AMD has unveiled Helios, a rackscale AI system built around its latest Instinct MI450 accelerators, and Microsoft is already on board as a launch customer. The company says Azure will ramp Helios at scale, with shipments to Microsoft and other customers beginning in the second half of 2026.
What Helios packs
A single Helios rack integrates 72 GPUs and up to 31 TB of HBM4 memory, delivering peak throughput of up to 2.9 exaflops in FP4 precision. AMD is positioning the system for both training and inference of large models, and claims roughly 30% more tokens per dollar versus the leading alternative — a performance-per-dollar pitch that has become central to its AI strategy.
The Microsoft deal
AMD and Microsoft jointly announced that Azure will deploy Helios at scale. The partnership gives AMD a named customer with a signed purchase order and a shipment schedule, which management has flagged as important for backlog quality. The H2 2026 timeline puts the system roughly two years out from today's announcement.
Margins hinge on HBM and packaging
AMD's AI gross margins will depend heavily on HBM cost, memory yields, advanced packaging, and the mix of racks sold. The company hasn't given specific margin targets for Helios, but the economics of high-bandwidth memory and chiplet integration remain a key variable in the profitability of its data center GPU business.
Software and ecosystem readiness
ROCm maturity, broad framework support, and reference deployments are meant to reduce friction for customers adopting AMD's hardware. The company has been working to close the software gap with Nvidia's CUDA ecosystem, and Helios will rely on that progress to win workloads in production environments.
With a major cloud customer locked in and a clear performance-per-dollar message, AMD now faces the challenge of delivering on the H2 2026 timeline while managing the cost structure that will determine whether Helios is a profitable product.
