Microsoft Taps Rack-Scale AI System to Challenge Nvidia

After a decade-long strategic resurgence, chip giant Advanced Micro Devices (AMD) is poised to disrupt the burgeoning artificial intelligence hardware market with its inaugural rack-scale AI system, codenamed “Helios.” This formidable offering is set to enter the fray, challenging the dominance of Nvidia’s established AI infrastructure and securing significant traction with a growing roster of high-profile clients, notably including Microsoft.

Helios represents AMD’s most ambitious play to date in the AI compute arena, aiming to directly rival Nvidia’s highly sought-after Grace Blackwell and Vera Rubin systems. This strategic move signals AMD’s intent to present the world’s most valuable chipmaker with its most substantial competitive threat in years, potentially reshaping the landscape of AI hardware procurement.

Microsoft officially announced its intention to integrate the Helios system into its extensive data center operations. This endorsement places Microsoft alongside other tech titans such as Meta, OpenAI, and Oracle, all of whom are aggressively expanding their AI compute capabilities to fuel the development and deployment of next-generation artificial intelligence applications. AMD plans to commence shipments of the Helios system to its customers, including Microsoft, later this year, although specific financial terms and the exact scale of compute capacity involved have not been publicly disclosed.

“We are continuously expanding the Azure infrastructure portfolio with solutions like AMD Helios to equip our customers with the unparalleled performance, scalability, and choice essential for building and running the future of AI applications,” stated Microsoft CEO Satya Nadella in a press release. The Helios system is slated to power frontier model inference for Microsoft, its AI clientele, and bolster Azure’s comprehensive AI services. Furthermore, Microsoft will introduce two new computing instances leveraging AMD’s latest “Venice” central processing units (CPUs) – one optimized for agentic AI and sophisticated data pipelines, and another designed for advanced semiconductor design workflows.

This collaboration underscores a long-standing and evolving partnership between AMD and Microsoft. AMD’s silicon has been integral to Microsoft’s Surface PCs and Xbox gaming consoles for numerous years. In a significant precursor to Helios, Microsoft was the first major adopter of AMD’s MI300X graphics processing unit (GPU) in 2023, a chip that demonstrably challenged Nvidia’s AI offerings. Microsoft also deploys its proprietary Maia chips within its own data centers, showcasing a multifaceted approach to AI infrastructure.

The insatiable demand for AI compute power is a defining characteristic of the current technology landscape. Microsoft, like its industry peers, is relentlessly pursuing greater computational resources, particularly as it intensifies its internal model development and allocates significant capacity to research and development initiatives. In June, the tech giant unveiled seven internally developed AI models, underscoring its commitment to expanding its AI ecosystem. While Microsoft’s AI endeavors have yielded a spectrum of results, from its 365 Copilot AI assistant to its GitHub Copilot coding agent, the company’s overall AI strategy continues to be a focal point of market analysis.

Microsoft’s adoption of Helios positions AMD as a critical player in the AI acceleration market, catering to a growing number of major technology firms. AMD reports that eight of the top ten AI companies now utilize its Instinct GPUs for their workloads. This esteemed clientele includes OpenAI, Cohere, and Elon Musk’s ventures, such as SpaceXAI, which is affiliated with SpaceX. This broad adoption signals a significant shift in market dynamics, with AMD establishing itself as a viable alternative to Nvidia’s entrenched position.

The momentum behind AMD’s AI push was further amplified in February when Meta announced its intention to deploy up to 6 gigawatts of AMD GPUs over time, commencing with 1 gigawatt integrated into Helios racks by the end of the year. OpenAI and Oracle have also made substantial commitments to deploy Helios this year, alongside Tata Consultancy Services, India’s largest IT services company, which has pledged to utilize the system. This widespread commitment from industry leaders underscores the perceived value and potential of AMD’s Helios offering.

Driving Down the Cost Per Token

Named after the Greek deity who chariot-pulled the sun across the sky, Helios embodies AMD’s integrated approach, consolidating its in-house expertise across GPUs, CPUs, networking, and software into a single, cohesive system. “Our primary focus is on delivering the most advantageous total cost of ownership, achieving the lowest cost per token, comprehensively,” stated Forrest Norrod, head of AMD’s data center division, in an exclusive conversation with CNBC regarding the first-generation Helios system. “And our customers are confirming that we are successfully meeting this objective.”

Earlier in May, AMD CEO Lisa Su shared with CNBC’s Jim Cramer that Helios offers “significant advantages” over Nvidia’s rack-scale systems, particularly in the realms of inference, memory bandwidth, and overall memory capabilities. While AMD has remained tight-lipped about the precise pricing, industry analysts at The Futurum Group estimate Helios to cost between $5 million and $5.5 million. This figure is juxtaposed against their estimates for Nvidia’s second-generation rack-scale system, Vera Rubin, which is projected to range from $3.5 million to $4 million. In terms of physical specifications, Helios, weighing up to 7,000 pounds, is also notably wider and heavier than Nvidia’s Vera Rubin.

Nvidia currently commands an overwhelming majority of the data center GPU market, holding over 95% share according to The Futurum Group. AMD, while possessing a more modest 4.5% market share, views Helios as a pivotal opportunity to significantly alter this dynamic. “There is a strong possibility that AMD can achieve substantial success, potentially capturing 20% to 25% of the market. This represents hundreds of billions of dollars in potential revenue,” commented Daniel Newman, an analyst and CEO of The Futurum Group. This optimism reflects a broader sentiment that AMD is not just participating but actively aiming to redefine the competitive landscape.

The strategic importance of the data center segment for AMD is evident in its financial performance. In the first quarter of 2026, data centers constituted the majority of AMD’s revenue, marking a substantial 57% year-over-year increase. AMD has indicated its intention to generate tens of billions in data center AI revenue starting in 2027, with Helios expected to be the primary driver of this growth. While Intel continues to hold a commanding lead in the data center CPU market, AMD has been steadily eroding its market share. This established strength in the CPU market provides AMD with a distinct advantage over Nvidia, which only launched its inaugural server CPU in 2021 and has recently reaffirmed its strategic focus on this segment.

Forging a ‘Very Different AMD’

Norrod described Helios as “our baby,” offering CNBC an in-depth look at the system’s core components. Each of its 18 compute trays is equipped with four Instinct GPUs, centrally managed by a single EPYC central processing unit. It was the introduction of these EPYC data center CPUs that propelled AMD’s significant recovery in the data center market after a decade of competitive stagnation. In 2003, AMD enjoyed a commanding presence in the data center market with its Opteron chips, capturing nearly a quarter of the market within three years. However, a series of product delays and strategic missteps led to a dramatic decline in market share, culminating in significant layoffs and shrinking revenue by the time Lisa Su assumed leadership.

“Under Lisa’s leadership over the past 12 years, it has been a profoundly transformed AMD,” Norrod affirmed, highlighting the company’s remarkable turnaround. This resurgence began in earnest with the unveiling of the first EPYC server CPU in 2017. Norrod elaborated on AMD’s strategic approach: “One of the key initiatives we undertook was the detailed outlining of our product roadmap for three generations, which is a highly unusual strategy. And, crucially, we delivered precisely what we promised.”

The execution of this roadmap included the planned launch of Helios, powered by the current generation of MI400 series GPUs. Each Helios tray also incorporates up to 12 networking chips, utilizing technology acquired through AMD’s strategic purchase of Pensando in 2022. This acquisition was one of several strategic moves in recent years that have been instrumental in enabling the development of Helios. AMD’s most significant acquisition to date was that of programmable chip company Xilinx for nearly $50 billion in 2022. Further bolstering its infrastructure capabilities, AMD acquired server builder ZT Systems for approximately $5 billion in 2025. Concurrently, a series of software company acquisitions have strengthened AMD’s ability to develop ROCm, its open-source alternative to Nvidia’s widely adopted CUDA software ecosystem.

Neil Shah, an analyst at Counterpoint Research, views AMD’s Helios chips as “on par” with Nvidia’s GPUs and CPUs, emphasizing that the “secret sauce lies in the software and optimization.” He further noted, “With CUDA, Nvidia possesses a more mature and expansive ecosystem, giving it a significant advantage over AMD.” Nonetheless, Helios presents AMD with a significant opportunity to make substantial inroads into the AI hardware market, contingent on the success of its early deployments. “The critical question will be: is AMD achieving market traction due to technological superiority, or is it simply capitalizing on an extreme capacity constraint where demand outstrips supply, leading customers to acquire whatever they can obtain?” questioned Newman.