AMD Moves Helios AI Racks Into Production Ahead of 2026 Shipments

AMD has moved its Helios rack-scale artificial intelligence platform into production as it prepares to begin customer shipments in the second half of 2026.

The system is built around 72 Instinct MI455X accelerators, each equipped with 432 gigabytes of HBM4 memory. A complete rack provides 31 terabytes of HBM4 memory and up to 2.9 exaFLOPS of FP4 performance.

The company is targeting large-scale AI training and inference deployments. Microsoft plans to use Helios on Azure for frontier-model inference and Azure AI services, while OpenAI expects to begin bringing the platform online in the fourth quarter. Anthropic has announced plans with AMD covering up to 2 gigawatts of MI455X deployments in Helios systems.

The production ramp places added importance on access to HBM4 and advanced packaging, the process used to combine memory stacks with processors. AMD’s 2025 annual report identified supply-chain coordination as part of the company’s effort to ramp the MI400 series and Helios.

Samsung is AMD’s primary HBM4 partner under a memorandum of understanding signed in March. The agreement covers HBM4 for MI455X accelerators and advanced DRAM for AMD’s next-generation EPYC processors.

AMD is using Helios to compete with Nvidia at the rack level, where customers buy integrated systems rather than individual accelerators. The platform uses open standards, including UALink and Ethernet, and can scale from a single rack to larger AI clusters.

The immediate test will be whether AMD can convert announced deployments into volume shipments while securing enough HBM4 and advanced-packaging capacity to meet demand.

Customer shipments are expected to begin in the second half of 2026.