Microsoft Has Millions of AI Chips. Can It Put Them to Work?

Rapid AI infrastructure expansion at Microsoft is facing questions over how much of its planned computing capacity is actually operational.

Advertisement

A Guardian investigation published on August 17 reported that internal Microsoft documents showed the company had about 2.2 million AI chips installed. Microsoft had reportedly targeted 1.8 million chips by the end of 2024.

The figures do not point to a simple chip shortage. Instead, the investigation raises questions about data-centre construction, electricity availability and the time needed to bring new AI infrastructure online.

Microsoft has disputed the Guardian’s estimates, saying they were inaccurate and based on incorrect assumptions.

$280 Billion AI Infrastructure Investment

Advertisement

According to the Guardian report, Microsoft has invested about $280 billion in AI infrastructure since 2022. The company has also said it added around 5GW of data-centre capacity over the past two years.

Also Read: How AI Is Changing Workforce Skills and Jobs at Infosys

However, the Guardian’s investigation raises questions about how much of this capacity is actually operational.

University of California, Riverside professor Shaolei Ren estimated that Microsoft’s AI capacity in 2024 may have been closer to 1.2GW. He estimated that, combined with the reported 5GW addition, Microsoft would need roughly 4 million AI chips. These are external estimates, not figures confirmed by Microsoft.

Microsoft disputed the Guardian’s calculations, saying the estimates were based on incorrect information and assumptions. However, the company did not specify which figures it considered incorrect.

Advertisement

Data Centres and Power Are Critical

The development points out a broader challenge facing the AI industry. Buying advanced chips is only one part of building AI capacity. Companies also need data centres, electricity, cooling systems, networking and other infrastructure to operate those chips at scale.

Microsoft has also acknowledged the growing need for physical infrastructure and reliable power as AI demand increases. Its recent sustainability report notes that access to affordable and reliable electricity is becoming an important requirement for AI and cloud expansion.

The Guardian investigation also examined Microsoft’s Fairwater data-centre project as part of its assessment of the company’s AI infrastructure rollout.

At the same time, Microsoft continues to expand its data-centre footprint. Its India South Central cloud region in Hyderabad has now become operational, taking Microsoft’s total number of cloud regions in India to four.

Microsoft Disputes the Estimates

Microsoft has rejected the Guardian’s calculations, saying they were inaccurate and based on assumptions that led to incorrect conclusions.

The company has not publicly provided a complete alternative breakdown of its AI-chip inventory, making it difficult for outsiders to independently establish the exact number of AI accelerators Microsoft operates.

Nvidia also does not publicly disclose detailed chip sales to individual customers, adding to the difficulty of independently verifying Microsoft’s total AI-chip deployment.

Therefore, the 2.2 million figure reported from internal documents should be treated as a reported number, while estimates of Microsoft’s potential GPU requirement or operational power capacity should be clearly identified as external calculations.

The Bigger AI Infrastructure Challenge

Microsoft’s situation shows a larger issue for the AI industry. As companies commit billions of dollars to AI infrastructure, the ability to expand computing capacity depends not only on access to GPUs but also on power availability, data-centre construction and the ability to bring new facilities online.

For enterprises planning large AI deployments, this means access to computing capacity could increasingly depend on where infrastructure is available and when additional capacity becomes operational.

Microsoft’s AI investment therefore continues at a massive scale, but the latest questions show that the AI race is not simply a race to buy more chips. The ability to turn those chips, buildings and power into usable computing capacity could be just as important.