Inside Helios: AMD's 7,000-Pound, $5 Million Bet on Breaking Nvidia's Grip

AMD is installing the machine it hopes will end Nvidia's near-monopoly on AI computing. According to CNBC's on-site teardown — described in the report as the world's first up-close teardown of a Helios system — the rack is a "7,000-pound, $5 million behemoth" packing 72 GPUs and 18 CPUs into a single cabinet.

"This is Helios. It's our baby. It's 72 GPUs and 18 CPUs per rack," an AMD engineer told the crew. The report places the never-before-shown "130,000 ft" mega lab at a data center outside Austin, Texas, while the first fully functional Helios rack CNBC observed being installed was at a Rockdale, Texas data center. Data center lead Forrest Norrod walked through the system's 18 compute trays — "these nine here and these nine here that are in the center connected with six switches." On each tray, four GPUs sit beneath copper cold plates with liquid running through them and connect to a single "Venice" CPU at the front.

Microsoft joins the customer list

The most newsworthy element is a customer disclosure timed to publication: Microsoft has signed up for Helios systems in its data centers. Microsoft was also first to adopt AMD's MI300X GPU in 2023. That places it alongside Meta, OpenAI and Oracle, which the report says have all signed deals to begin deploying AMD's system this year. India's largest IT company, TCS, is described as committing to use Helios eventually.

Meta's commitment is the largest disclosed: up to 6 gigawatts of AMD GPUs, starting with 1 gigawatt deployed on Helios racks later this year — an announcement the report credits with a 7% bump in AMD's stock in February. The Rockdale rack CNBC saw is the configuration Meta will deploy, with some 85 racks coming now and more expansion planned.

The price gap AMD won't discuss

Here the reporting gets uncomfortable for AMD. The company won't comment on how much Helios costs, saying it varies depending on configuration. The Futurum Group's estimate, cited in the report, puts each Helios system at about $5 million against $3.5 to $4 million for Nvidia's Vera Rubin. If that estimate holds, AMD is entering the market as the more expensive option — an unusual position for a challenger. Asked directly what would make a customer pay more, an AMD executive said, "I can't comment on Vera Rubin's cost," then pivoted to "the best total cost of ownership, the lowest cost per token all-in," saying customers report AMD is achieving it. That is an assertion, not a demonstration, and no benchmark data appears in the report to support it.

Scale, power and water

Asked how many new hires it took to pull Helios off, an executive deflected to scale instead: "we have a town hall where we meet with everybody working on Helios every quarter, and there's 8,000 people in that meeting." Each rack draws roughly 225,000 to 245,000 watts. The liquid-cooled installation runs "about 750 gallons per minute of chilled water" — which AMD says is a closed-loop system, adding, "we're not wasting any water."

On supply, the MI455 GPUs are manufactured in Taiwan on TSMC's most advanced 2nm node, and AMD says it is already building data center chips at TSMC Arizona. Because Nvidia has reserved the majority of TSMC's leading advanced-packaging capacity, AMD committed $10 billion in May to other Taiwanese packaging firms such as ASE. Each MI455 carries 12 stacks of high-bandwidth memory, up to 432GB per GPU; AMD says close relationships with all three major memory suppliers have let it secure what it needs despite the global shortage.

What remains unproven

AMD says its debut system performs better, more efficiently, and is more customizable than both Grace Blackwell and Vera Rubin, and frames its openness — open standards, the open-source ROCm stack, support for PyTorch, vLLM and SGLang — as the key differentiator against Nvidia's proprietary CUDA. But an analyst quoted in the report undercuts that framing: "with CUDA, I think Nvidia has a bigger ecosystem and it's quite ahead versus AMD. So otherwise I think with respect to particular CPU or GPU I think they are on par but the secret sauce is in the software and optimization." Another frames the real question bluntly: "Is AMD winning because they're technologically superior? Does AMD win because there's just such a constraint on capacity that if they can build it, someone will buy it?"

The gap is enormous. Nvidia controls more than 95% of the data center GPU market, per the report. One analyst sees "a serious case in which AMD does great and can get to 20 and 25%. And by the way, this is hundreds of billions of dollars of revenue." AMD's executives decline to name a target, saying only, "You don't play in markets to be a single digit player." AMD says data center revenue is up 57% year over year, and an executive states it is the company's fastest growing segment and, as of Q1 2026, the majority of its revenues. AMD expects tens of billions of dollars in data center AI revenue starting next year, the majority coming from Helios.

This account was compiled from a single source: CNBC's on-site Helios teardown at AMD's Rockdale and Austin, Texas facilities. Because no second outlet in our index independently covers the Helios launch, the cost estimate, deployment timelines, market-share figures and performance claims should be treated as reported-but-uncorroborated. Figures — particularly the Futurum cost estimate and customer deployment schedules — may change as more outlets report and as systems actually ship.