — each Helios rack offers roughly 150GB more HBM4 per GPU than Nvidia's Vera Rubin is expected to carry, a potential edge for memory-intensive inference workloads like large language models. The system is built on the Open Rack Wide v3 specification developed with Meta and is designed for OCP-compliant integrators.WEBRONAQ+2

A YouTube video published on July 22 also indicated that Anthropic has committed to up to 2 gigawatts of AMD GPU deployments, with AMD investing up to $5 billion in the AI company — though AMD has not confirmed this in official press materials.youtube