Microsoft's ambitious expansion of its artificial intelligence infrastructure may be hampered by a significant shortfall in the number of advanced chips deployed, according to a Guardian investigation. Internal documents indicate the company has installed 2.2 million AI chips, a figure that appears to be considerably lower than what its stated data center capacity would suggest. This apparent gap raises questions about the pace of its AI development and the operational readiness of its new facilities. The investigation highlights a discrepancy between Microsoft's public pronouncements on its AI capabilities and the actual number of high-performance chips powering its operations. While Microsoft reportedly aimed to have 1.8 million AI chips in place by the end of 2024, nearly two years later, the installed base has reached 2.2 million. This figure, however, is less than half of what some experts had estimated based on the company's announced data center expansion. One projection suggested that 10 gigawatts of AI data center capacity, which Microsoft has stated it is adding, would require approximately 6.4 million graphics processing units (GPUs). Even a more conservative estimate of 1.2 gigawatts of AI capacity in 2024, with an additional 5 gigawatts added since, would necessitate around 4 million chips. Microsoft has stated that the calculations in the investigation are based on incorrect information, but the company has not specified which figures are inaccurate. Some analysts within the industry have noted that the chip numbers are lower than anticipated from Microsoft. Sources within the company suggest that the total number of AI chips has seen little change over the past year. The global race to develop and deploy artificial intelligence necessitates a substantial build-out of data centers, which in turn require vast quantities of expensive, specialized chips, primarily from Nvidia. The opacity surrounding chip supply chains, as Nvidia does not publicly disclose sales figures to specific clients, makes it difficult for external observers to gauge the true scale of AI development. A portion of the discrepancy might be attributed to Microsoft's partnership with OpenAI. The precise financial and operational terms of this collaboration are not public, and it is possible that some of Microsoft's data center deployments related to OpenAI are not reflected in the documents reviewed by The Guardian. Microsoft has also been developing its own AI silicon, such as the Maia 200 AI accelerator introduced in early 2026, with the Maia 300 expected later in the year. The company is reportedly in discussions with TSMC to secure manufacturing capacity for over 300,000 next-generation Maia chips by 2027, with long-term targets potentially exceeding one million units. This strategic move aims to reduce reliance on third-party GPUs and optimize data center infrastructure. However, Microsoft's in-house chip production is seen as lagging behind competitors like Amazon and Google. The broader semiconductor industry continues to face supply chain constraints. The CoWoS packaging capacity at TSMC, essential for assembling advanced AI chips, is fully allocated through at least mid-2027. High-bandwidth memory (HBM) production also lags behind demand, creating bottlenecks that affect overall GPU output. These constraints contribute to increased rental costs for AI chips, with H100 GPU rental prices reportedly rising significantly in recent months. Furthermore, the demand from AI data centers is projected to consume a substantial portion of memory chip production in the coming years, potentially impacting availability and pricing across various technology sectors.