[NEW YORK] Microsoft’s deal with neocloud company Nebius Group will provide computing power to internal teams creating large language models and a consumer AI assistant, according to people familiar with the situation.
The arrangement, which is worth as much as US$19.4 billion, sparked a rally in Nebius shares when it was outlined Sep 8, but the announcement was short on specifics.
As part of the deal, Microsoft will get access to more than 100,000 of Nvidia’s latest GB300 chips, said the people, who requested anonymity to discuss an internal matter.
The strategy is aimed at coping with a shortage of AI data centre capacity and letting Microsoft free up its own server farms to provide lucrative AI services to customers.
This is just one of a string of deals Microsoft has struck with so-called neoclouds – a new category of small infrastructure providers that lease out AI-focused computing power.
Microsoft has inked more than US$33 billion in commitments to such providers, including Nebius, CoreWeave, Nscale and Lambda. The deals demonstrate the company’s growing willingness to use relative newcomers for important infrastructure.
BT in your inbox
Start and end each day with the latest news stories and analyses delivered straight to your inbox.
Nebius and the other neocloud firms declined to comment.
Cloud vendors have generally operated their own data centres, but Microsoft is struggling to bring sufficient computing capacity online. Renting access to servers from neoclouds speeds things up since they have already solved logistical challenges including obtaining sufficient power and chips.
“We are in very much land-grab mode in the AI space,” said Scott Guthrie, who leads Microsoft’s cloud efforts. “We’ve made the decision that we don’t want to be constrained in terms of capacity.”
SEE ALSO
The advent of power-guzzling generative AI has put enormous strain on data centre infrastructure. So Microsoft is freeing up capacity in its own data centres for cloud clients by running some computing for internal and OpenAI work at neocloud facilities, Guthrie said.
For example, the first foundation AI models built under consumer AI chief Mustafa Suleyman were trained at a CoreWeave data centre near Portland, Oregon, according to people familiar with the work.
In a nutshell, the strategy lets Microsoft use its own servers to sell more AI services at a time when computing capacity is tight and investors are clamouring for evidence that huge investments in the technology are paying off.
Renting data centre access also gives Microsoft more financial flexibility. For example, the company can categorise some costs as operational expenses, rather than capital costs, which offers potential benefits for cash flow, taxes and the way profit is reported to Wall Street, said Bernstein analyst Mark Moerdler.
Neocloud deals also let Microsoft pivot more quickly than when operating its own data centres, he added.
The company is not only using neocloud servers to train AI models. Recent deals with Nscale in the UK and Norway will help Microsoft deliver AI services in those regions, Guthrie said.
Early in the AI boom, the company announced it would rent capacity from Oracle to offer an AI-infused version of the Bing search engine.
Rival cloud services companies, such as Amazon.com, haven’t announced work with neoclouds to the same degree as Microsoft. Google is renting some capacity from CoreWeave as part of its work with OpenAI, Reuters has reported. Moerdler said that’s probably because Microsoft has more AI-related demand than its competitors.
Guthrie pointed to surging AI demand from both Microsoft products such as GitHub Copilot and from OpenAI, which has hundreds of millions of ChatGPT users. “Amazon doesn’t have anywhere near that scale. Nor does Google, in terms of total users with intensity, especially during business hours,” he said.
Even as Microsoft outsources some AI computing to neocloud providers, the company continues to spend aggressively on its own facilities. On Tuesday, Microsoft announced a second phase of development at a data centre in Racine, Wisconsin.
The expansion will bring the training-focused site’s total utility power capacity to at least 900 megawatts, according to people familiar with the plans, almost as much power produced by a typical nuclear reactor.
Earlier this year, Microsoft paused or stepped away from multiple projects as the company evaluated its computing needs. Guthrie said Microsoft will continue to tweak its infrastructure plans, accelerating or pulling back on certain projects or regions as it adjusts to demand and regulations. BLOOMBERG
