Microsoft announced on Monday that it will deploy AMD's Helios rack-scale equipment in its Azure data centers, joining Meta, OpenAI, and Oracle as early customers for this new system.
Helios represents AMD's first internally developed, rack-scale AI integrated solution and is the company's most significant product to date in a direct challenge to NVIDIA. The equipment is scheduled for shipment and delivery later this year.
AMD has launched a competing AI rack system to directly confront NVIDIA, securing major customer orders from both Microsoft and Meta.
After a decade of strategic recovery, chipmaker Advanced Micro Devices is preparing to deliver its first rack-scale artificial intelligence integrated system, named Helios. The list of contracted clients continues to grow, with Microsoft officially joining the procurement roster.
This product is the first serious competitor to NVIDIA's highly successful Grace Blackwell and Vera Rubin rack systems, marking the first time in years that the world's highest-valued chip company is facing substantial industry competition.
Microsoft disclosed its procurement plans on Monday, stating it will implement Helios hardware in its own data centers. As technology firms scramble for computing resources, Meta, OpenAI, and Oracle have also finalized procurement agreements. AMD will begin bulk shipments to Microsoft and other customers later this year. The specific financial terms and scale of the computing power purchased were not disclosed.
Microsoft CEO Satya Nadella stated in the official announcement: "We are incorporating AMD Helios to expand Azure's computing infrastructure, providing the performance, scalability, and hardware diversity our customers need to build and run the next generation of AI applications."
This new Helios system will support Microsoft's and its cloud customers' cutting-edge large language model inference tasks, while also powering the full suite of Azure cloud AI services. Microsoft will also launch two new cloud compute instances based on AMD's latest Venice CPU, one targeting agent AI and data pipeline workloads, and the other serving chip design and simulation scenarios.
The collaboration between the two companies has a long history: AMD chips have been featured in Microsoft's Surface PCs and Xbox gaming consoles for years. In 2023, Microsoft was also the first major industry player to deploy AMD's MI300X AI accelerator GPU, a chip designed at the time to compete with NVIDIA's AI graphics cards. Additionally, Microsoft's data centers deploy its own internally developed Maia AI chips.
Like its industry peers, Microsoft currently faces a massive shortfall in computing power demand, driven by the accelerated iteration of its in-house large models and increased compute quotas for its research divisions. The company released seven proprietary large models in June alone. However, the market performance of Microsoft's AI products has been mixed, with tools like the Office 365 Copilot assistant and GitHub Copilot receiving varied reviews and commercial results. Among the "Magnificent Seven" U.S. tech stocks, Microsoft has had the worst stock performance this year.
An increasing number of leading companies are beginning to adopt AMD products for AI acceleration workloads. Official AMD data shows that eight of the world's top ten AI giants are using AMD Instinct series GPUs to run their operations, including clients like OpenAI, Cohere, and the AI division of Elon Musk's SpaceX.
In June 2026, at a data center lab in Rockdale, Texas, AMD engineers completed the assembly and debugging of the first Helios rack AI system. The product supports four customized configuration versions; the model pictured is the one selected by Meta for deployment this year.
Meta announced in February that it would deploy AMD computing clusters with a total capacity of up to 6 gigawatts in phases, with the first 1-gigawatt phase relying on Helios racks. OpenAI and Oracle have also finalized plans for large-scale Helios deployments this year, and Indian IT giant Tata Consultancy Services (TCS) has also signed a procurement agreement.
Named after the Greek god who drove the chariot of the sun across the sky, the Helios system deeply integrates four core AMD-developed components: GPU cards, CPU processors, high-speed interconnect networks, and a complete supporting software stack.
AMD Data Center Solutions Group head Forrest Norrod stated in an interview: "The core goal of our first-generation rack system is to deliver the optimal total cost of ownership (TCO) and the industry's lowest cost per token across the entire pipeline. Customer test data confirms we have achieved this goal."
AMD CEO Lisa Su previously indicated that Helios holds significant performance advantages over NVIDIA rack products in terms of model inference, memory bandwidth, and memory capacity.
AMD has not disclosed official pricing for Helios. Industry research firm Futurum estimates the price for a single Helios unit ranges between $5 million and $5.5 million. In contrast, NVIDIA's second-generation Vera Rubin rack system is estimated to cost only $3.5 million to $4 million. Additionally, a fully configured Helios unit can weigh up to 7,000 pounds, making it larger and heavier than its NVIDIA competitor.
Futurum data shows that NVIDIA currently holds over 95% of the data center GPU market share, with AMD at just 4.5%. Helios has the potential to change this lopsided landscape.
Futurum Group CEO and industry analyst Daniel Newman believes: "AMD has a real opportunity to increase its market share to 20-25%, representing a revenue growth opportunity of hundreds of billions of dollars."
In the first quarter of 2026, the data center business became AMD's largest revenue source, with sales surging 57% year-over-year. AMD indicated that starting in 2027, its data center AI business is expected to generate tens of billions of dollars in revenue, with the vast majority contributed by the Helios system.
In the data center CPU arena, Intel remains the leader, but AMD's market share has been steadily climbing. AMD's in-house CPU design capability is a core differentiator from NVIDIA, which only launched its first server CPU in 2021 and recently adjusted its strategy to refocus on processor business.
Official Assessment: A Transformed AMD
Norrod, while showing reporters the system's core chips, described Helios as "our labor of love." Each rack contains 18 compute trays, with each tray housing four Instinct accelerator GPUs managed by a single EPYC CPU.
It was the EPYC series of data center CPUs that helped AMD reclaim its voice in the data center market, which it had lost for a decade.
In 2003, AMD launched a groundbreaking data center processor, quickly capturing nearly a quarter of the market. However, subsequent product delays and strategic missteps led to large-scale layoffs and shrinking revenue until Lisa Su took over the company and reversed its decline.
Norrod commented: "The 12 years under Lisa Su's leadership have seen this company undergo a complete transformation."
AMD officially launched its first-generation EPYC server CPU in 2017, marking a turning point for the company's performance.
"We publicly announced a complete three-generation product roadmap back then, which was highly unusual in the industry, and we strictly followed through on delivering all those products."
In the Helios product design, the hardware is paired with the current mainstream MI400 series GPUs. A single Helios tray can contain up to 12 networking chips, a technology derived from AMD's 2022 acquisition of Pensando.
Several strategic acquisitions in recent years have provided AMD with the full-stack capabilities needed to develop Helios:
The nearly $50 billion acquisition of FPGA leader Xilinx in 2022, AMD's largest-ever deal.
The approximately $5 billion acquisition of server manufacturer ZT Systems in 2025.
The acquisition of several software companies and the development of the ROCm open-source software ecosystem, positioning it against NVIDIA's dominant CUDA software system.
Counterpoint Research analyst Neil Shah noted: "AMD Helios' hardware performance is now on par with NVIDIA's GPUs and CPUs. The real killer feature of this product lies in its supporting software and deep system optimization."
"The CUDA ecosystem, built up over many years, has deep moats. NVIDIA still holds a significant lead over AMD in software."
The launch of Helios creates a window of opportunity for AMD to catch up, with the ultimate outcome depending on the actual performance seen after large-scale deployment by major customers.
Newman poses a key industry question: "Will AMD's future market share gains be driven by genuine technical advantages, or simply by the current global shortage of computing power — where as long as AMD can keep up with production, its products will sell?"
Comments