Microsoft's AI ambitions are facing a significant hurdle: a shortage of chips. The company's plans to install 1.8 million AI chips by the end of 2024 have been met with a reality check, as internal documents reveal a shortfall of nearly half that target. This discrepancy highlights a deeper issue: the complex interplay between AI development, chip manufacturing, and the global supply chain.
The chips in question are small but powerful, made by Nvidia, one of the world's most valuable companies. Nvidia's supply chain is a closely guarded secret, and the company doesn't disclose sales figures or client information. This lack of transparency makes it difficult to gauge the true state of the AI arms race.
Microsoft, a key player in this race, has been investing heavily in AI infrastructure. Since 2022, the company has poured over $280 billion into land, buildings, and computational infrastructure, with a focus on AI. However, estimating the exact number of datacentres built and their operational status is challenging.
Microsoft's claims of adding 5GW of datacentre capacity over two years and having hundreds of datacentres on five continents are intriguing. But these figures don't tell the whole story. The company's sustainability reports suggest a lower AI capacity, and experts question the accuracy of Microsoft's announcements.
The Fairwater project in Wisconsin, Microsoft's largest AI development, is a case in point. Initially a multi-gigawatt, multibillion-dollar investment, it has only partially operational three years later. This slow rollout raises questions about the availability of chips and the challenges of integrating them into the infrastructure.
Nvidia's balance sheets show impressive chip sales, but the question remains: where are the chips Microsoft needs? The company's CEO, Satya Nadella, acknowledged the issue of electrical power and the need for warm shells to plug in chips, rather than a shortage of chips themselves.
Calculating the number of chips from AI capacity figures is complex. Microsoft's internal documents indicate a mix of H100, A100, and Blackwell chips. A conservative estimate suggests 6.4 million chips for 10GW of AI datacentre capacity. However, this doesn't account for cooling systems and other equipment that also consume electricity.
The AI arms race is a multifaceted challenge, with supply chain secrecy, infrastructure development, and the intricate relationship between chips and power. As Microsoft navigates these complexities, the company's AI plans may be held back not by a shortage of ambition but by the intricate logistics of chip procurement and integration.
