The Guardian investigation showMicrosoft’s statements about the company’s artificial intelligence capabilities are inconsistent with data on the number of chips in operation. According to reports, Microsoft previously expected to install 1.8 million artificial intelligence accelerators in global data centers by the end of 2024, but according to internal documents, it has now deployed “only” 2.2 million accelerators in its $280 billion artificial intelligence expansion plan. That’s only half of what experts predicted.
This data discrepancy may mean that the latest data centers are not yet operational, or that there are not enough wafers available. Assessing the company’s actual progress is difficult, in part because information in NVIDIA’s supply chain is confidential. Like its customers, NVIDIA itself rarely discloses how many chips it sells and to whom, and if it doesn’t, it’s difficult to determine whether the AI industry is actually growing as advertised.
In the past two years, Microsoft has announced the rapid construction of AI infrastructure. It was reported last year that the company would double the size of its data center network by mid-2027. Since 2022, it has allocated about $280 billion to buy land, buildings and computing artificial intelligence infrastructure. last season More than US$41 billion has been invested. However, it is difficult to determine exactly how many data centers were built with this money.
Image source: Microsoft
Progress can be assessed using indirect data. Microsoft’s own reports show that the company has added 5 gigawatts of data center capacity as part of its AI infrastructure buildout. The company currently says it has hundreds of data centers across five continents. In fact, the capacity should be higher because the company has been building out its AI infrastructure since 2022.
Microsoft’s internal presentation of 2024 shows that even then, the company’s data center installation capacity will reportedly reach 5 gigawatts. Therefore, the current total capacity is about 10 GW, but there is no data to indicate whether all these data centers support AI services or whether some data centers serve other cloud services. However, Microsoft itself claims that the vast majority of capital expenditures in recent years have been on artificial intelligence infrastructure.
If we assume that Microsoft has 10 GW of AI data centers, then it should have about 6.4 million AI accelerators, but some experts believe the situation is different. According to some evidence, Microsoft’s data center AI capacity in 2024 is 1.23GW, but if it adds another 5GW, it will require about 4 million AI accelerators. Experts believe that third-party data on sustainability should first and foremost be trustworthy because they are verified by independent agencies.
Microsoft itself said the Guardian’s calculations were incorrect, although they didn’t say what exactly the error was. Some experts believe that judging from the company’s annual report, the progress of artificial intelligence production capacity construction is slower than expected. Reporters reported that sources within the company confirmed that the total number of artificial intelligence chips “has remained almost unchanged” over the past year. Part of the ‘shortage’ may be caused by partnerships with OpenAIbecause some Microsoft data centers associated with the company may not appear in internal documents.
In addition, some large data center projects may not yet be fully operational. For example, Microsoft’s largest artificial intelligence project in the United States is Fairwater Wisconsin and Georgia. The company said in April that its Wisconsin data center was “coming online,” but satellite images showed the facility was only partially operational. In May, Microsoft admitted that Fairwater wasn’t yet available.
Internal documents also show that the company has fewer NVIDIA Blackwell chips than expected based on NVIDIA’s public statements. According to indirect data, Microsoft should have about 1 million of these chips, but the number actually installed is less than half that number. Microsoft CEO Satya Nadella said the main problems are power shortages and data center construction issues. Those ones. A company may have a large inventory of wafers but no existing facilities to house them.
As a Microsoft representative said, the company uses a variety of chips, ranging from own NVIDIA, AMD and Intel develop hardware, among others, so the Guardian’s estimate is inaccurate and “based on incorrect assumptions, leading to incorrect conclusions.” The Guardian used a method developed by scientists whereby the total power is divided by the energy consumption of an accelerator. Microsoft most commonly NVIDIA H100power consumption is about 700W. If the capacity of Microsoft’s data center is 10GW, simply dividing it into 12 million chips. However, Microsoft also has a lot of less powerful features A100 and more energy consuming blackwell.
Of course, such estimates do not take into account the use of cooling systems and the energy consumption of other equipment. It is estimated that in artificial intelligence data centers, only about 80% of the total energy consumption is consumed directly on the chips. In other words, 10 GW of electricity may have 8 GW of computing power. Therefore, a Microsoft report talked about the use of 89% artificial intelligence chips.
Furthermore, not all chips are AI accelerators. Memory modules, drives, processors, etc. all require energy. A server with eight H100s consumes up to 10 kW. If you divide 8 GW by 10 kW, you get 800,000 servers or about 6.4 million chips. Experts believe this is a “conservative” estimate because in practice, companies like Microsoft “overload” their data centers by installing chips beyond what a site can support.
source: