AMD Launches Helios Rack-Scale System, Takes Aim at Nvidia in AI Infrastructure

AMD today announced the full production launch of its Helios rack-scale AI system, with major customers including Microsoft and Anthropic already committed to deployment.

This article was drafted with AI assistance from multiple sources and was reviewed and approved by a human editor before publication.

AMD today announced the full production launch of its Helios rack-scale system at the AMD Advancing AI 2026 event in San Francisco, directly challenging Nvidia's dominance in AI infrastructure. The system, which was first revealed in 2025 and shown at CES 2026 in January, is now in full production as of today.

Helios measures 1.2 meters wide and 1.3 meters deep, based on the Open Rack Wide (ORW) format with 44 Open Units (OU). One Open Unit equals 48 mm in height, compared to the standard 44.45 mm. The rack contains 18 compute trays, each housing four Instinct accelerators and one Epyc CPU, for a total of 72 GPUs and 18 CPUs. Six switch trays sit in the middle, each with two Ethernet switches based on Broadcom Tomahawk technology, for a total of 12 switches connected to all 72 GPUs.

The GPU used is the Instinct MI455X, paired with an AMD Epyc 9G76 CPU featuring 96 cores (8×12 Zen-6) running at 5 GHz. The system includes 1 TB of RAM via 16 DIMM slots, each 64 GB. Helios also features the first 800 Gbit Ethernet solutions, named Vulcano, fabricated in 3 nm, and includes Salina DPUs from Pensando, which AMD acquired several years ago.

AMD acquired ZT Systems for $4.9 billion, contributing rack expertise and partnerships to the Helios development. The Switch Blade component weighs 170 lb (77 kg).

Major customers are already committed to deploying Helios. Microsoft CEO Satya Nadella said on July 20 that Azure infrastructure will expand with Helios. Anthropic and AMD announced a strategic partnership on July 22 to deploy up to 2 GW of GPUs via Helios. Other customers include OpenAI, Meta, and Oracle. Helios is expected to ship later in 2026.

AMD also introduced the Venice-X CPU, designed for data centers, with an expected launch in 2027. AMD CEO Lisa Su said the AI accelerator market will reach approximately $1.4 trillion by 2030, approaching the size of the entire semiconductor market today. She added that GPUs will make up the vast majority of the AI accelerator market due to workload changes favoring programmability.