AMD's Helios Takes On Nvidia: Microsoft Joins the Queue for a New AI Supercomputer

Key points

• AMD's Helios, a rack-scale AI computing system, will begin shipping to customers including Microsoft later in 2026.
• Microsoft announced on Monday 30 June 2026 it will deploy Helios in its Azure cloud data centres.
• Meta committed in February 2026 to deploying 1 gigawatt of Helios capacity, with plans for up to 6 gigawatts over time.
• Analyst firm Futurum Group estimates a single Helios system costs between $5 million and $5.5 million, compared to $3.5 million to $4 million for Nvidia's rival Vera Rubin system.
• Nvidia currently controls more than 95 percent of the data centre GPU market, according to Futurum Group; AMD holds roughly 4.5 percent.

Picture a single machine the size of a small car, weighing up to 7,000 pounds, packed with chips that let AI companies process millions of requests at once. That is AMD's Helios, and it is about to land in some of the world's biggest data centres.

Microsoft announced this week it will add Helios to its Azure cloud infrastructure, the vast network of computers that powers everything from Office apps to the AI assistant Copilot. Microsoft joins Meta, OpenAI, Oracle and India's Tata Consultancy Services in placing early orders. AMD will begin shipping the systems later this year.

CNBC, which first reported the story, got an exclusive look inside a working Helios prototype at AMD's data centre lab in Rockdale, Texas.

What does this mean for people who use AI services?

For most people, the immediate effect is indirect: more Helios systems in data centres means AI companies can handle more requests, faster and at lower cost per query. That can translate into cheaper, quicker AI tools over time, though no price cuts for consumers have been announced.

Helios packs four Instinct GPUs (graphics processing units, the specialised chips that do the heavy number-crunching AI workloads demand) and one EPYC CPU (the central processor that co-ordinates the work) onto each of 18 computing trays per rack. AMD built the GPUs, the CPUs, the networking chips and the software itself, keeping the whole stack in-house.

AMD data centre chief Forrest Norrod told CNBC the company's goal is "the lowest cost per token," meaning lower cost per individual piece of AI output, like a word generated by a chatbot. AMD CEO Lisa Su said in May that Helios offers "significant benefits" over Nvidia's systems for inference (running a finished AI model to answer questions) and for memory handling.

Nvidia still towers over the market. That 95 percent share is the product of years of dominance, and its CUDA software ecosystem, a set of tools developers use to write code for Nvidia chips, is deeply embedded across the AI industry. AMD has its own open-source alternative called ROCm, but Counterpoint Research analyst Neil Shah says Nvidia's ecosystem is "quite ahead."

Analyst Daniel Newman of Futurum Group sees a realistic path for AMD to reach 20 to 25 percent market share. In an industry worth hundreds of billions of dollars, even that slice would be transformative for AMD.

AMD's data centre revenue grew 57 percent year-over-year in the first quarter of 2026. The company says it expects tens of billions in data centre AI revenue from 2027, with Helios as the centrepiece. Whether that holds depends on how the first real-world deployments perform.