AMD has officially launched Helios, its first fully integrated rack-scale AI system designed to challenge NVIDIA's dominance of the data center GPU market. Announced on July 20, 2026, the platform combines 72 Instinct MI455X GPUs with 6th-generation EPYC "Venice" processors and 31 terabytes of HBM4 memory per rack — and has already secured deployment commitments from Microsoft Azure, Meta, OpenAI, and Oracle.
Key Highlights
- 72 Instinct MI455X GPUs (CDNA 5 architecture) per rack
- 31TB HBM4 memory with 260TB/s rack-scale bandwidth
- 1.4 exaFLOPS FP8 compute and 2.9 exaFLOPS FP4 compute
- 6th-generation EPYC "Venice" CPUs for system management
- Microsoft Azure deploying ND MI455X v7 VM instances for inference
- Meta targeting 1 gigawatt of Helios capacity in H2 2026
- Oracle rolling out 50,000 MI455X GPUs from Q3 2026
- Shipments begin H2 2026; mass production expected Q2 2027
A Full-Stack Challenger to NVIDIA
Unlike previous AMD GPU launches that competed chip-by-chip, Helios takes a system-level approach — packaging GPUs, CPUs, and networking into a single integrated rack solution. The platform uses UALink over Ethernet for intra-rack communication and Pensando Ethernet for inter-rack connectivity, delivering 260TB/s of aggregate bandwidth at rack scale and 43TB/s between racks.
AMD positions Helios as an open-architecture alternative to NVIDIA's proprietary Grace Blackwell and Vera Rubin systems. Hardware manufacturing is handled by Celestica and Super Micro, giving cloud providers and hyperscalers supply chain flexibility that does not exist within NVIDIA's closed ecosystem.
Microsoft Azure Goes All-In
Microsoft will deploy Helios racks inside Azure data centers and launch new ND MI455X v7 VM instances targeting large-scale inference, search, reasoning, and agentic workloads. Two additional VM families built on EPYC Venice are also coming: HDv2 for AI data pipelines (approximately 500 cores, 4TB RAM) and HXv2 for chip design simulation (176 cores, over 5GHz).
Microsoft CEO Satya Nadella framed the deployment as part of Azure's push to give customers "the performance, expansion space, and diversified hardware options required to build and run next-generation AI applications." The companies share a longstanding partnership: AMD chips power Microsoft Surface PCs and Xbox consoles, and Azure has deployed AMD hardware since the MI300X generation.
Meta, Oracle, and OpenAI Join the Customer List
Meta is deploying a custom MI450-based variant of Helios at scale, with a first-phase target of 1 gigawatt of AI compute capacity launching in H2 2026. Oracle Cloud has committed to rolling out 50,000 AMD GPUs starting Q3 2026. OpenAI has also finalized plans for large-scale Helios deployments, adding the leading AI lab to AMD's growing enterprise roster.
Market Context: From 4.5% to a Credible Rival
NVIDIA currently holds more than 95% of the data center GPU market, while AMD occupies approximately 4.5%. Helios represents AMD's most deliberate attempt yet to close that gap. AMD's data center segment already posted 57% year-over-year revenue growth in Q1 2026, and eight of the world's ten largest AI companies reportedly use AMD Instinct GPUs in some capacity.
Industry analysts project that AMD's rack-scale push could help grow its data center GPU market share to between 20% and 25% over the coming years — a shift that would represent hundreds of billions of dollars in additional revenue.
What to Watch Next
AMD is hosting its "Advancing AI 2026" keynote on July 23, 2026, where additional technical details, performance benchmarks, and potentially new customer announcements are expected. Azure pricing, regional availability, and specific VM SKUs for the ND MI455X v7 instances have not yet been disclosed.
Source: Cryptobriefing