Microsoft, OpenAI may be dreaming of $100B 5GW AI 'Stargate' supercomputer

Play it again, Sam

OpenAI is believed to be in talks with Microsoft to construct a massive supercomputer code-named Stargate containing millions of AI accelerators at a cost of up to $100 billion.

Citing sources, The Information reports [paywall] that Microsoft is deciding whether to finance the project, which might cost as much as 100 times that of setting up a modern datacenter and be ready by 2028. We imagine a lot of Redmond's dosh will be blown on the super's machine-learning accelerators which can cost tens of thousands of dollars apiece, depending on the supplier.

Microsoft has staked the future of its AI products on OpenAI's models, for which it furnishes the GPU clusters necessary to train and run that software. In January last year, Microsoft announced the third-phase of its partnership with the AI juggernaut, which included investing billions of dollars into infrastructure to maintain its status as OpenAI's exclusive cloud provider.

Redmond is said to be developing another supercomputer for the fourth-phase of this AI collaboration, due to launch in 2026.

Stargate would reportedly be the largest in a series of AI supercomputers built by Microsoft for OpenAI over the next six years and could consume as much as five gigawatts when complete.

Given the timeline, it's hard to say whose accelerators Stargate will end up using. Microsoft currently employs Nvidia, AMD, and its own Maia 100 accelerators for a variety of workloads including the GPT-4 and GPT-3.5 models. Meanwhile OpenAI is rumored to be developing its own AI chips.

If built today, a five gigawatt supercomputer could accommodate in excess of 40,000 of Nvidia's DGX GB200 NVL72 rack systems containing 2.88 million Blackwell GPUs capable of 14,400 exaFLOPS of dense FP8 performance.

Finding enough carbon-free energy to power such a system without completely derailing Microsoft's aggressive eco-promises could, however, prove troublesome. Even the largest nuclear power plant in the US, the Palo Verde Generating Station west of Phoenix, Arizona, can only muster 3.3 gigawatts of electric power from its three reactors.

In order to satiate its own energy demands, Amazon recently paid $650 million for Cumulus Data's nuclear-powered datacenter facilities in Pennsylvania. The multi-year deal will provide AWS access to as much as 960 megawatts of power from the 2.5 gigawatt Susquehanna site — less than a fifth of what Microsoft is apparently considering.

It's not clear whether Stargate would occupy a single building or be spread across one or many sites. However, access to power may necessitate a distributed approach. OpenAI supremo Sam Altman is certainly a big dreamer. Previous reports have alleged that he was seeking $7 trillion in funding for a massive network of accelerator chip fabs, though he later shrugged off those claims telling folks not to believe everything you read on the internet. There's no guarantee Stargate will turn out to be anything more than a concept.

The Register reached out to Microsoft and OpenAI for comment.

Talks regarding the construction of Stargate reportedly date back to the middle of last year. Meanwhile, Microsoft's close collaboration with OpenAI may have played into the software titan's decision to explore small modular reactors (SMR) to power future sites.

SMRs are just what they sound like: Miniaturized fission reactors capable of producing tens to hundreds of megawatts depending on the design. Tiny reactors are nothing new — they've been powering submarines going back to 1954 — but the argument for SMRs is that they will, in theory, be less expensive to manufacture and operate. These qualities have made them a tantalizing option for datacenter operators.

Last year, we reported that Microsoft had hired the former director of nuclear strategy & programs at Ultra Safe Nuclear Corporation (USNC) to head up a division tasked with developing a nuclear strategy. But we'll note that commercial SMRs, like those being developed by NuScale or Altman-backed Oklo, have yet to hit the market.

Another interesting wrinkle in this is that the 2028 timeline aligns with Microsoft's previously announced power purchase agreement with another Altman backed energy startup: Helion Energy, which is developing a fusion reactor. With that said, when we spoke with Helion last year about its tech, it was only targeting 50 megawatts of output, far from the five gigawatts reportedly needed to power Stargate. ®

Want more? Check out The Next Platform... $100 billion doesn't sound so steep for moneybags Microsoft.

More about

TIP US OFF

Send us news


Other stories you might like