After a decade-long comeback, chip large Advanced Micro Devices is getting ready to ship its first rack-scale system for artificial intelligence, known as Helios, to a rising checklist of consumers that now contains Microsoft.
It is the primary rival to Nvidia‘s wildly well-liked Grace Blackwell and Vera Rubin techniques, and is aiming to present the world’s most respected chipmaker its first actual competitors in years.
Microsoft introduced Monday it should use the Helios system in its information facilities, becoming a member of Meta, OpenAI, Oracle and others in a race to seize as a lot compute as doable.
AMD will start delivery to clients, together with Microsoft, later this yr.
Particulars about monetary phrases or the quantity of compute capability weren’t disclosed.
“We’re increasing the Azure infrastructure portfolio with AMD Helios to present clients the efficiency, scale and selection they should construct and run the subsequent technology of AI purposes,” Microsoft CEO Satya Nadella wrote in a press launch.
The brand new Helios system will energy frontier mannequin inference for Microsoft, its AI clients and help Azure AI companies. Microsoft will even add two new computing situations run on AMD’s newest “Venice” central processing models, or CPUs, one for agentic AI and information pipelines, and one other for semiconductor design.
It is the continuation of a longtime partnership, with AMD chips powering Microsoft’s Floor PCs and Xbox gaming consoles for a few years. In 2023, Microsoft was additionally the first to adopt AMD’s MI300X graphics processing unit, or GPU, that rivaled Nvidia’s AI chips. Microsoft additionally deploys its personal Maia chips in its information facilities.
Like its friends, Microsoft wants as a lot compute as doable, particularly because it ramps up its personal mannequin improvement and allocates extra computing capability to analysis and improvement. In June, it announced seven models constructed in-house. Microsoft’s AI efforts to this point have seen mixed results, from its 365 Copilot AI assistant to its GitHub Copilot coding agent. It is the worst-performing “Magnificent Seven” inventory to date this yr.
Microsoft is a part of a rising variety of huge corporations turning to AMD for AI acceleration. AMD says eight of the highest 10 AI corporations run workloads on its Intuition GPUs, together with OpenAI, Cohere and Elon Musk‘s SpaceXAI, which is a part of SpaceX.
An AMD worker installs a the primary Helios rack-scale AI system in a knowledge middle lab in Rockdale, Texas, on June 24, 2026. Helios is available in 4 customizable configurations, and that is the one Meta will deploy later this yr.
Andrew Evers | CNBC
In February, Meta introduced it will use up to 6 gigawatts of AMD GPUs over time, beginning with 1 gigawatt deployed on Helios racks later this yr. OpenAI and Oracle additionally made main commitments to deploy Helios this yr, with India’s largest IT company, Tata Consultancy Services, committing to make use of it as properly.
CNBC acquired the world’s first detailed look inside a Helios system, from the Texas information middle lab the place it is being developed and examined.
‘Lowest price per token’
Named for an historic Greek god who pulls the solar throughout the sky with the assistance of 4 horses, Helios brings collectively 4 issues AMD does in-house: GPUs, CPUs, networking and software program.
“We’re very centered on offering the most effective whole price of possession, the bottom price per token, all in,” information middle head Forrest Norrod advised CNBC about AMD’s first-generation system. “And our clients are telling us that we’re attaining that.”
In Could, AMD CEO Lisa Su advised CNBC’s Jim Cramer that Helios has “important advantages” over Nvidia’s rack-scale techniques, “if you’re speaking about inference and if you’re speaking about reminiscence bandwidth and reminiscence capabilities.”
Whereas AMD would not touch upon price, the Futurum Group estimates Helios will price between $5 million and $5.5 million. That is in contrast with Futurum estimates of $3.5 million to $4 million for Nvidia’s second-generation rack-scale system, Vera Rubin.
At as much as 7,000 kilos, Helios can be wider and heavier than Nvidia’s Vera Rubin.
Nvidia controls greater than 95% of the information middle GPU market, in accordance with the Futurum Group. AMD solely holds some 4.5% of the market, however Helios may change that.
“I believe there is a critical case wherein AMD does nice and may get to twenty% and 25%. And by the best way, that is a whole bunch of billions of {dollars} of income,” stated Daniel Newman, analyst and CEO of the Futurum Group.
Within the first quarter of 2026, information facilities made up nearly all of AMD’s income, up 57% yr over yr. AMD advised CNBC that it plans to guide tens of billions in information middle AI income beginning in 2027, the bulk coming from Helios.
In information middle CPU market share, Intel stays the clear chief, however AMD has steadily been gaining floor. This CPU management units AMD other than Nvidia, which launched its first server CPU in 2021 and shifted methods to renew focus on the chips this yr.
‘A really totally different AMD’
Norrod known as Helios “our child,” as he confirmed CNBC the system’s core chips. Every of its 18 compute trays has 4 Intuition GPUs powered by a single EPYC central processing unit.
It was these EPYC information middle CPUs that helped AMD regain a decade of misplaced management within the information middle market.
In 2003, AMD had a groundbreaking information middle CPU that helped it rapidly gain nearly a quarter of the market, however that slice withered away following a sequence of delays and missteps that led to main layoffs and shrinking revenue by the point Su took the helm.
“Beneath Lisa’s management for the final 12 years, it has been a really totally different AMD,” Norrod stated.
Issues circled after the corporate unveiled the primary EPYC server CPU on stage in 2017.
“One of many issues that we did is we laid out our highway map intimately for 3 generations, which could be very uncommon,” he stated. “And we delivered precisely what we stated.”
A part of AMD’s highway map included plans to launch Helios with the present MI400 sequence of GPU.
Every Helios tray additionally has as much as 12 networking chips made with know-how AMD acquired when it bought Pensando in 2022.
It was one among a number of acquisitions that has helped allow Helios improvement in the previous few years.
AMD’s largest buy to this point was programmable chip firm Xilinx for almost $50 billion in 2022. AMD additionally acquired server maker ZT systems for almost $5 billion in 2025, and a sequence of software program corporations that helped it develop ROCm, its open-source different to Nvidia’s broadly adopted CUDA software program ecosystem.
Counterpoint Analysis analyst Neil Shah stated AMD’s Helios chips are “on par” with Nvidia GPUs and CPUs, however the “secret sauce is within the software program and optimization.”
“With CUDA, I believe Nvidia has a much bigger ecosystem, and it is fairly forward versus AMD,” he stated.
With Helios, AMD has the chance to make substantial strides, relying on how properly early deployments fare.
“The query goes to be: Is AMD successful as a result of they’re technologically superior? Or does AMD win as a result of there’s simply such a constraint on capability that if they’ll construct it, somebody will purchase it?” Newman stated.
WATCH: First look at Vera Rubin, Nvidia’s AI system that’s 10 times more efficient than Grace Blackwell
— CNBC’s Jordan Novet contributed to this report.