Challenging Nvidia’s dominance! AMD launches the “Helios” rack-mounted AI system, and Microsoft Azure announces it will be the first to roll it out

According to CNBC, chipmaker AMD has officially unveiled its first rack-scale AI system, “Helios,” directly targeting NVIDIA’s market monopoly position. Technology giant Microsoft immediately announced that it will roll out the system in Azure data centers, providing stronger compute infrastructure for future frontier model inference and AI services.
(Background: A financial black hole emerges for OpenAI: a single-year loss of $38.5 billion; analysts warn that a domino collapse is about to begin)
(Additional background: 142 anti-AI protests across the U.S. are intensifying, opposing disorderly expansion of AI data centers)

Table of contents

Toggle

• Microsoft Azure to be deployed first, expanding AI infrastructure
• Named after ancient Greek mythology, full hardware specifications revealed
• Aims to win 25% market share; Meta has begun testing in Texas

As the AI arms race continues to heat up, competition among chip makers has entered a new phase. On July 20, 2026 (Taipei time), according to an exclusive report by CNBC, after a decade-long transformation and revival, AMD is preparing to begin shipping its first rack-scale AI system named “Helios” in the second half of 2026. The product is seen as AMD’s most direct and powerful weapon so far in its fight against NVIDIA’s Grace Blackwell and Vera Rubin systems.

Microsoft Azure to be deployed first, expanding AI infrastructure

As AMD’s long-term partner, Microsoft CEO Satya Nadella announced on Monday that it will expand Azure’s infrastructure lineup with AMD Helios, enabling customers to build and run next-generation AI applications. Nadella said Helios will be mainly used for frontier model inference, supporting Azure AI services, and it will add two compute instances based on AMD’s latest “Venice” CPUs, respectively for agentic AI, data pipelines, and areas such as semiconductor design.

This marks a further deepening of Microsoft’s alliance with AMD after 2023, when it became among the first customers to adopt AMD MI300X GPUs. Although Microsoft is also actively developing its own chip called Maia, amid internal demand for as many as 7 AI models, Microsoft still has to actively seek additional external compute resources.

Named after ancient Greek mythology, full hardware specifications revealed

It is said that “Helios” takes its naming inspiration from the ancient Greek deity who harnessed four horses to pull the sun, symbolizing AMD’s deep integration of its four core technologies: GPU, CPU, networking, and software. In terms of hardware specifications, each Helios rack is equipped with 18 compute trays (tray). Each tray includes 4 Instinct GPUs, 1 EPYC CPU, and up to 12 networking chips from the company Pensando, acquired in 2022.

Based on estimates from the research firm Futurum Group, this massive system weighing 7,000 pounds and with a large footprint is priced at about $5 million to $5.5 million. By comparison, NVIDIA’s Vera Rubin system is about $3.5 million to $4 million. However, AMD CEO Lisa Su previously emphasized that Helios is designed to deliver the best total cost of ownership (TCO) and the lowest cost per token, with notable advantages particularly in model inference, memory bandwidth, and capacity.

Aims to win 25% market share; Meta has begun testing in Texas

At present, NVIDIA holds more than 95% of the overwhelming share in the data center GPU market, while AMD has about 4.5%. However, market analysts are optimistic that the launch of Helios could substantially raise AMD’s market share to 20% to 25%. AMD officials also said that among the world’s top 10 AI companies, 8 are already using its Instinct GPU series products. In addition to Microsoft, they include OpenAI, Oracle, Tata Consultancy Services, and major player xAI, led by Elon Musk.

Notably, Meta, one of the early customers, plans to use up to 6 GW of AMD GPUs, and this year it already deployed 1 GW of compute resources on Helios. In on-site checks by CNBC reporters, the customized version Meta plans to deploy is scheduled to be installed and tested at Rockdale Labs in Texas as early as June 2026. As hardware specifications gradually catch up with competitors, whether AMD can break through successfully by leveraging the open-source ROCm software ecosystem will be the next key focus of this AI chip battle.