AMD Launches Helios Rack-Scale AI System to Challenge Nvidia's Data Center Dominance

Advanced Micro Devices officially launched Helios, its first fully integrated rack-scale AI system, at its Advancing AI 2026 conference. The system connects 72 Instinct MI455X GPUs with 18 6th Gen EPYC “Venice” CPUs, AMD Pensando networking and ROCm software. AMD says Helios provides up to 2.9 exaflops of peak theoretical FP4 compute and claims up to 15% more peak FP4 compute than Nvidia’s Vera Rubin NVL72, based on AMD calculations. AMD also estimates that Helios could deliver up to 30% more tokens per dollar for a modeled Kimi K2 Thinking workload using projected pricing. Hardware partners are expected to begin shipping Helios by the end of the third quarter of 2026. Microsoft plans to deploy Helios on Azure, while Anthropic plans to deploy up to two gigawatts of AMD Instinct GPUs in Helios systems, with the first gigawatt scheduled to begin deployment in the first half of 2027.

AMD Helios rack system launch

• ▪Advanced Micro Devices officially launched Helios, its first fully integrated rack-scale AI system, at its Advancing AI 2026 conference in San Francisco on July 23, 2026.
• ▪Anthropic partnered with AMD to deploy up to 2 gigawatts of Instinct GPUs in Helios systems, with the first gigawatt of deployment starting in the first half of 2027.
• ▪AMD expects to begin shipping the Helios rack-scale system to customers in the second half of 2026.
• ▪Microsoft expanded its Azure partnership with AMD to include planned Helios deployments beginning in the second half of 2026.
• ▪The AMD Helios rack-scale system combines 72 Instinct MI455X GPUs, 18 next-generation EPYC Venice server processors, and AMD Pensando networking.

AI infrastructure demand growth

• ▪AMD said monthly inference-token volume increased 158-fold over a two-year period as artificial intelligence moved from experimentation into products and services.
• ▪The rise of agentic AI is driving a step change in compute demand because a single request triggers multiple reasoning steps, sub-agents, and tool calls.
• ▪AMD projects the AI accelerator market will reach approximately $1.4 trillion by the year 2030.

MI455X GPU specifications

• ▪AMD Vice President of AI Software Anush Elangovan said the Instinct MI455X achieved 20 petaflops of FP4 performance in AMD’s real-world testing, approximately half of its peak theoretical performance.
• ▪According to AMD, the Instinct MI455X GPU carries 432 GB of HBM4 memory and provides up to 23.3 TB/s of memory bandwidth.
• ▪The AMD Instinct MI455X GPU features 24 chiplets, including eight compute dies fabricated on TSMC's 2-nanometer process technology.

Helios versus Nvidia Vera Rubin

• ▪According to AMD, the Helios rack provides up to 2.9 exaflops of peak theoretical dense FP4 compute, 31 TB of HBM4 memory and 1.7 PB/s of aggregate HBM bandwidth.
• ▪Based on AMD calculations and published specifications, AMD claims Helios provides up to 15% more peak theoretical FP4 compute, 50% more HBM capacity and 50% more scale-out bandwidth than Nvidia's Vera Rubin NVL72 rack.
• ▪AMD estimates that Helios could deliver up to 30% more tokens per dollar than Nvidia's Vera Rubin NVL72 for a modeled Kimi K2 Thinking workload, using projected system pricing.

AMD ROCm software ecosystem

• ▪AMD and Anthropic are launching a multi-year engineering collaboration to use Claude AI to optimize workloads for AMD Instinct GPUs and accelerate ROCm development.
• ▪AMD said ROCm.ai combines the ROCm command-line interface, AMD Skills and AI-powered software optimization into a unified developer experience across AMD platforms.