After a decade-long comeback, semiconductor giant Advanced Micro Devices is preparing to ship its first artificial intelligence rack-scale system, called Helios, to a growing list of customers that now includes Microsoft.
It’s the first rival to Nvidia’s hugely popular Grace Blackwell and Vera Rubin systems, and aims to give the world’s most valuable chipmaker its first serious competition in years.
Microsoft announced Monday that it will use Helios systems in its data centers, meta, OpenAIOracle, and others are competing to capture as much computing as possible.
AMD plans to start shipping it to customers including Microsoft later this year.
Details regarding financial terms and amount of computing power were not disclosed.
“With AMD Helios, we are expanding our Azure infrastructure portfolio to give customers the performance, scale, and choice they need to build and run next-generation AI applications,” Microsoft CEO Satya Nadella said in a press release.
The new Helios system powers frontier model inference for Microsoft and its AI customers and supports Azure AI services. Microsoft also plans to add two new compute instances running on AMD’s latest Venice central processing units (CPUs), one for agent AI and data pipelines and one for semiconductor design.
This is a continuation of a long-standing partnership that has seen AMD chips power Microsoft’s Surface PCs and Xbox game consoles for years. In 2023, Microsoft first adopted AMD’s MI300X graphics processing units (GPUs), which rival Nvidia’s AI chips. Microsoft is also deploying its own Maia chips in its data centers.
Like its competitors, Microsoft needs as much computing as possible, especially as it ramps up its own model development and allocates more computing power to research and development. In June, the company announced seven in-house models. From the 365 Copilot AI Assistant to the GitHub Copilot Coding Agent, Microsoft’s AI efforts to date have had mixed results. It has been the worst performer of the Magnificent Seven stocks so far this year.
Microsoft is part of a growing number of large companies turning to AMD to accelerate AI. AMD says eight of the top 10 AI companies run workloads on its Instinct GPUs, including OpenAI, Cohere and Elon Musk’s SpaceXAI, part of SpaceX.
AMD employees install the first Helios rack-scale AI system in a data center lab in Rockdale, Texas, on June 24, 2026. Helios has four customizable configurations, which Meta will roll out later this year.
Andrew EversCNBC
Meta announced in February that it would use up to 6 gigawatts of AMD GPUs over time, starting with 1 gigawatt coming to Helios racks later this year. OpenAI and Oracle also made major commitments to deploy Helios this year, and India’s largest IT company Tata Consultancy Services also committed to using Helios.
CNBC got the world’s first detailed look inside the Helios system from the Texas data center lab where it is developed and tested.
“Lowest cost per token”
Helios is named after the ancient Greek god who draws the sun into the sky with the help of four horses. Helios brings together four functions that AMD does in-house: GPU, CPU, networking, and software.
“We’re all very focused on delivering the highest total cost of ownership, lowest cost per token,” data center director Forrest Norrod told CNBC about AMD’s first-generation systems. “And our customers tell us we’re achieving that.”
In May, AMD CEO Lisa Su told CNBC’s Jim Cramer that Helios has “significant advantages” over Nvidia’s rack-scale systems “both in terms of inference and memory bandwidth and memory capabilities.”
AMD declined to comment on cost, but Futurum Group estimates Helios will cost between $5 million and $5.5 million. This compares to Futurum’s estimate of $3.5 million to $4 million for Vera Rubin, Nvidia’s second-generation rack-scale system.
At up to 7,000 pounds, Helios is wider and heavier than Nvidia’s Vera Rubin.
According to Futurum Group, Nvidia controls more than 95% of the data center GPU market. AMD only owns about 4.5% of the market, but Helios could change that.
“I think there’s a serious case for AMD to do well and potentially reach 20%, 25%. By the way, that’s hundreds of billions of dollars in revenue,” said Daniel Newman, analyst and CEO of Futurum Group.
In the first quarter of 2026, data centers accounted for the majority of AMD’s revenue, increasing 57% year-over-year. AMD told CNBC it plans to record tens of billions of dollars in data center AI revenue starting in 2027, the bulk of which will come from Helios.
In terms of data center CPU market share, intel Although still the clear leader, AMD is steadily gaining ground. This CPU leadership sets AMD apart from Nvidia, which launched its first server CPUs in 2021 and pivoted its strategy to focus on chips again this year.
“A completely different AMD”
Norod called Helios “our baby” while showing CNBC the system’s core chip. Each of the 18 compute trays has four Instinct GPUs with a single EPYC central processing unit.
These EPYC data center CPUs helped AMD regain the leadership it lost in the data center market for a decade.
In 2003, AMD developed a revolutionary data center CPU that quickly helped the company capture nearly a quarter of the market, but a series of delays and failures decimated some of it, leading to massive layoffs and shrinking revenues by the time Su took the helm.
“Over the past 12 years, under Lisa’s leadership, AMD has changed significantly,” said Norrod.
Things took a turn for the better after the company unveiled its first EPYC server CPU on stage in 2017.
“One of the things we’ve done is we’ve developed a detailed roadmap for three generations, which is very unusual,” he said. “And we did exactly what we said we would.”
Part of AMD’s roadmap included plans to launch Helios with current MI400 series GPUs.
Each Helios tray also has up to 12 networking chips made with technology AMD acquired when it acquired Pensando in 2022.
This is one of several acquisitions that have helped enable the development of Helios over the past few years.
AMD’s biggest acquisition to date was programmable chip company Xilinx in 2022 for nearly $50 billion. AMD also acquired server maker ZT Systems in 2025 for nearly $5 billion, along with a series of software companies that helped develop ROCm, an open source alternative to NVIDIA’s widely adopted CUDA software ecosystem.
Neil Shah, an analyst at Counterpoint Research, said AMD’s Helios chips are “on par” with Nvidia’s GPUs and CPUs, but “the secret sauce is in the software and optimization.”
“With CUDA, I think Nvidia has a larger ecosystem and is much more advanced than AMD,” he said.
With Helios, AMD has the opportunity to make significant advances depending on how well the initial rollout goes.
“The question becomes: Will AMD win because it’s better technologically? Or will AMD win because it’s so capacity-constrained that if it can be developed, someone will buy it?” Newman said.
WATCH: First look at Vera Rubin, Nvidia’s AI system 10x more efficient than Grace Blackwell
—CNBC’s Jordan Novet contributed to this report.
