The chips are quite small and some can be held in the palm of a hand. They are fundamental to the development of artificial intelligence models \u2013 and the world\u2019s biggest technology companies need vast numbers of them to keep ahead.<\/p>\n
Microsoft is one of them. And, on paper, it seems to have a problem. A Guardian investigation has found an apparent discrepancy between what the company has said about its AI capacity \u2013 and the number of advanced AI chips it has in operation.<\/p>\n
It is not a small shortfall either. Microsoft reportedly<\/a> targeted having 1.8m AI chips installed in its datacentres around the globe by the end of 2024. Nearly two years on, in the middle of a $280bn (\u00a3208bn) expansion, the company has 2.2m AI chips installed, according to internal documents seen by the Guardian. This is less than half the number some experts had imagined.<\/p>\n Put simply, the global AI arms race requires a massive build-out of datacentres that run on extremely expensive chips. The apparent discrepancy over the chips suggests Microsoft\u2019s newest datacentres may not be fully operational or, if they are, they do not have the chips they need.<\/p>\n Nvidia\u2019s microchips are integral to the boom in datacentre development. Photograph: NurPhoto\/Getty Images<\/p>\n This highlights something even more fundamental about charting the progress being made in the development of AI technologies. The chips that power AI are made by Nvidia, one of the two most valuable companies in the world<\/a>. Its supply chain is one of the most tightly held secrets in the entire industry.<\/p>\n With almost no exceptions, Nvidia<\/a> does not report how many of these chips it sells or to whom. Its clients, the world\u2019s biggest tech companies, in turn do not reveal how many they have. Without this information, it is very hard for anyone to know whether AI is booming or not.<\/p>\n Microsoft: a power vacuum?<\/p>\n In the past two years, Microsoft says it has built AI infrastructure at breakneck speed. Its chief executive, Satya Nadella, said last year it would double its global datacentre footprint by mid-2027. Since 2022 it has ploughed roughly $280bn into the land, buildings and computational infrastructure to build AI. This includes more than $41bn<\/a> in the past quarter.<\/p>\n But it is difficult to estimate how many datacentres Microsoft has built with this money.<\/p>\n It is possible to assess the progress that the company is making by looking at what it has announced publicly, with a particular eye on the power it needs. Datacentres<\/a> need electricity, so one way of estimating how many datacentres are operational is to add up the energy Microsoft has at its disposal \u2013 its AI capacity.<\/p>\n Microsoft\u2019s own claims, set out in annual reports<\/a> and<\/a> quarterly<\/a> earnings<\/a>, suggest it has added 5GW of datacentre capacity over the past two years as part of its AI build-out. It says it now has hundreds of datacentres on five continents.<\/p>\n Five gigawatts is a dizzying amount of energy \u2013 it is four times the size of the largest datacentre park in Europe. But Microsoft\u2019s total capacity should be even greater than this; it has been building AI infrastructure since 2022. How much greater is an open question.<\/p>\n The hardware inside a Microsoft datacentre campus. Photograph: Audrey Richardson\/Reuters<\/p>\n In an investor presentation from 2024, Microsoft reportedly<\/a> claimed to have 5GW of datacentre capacity already installed. That would suggest it could now have a total of 10GW of capacity. It is unclear if all of these are AI datacentres \u2013 some could be for other cloud services. But Microsoft\u2019s own statements indicate that the overwhelming focus of its capital expenditures in recent years has been to build AI infrastructure.<\/p>\n Ten gigawatts of AI datacentres would suggest Microsoft should have roughly 6.4m graphics processing units (GPUs). Shaolei Ren, a professor at the University of California, Riverside, said Microsoft\u2019s sustainability reports, which contain figures for its electricity usage and are published separately from its financials, painted a different picture.<\/p>\n He said these reports suggest Microsoft\u2019s AI capacity in 2024 was probably closer to 1.2GW. But even this lower figure would indicate Microsoft would need roughly 4m AI chips \u2013 if it added 5GW of AI datacentres in the past two years.<\/p>\n \u201cAccording to their own metrics, Microsoft could be correct. But it isn\u2019t clear what they mean when they say they have added datacentre capacity. They are giving insufficient context,\u201d Ren said. \u201cThe sustainability reports are audited by a third party. They have more credibility than announcements.\u201d<\/p>\n An analyst who specialises in Nvidia said they thought Microsoft would have more chips, given its public statements. \u201cThey\u2019re low to me. They\u2019re less than I expected Microsoft would have,\u201d they said.<\/p>\n Microsoft insisted the Guardian\u2019s calculations were based on incorrect information. It did not offer any insight as to which of the Guardian\u2019s numbers were incorrect or why. What is clear is that Microsoft\u2019s build-out of AI capacity appears to be going far more slowly than its annual reports may suggest.<\/p>\n Ren said: \u201cIt may be plausible to secure or announce 1GW of power capacity within a single quarter on paper. But bringing that capacity online and actually using it for computing within the same quarter would be far more difficult.\u201d<\/p>\n Sources within Microsoft say the company\u2019s total number of AI chips has \u201cbarely moved\u201d over the past year.<\/p>\n Some of the apparent discrepancy may be explained by Microsoft\u2019s tie-up with OpenAI. The exact terms of their commercial partnership are not public, but this unit may account for some of Microsoft\u2019s datacentre deployments, which would not be in the documents the Guardian has seen.<\/p>\n A Microsoft datacentre in Middenmeer, the Netherlands. Photograph: ANP\/Shutterstock\u2018You may have a bunch of chips \u2026 you can\u2019t plug in\u2019<\/p>\n There is another factor: some of Microsoft\u2019s big projects appear to be far from operational.<\/p>\n Take Microsoft\u2019s largest AI development in the US, a pair of datacentres in Wisconsin and Georgia called Fairwater. In April, Nadella, Microsoft\u2019s chief executive, said<\/a> the Fairwater project in Wisconsin \u201cis going live\u201d.<\/p>\n Satellite footage<\/a> of the building from Epoch AI, however, appears to indicate only part of it is operational. In May, Microsoft admitted<\/a> to a Wisconsin newspaper that Fairwater was not yet online.<\/p>\n This is very common, said Ren. Initially it was a multi-gigawatt, multibillion-dollar investment. Three years later, only 300MW has been built.<\/p>\n Satya Nadella said last year Microsoft would double its global datacentre footprint by mid-2027. Photograph: Jeff Chiu\/AP<\/p>\n The internal document also indicates Microsoft has fewer of Nvidia\u2019s newest model of chip, the Blackwell, than one might expect given Nvidia\u2019s public announcements. Last March, Nvidia\u2019s chief executive, Jensen Huang, said orders<\/a> for Blackwells from Nvidia\u2019s top four customers \u2013 widely thought to be Amazon, Oracle, Microsoft and Google \u2013 amounted to 3.6m.<\/p>\n There was no breakdown given for this figure, but Microsoft has historically been one of Nvidia\u2019s largest customers. If this was still the case, that should put Microsoft\u2019s total Blackwell holdings at somewhere close to 1m chips. In fact, it has less than half of this amount installed.<\/p>\n Where are the chips, if not in the datacentres?<\/p>\n Nvidia\u2019s balance sheets appear to indicate that it has sold a great many chips; it posted a revenue of $215.9bn in February. Has Microsoft bought these but not installed them? How many, and are all of them in its possession?<\/p>\n Nadella appeared to gesture at this question on a podcast late last year called All Things AI, where he talked about Microsoft\u2019s datacentre build-out. The biggest problem, he said, was electrical power and building datacentres close enough to where power was located.<\/p>\n \u201cIf you can\u2019t do that, you may actually have a bunch of chips sitting in inventory that I can\u2019t plug in. In fact, that is my problem today. It\u2019s not a supply issue of chips. It\u2019s actually the fact that I don\u2019t have warm shells to plug into.\u201d<\/p>\n A Microsoft spokesperson said: \u201cOver several decades, Microsoft has built a global infrastructure to meet rapidly growing customer demand for cloud and AI services. Our datacentres combine custom silicon, AMD, Intel and Nvidia chips across multiple generations with the networking, storage and systems infrastructure required to operate at scale.<\/p>\n \u201cMicrosoft does not report on the volume of specific chips in its AI infrastructure. The estimates the Guardian has shared with us are inaccurate, drawing the wrong conclusions from incorrect assumptions.\u201d<\/p>\n Nvidia did not respond to a request for comment.<\/p>\n How to calculate numbers of chips from a company\u2019s \u2018AI capacity\u2019<\/p>\n The world\u2019s biggest technology companies give figures for their AI capacity in terms of power: gigawatts. One gigawatt powers between 700,000 and 1m homes. Meta says its controversial Hyperion datacentre in Louisiana will have 5GW of capacity. The UK company DataVita is planning a 1GW datacentre in Lanarkshire.<\/p>\n Converting these figures into chips means calculating how many chips can be run with that amount of power. The Guardian used the following methodology, reviewing these calculations with Abdeltawab Hendawi, a professor at the University of Rhode Island, and Ren.<\/p>\n To get a very broad approximation of how many chips there are in a datacentre, you could divide the power usage of that datacentre by the power usage of an AI chip \u2013 for example, an H100. A single H100 uses 700W. If Microsoft has 10GW of capacity, dividing this by 700 watts suggests it should have 12m chips.<\/p>\n H100s make up the bulk of the chips described in the internal document. It also indicates that Microsoft has A100s, which use less power, and Blackwells, which use more.<\/p>\n But this approximation does not account for several factors. First, datacentres have cooling systems and other equipment, which also use electricity. Ren estimates that in a given AI datacentre, 80% of the electricity goes to computer chips. This is roughly in accordance<\/a> with figures from the International Energy Agency, although the number depends on the efficiency of the datacentre. Eighty per cent of 10GW would suggest 8GW may actually be in use.<\/p>\n