AMD is preparing to ship Helios, its first rack-scale system tailored for large AI workloads, to a growing roster of customers that now includes Microsoft. The system brings together AMD’s GPUs, EPYC CPUs, networking and software in an effort to offer an alternative to Nvidia’s dominant rack products and to reduce the total cost of running inference and large-scale AI services.
Customers and commercial rollout
Microsoft confirmed it will deploy Helios in its data centers and add new Azure instances powered by AMD’s latest EPYC “Venice” CPUs for agentic AI, data pipelines and semiconductor design workloads. Microsoft joins Meta, OpenAI, Oracle and Tata Consultancy Services among the announced Helios users; other firms running AMD Instinct GPUs include OpenAI, Cohere and SpaceXAI.
AMD said it will begin shipping Helios to customers later this year. Financial terms and exact capacity commitments were not disclosed. Analysts and industry observers are watching early deployments closely, since large cloud and enterprise buys can influence long-term data center demand.
System design and capabilities
Named for the Greek sun deity, Helios is built from technologies AMD develops in-house: Instinct GPUs, EPYC server CPUs, networking silicon (including chips from AMD’s Pensando acquisition) and ROCm-based software. The rack design features multiple compute trays—each tray pairs an EPYC CPU with four Instinct GPUs—and includes dense networking components to link compute resources at rack scale.
AMD emphasizes total cost of ownership and “cost per token” for inference workloads. Company executives have highlighted Helios’ memory bandwidth and other architectural attributes when contrasting it with Nvidia’s rack offerings, positioning Helios for model inference and memory-intensive AI tasks.
Market context and performance claims
Nvidia currently controls a large majority of the data center GPU market, and industry estimates put that share above 90%. AMD’s share of data center GPUs is much smaller—around 4.5% according to cited research—but the vendor has gained traction with recent EPYC CPU wins and the rollout of Instinct accelerators.
Cost estimates from industry analysts vary: one research firm projects Helios hardware could be in the $5 million to $5.5 million range per rack, compared with estimated price points of roughly $3.5 million to $4 million for Nvidia’s second-generation Vera Rubin rack. Helios is also physically larger and heavier in some configurations, with weights quoted up to 7,000 pounds.
Strategic momentum and acquisitions
Helios is the product of a multi-year strategy that married AMD’s CPU resurgence with targeted acquisitions and software work. Following the return of competitive EPYC server processors, AMD bought Xilinx, Pensando and other firms to extend capabilities across programmable logic, networking and systems. These acquisitions helped assemble the components and software stack used in Helios.
ROC m, AMD’s open-source compute stack, figures into the company’s push to provide a fuller alternative to Nvidia’s CUDA ecosystem. Analysts say software and optimization will be a key differentiator for Helios; AMD’s hardware parity with Nvidia is notable, but the breadth and maturity of the software ecosystem remain important for large-scale deployments.
Financial impact and future outlook
Data centers accounted for the majority of AMD’s revenue in the first quarter of 2026, rising 57% year over year. AMD has signaled expectations for substantial AI-driven data center revenue, forecasting tens of billions in such revenue beginning in 2027, with Helios playing a primary role.
Analysts suggest a successful Helios rollout could materially change AMD’s position in the GPU market, with potential upside to market share and revenue if large cloud providers and enterprises continue to adopt the system. However, early performance and software support will determine how quickly customers shift workloads from established Nvidia platforms.
What to watch
Key indicators to monitor include the timing and scale of Helios deployments at Microsoft, Meta and OpenAI; customer reports on performance, cost per inference token and operational efficiency; and how quickly AMD’s ROCm ecosystem attracts third-party tools and optimizations. The interplay between AMD’s EPYC CPUs and Instinct GPUs in Helios racks will also be scrutinized as organizations evaluate trade-offs between cost, performance and software compatibility.
As the AI hardware landscape evolves, Helios represents AMD’s most significant bid yet to compete directly with Nvidia at the rack level. The outcome will depend on execution across manufacturing, customer deployments and software support in the months ahead.
Source: CNBC