Microsoft's AI ambitions are facing a critical challenge: a shortage of chips. This issue is not just a technical hurdle but a strategic one, with far-reaching implications for the company's future in the AI race. The question of whether Microsoft's AI plans are being held back by a lack of chips is more complex than it initially seems, and the answer lies in the intricate relationship between the company, its suppliers, and the broader market dynamics.
The Chip Shortage Conundrum
On the surface, the discrepancy between Microsoft's public claims and the actual number of AI chips in operation is striking. The company targeted 1.8 million AI chips by the end of 2024, but internal documents reveal a much lower figure of 2.2 million. This shortfall is not just a minor issue; it raises questions about the company's ability to execute its ambitious AI strategy. The AI arms race is intensifying, and Microsoft's slow build-out of AI capacity could put it at a significant disadvantage.
One of the key challenges is the supply chain for AI chips, which is dominated by Nvidia, one of the world's most valuable companies. Nvidia's supply chain is a closely guarded secret, and the company does not disclose the number of chips it sells or to whom. This lack of transparency makes it difficult to assess the true state of the AI chip market and the impact of the shortage on Microsoft's plans.
Microsoft's AI Build-Out
Microsoft has been investing heavily in AI infrastructure, with CEO Satya Nadella stating that the company would double its global datacentre footprint by mid-2027. The company has ploughed roughly $280 billion into land, buildings, and computational infrastructure since 2022, with over $41 billion invested in the past quarter alone. However, estimating the exact number of datacentres built with this money is challenging.
One way to assess Microsoft's progress is by looking at its public announcements and energy consumption. The company claims to have added 5 GW of datacentre capacity over the past two years, which is a staggering amount of energy. This capacity is four times the size of the largest datacentre park in Europe, and it suggests Microsoft has hundreds of datacentres on five continents.
However, Microsoft's own sustainability reports indicate that its AI capacity in 2024 was likely closer to 1.2 GW, which would require roughly 4 million AI chips if it added 5 GW of AI datacentres in the past two years. This discrepancy raises questions about the accuracy of Microsoft's public statements and the true state of its AI infrastructure.
The Role of Nvidia
Nvidia's dominance in the AI chip market is a critical factor in Microsoft's struggle. The company's public statements about chip orders and sales are often vague, and it does not provide detailed breakdowns of its customer base. This lack of transparency makes it difficult to assess the true demand for AI chips and the impact of the shortage on Microsoft's plans.
Microsoft's tie-up with OpenAI may also be a factor in the apparent discrepancy. The exact terms of their commercial partnership are not public, but this unit may account for some of Microsoft's datacentre deployments, which would not be reflected in the internal documents seen by the Guardian.
The Fairwater Project
Microsoft's largest AI development in the US, the Fairwater project in Wisconsin and Georgia, is a case in point. In April, Nadella announced that the project was 'going live', but satellite footage and internal documents indicate that only part of the facility is operational. This is a common issue in large-scale AI projects, where initial investments are massive, and it takes time to bring capacity online.
The Fairwater project also highlights Microsoft's reliance on Nvidia's newest chip model, the Blackwell. Nvidia's CEO, Jensen Huang, announced orders for Blackwells from the top four customers, including Microsoft, amounting to 3.6 million chips. However, Microsoft has installed fewer than half of this amount, which raises questions about the true demand for Blackwells and the impact of the shortage on the company's plans.
The Calculation Conundrum
Calculating the number of AI chips from a company's 'AI capacity' in terms of power is a complex task. The Guardian used a methodology that divides the power usage of a datacentre by the power usage of an AI chip, such as the H100, to get a broad approximation. This suggests Microsoft should have 12 million chips with 10 GW of capacity.
However, this approximation does not account for several factors, including the electricity used by cooling systems and other equipment, and the fact that not all chips in a datacentre are AI chips. Microsoft's own sustainability reports indicate that 89% of the electricity in its new datacentres powers the IT systems, with an 11% overhead.
Taking these factors into account, the estimate of 12 million chips is likely an overestimation. A more conservative estimate, based on the electricity used by computer chips, suggests Microsoft may have 6.4 million chips. This is a significant figure, but it still falls short of the company's public claims.
The Way Forward
Microsoft's AI ambitions are facing a critical challenge, and the shortage of chips is a significant factor in this struggle. The company's public statements about its AI capacity and chip orders are often vague, and the true state of its infrastructure is difficult to assess. The relationship between Microsoft, Nvidia, and the broader market dynamics is complex, and it is unclear how the company will navigate this challenge.
In my opinion, Microsoft's AI plans are being held back by a combination of factors, including the shortage of chips, the dominance of Nvidia, and the complexity of calculating the true number of AI chips. The company's public statements are often misleading, and the true state of its infrastructure is difficult to assess. As an expert, I believe that Microsoft needs to be more transparent about its AI strategy and the true state of its infrastructure if it is to succeed in the AI race.
