Microsoft's AI ambitions may be facing a significant hurdle: a shortage of chips. The company's plans to install 1.8 million AI chips by the end of 2024 seem to have hit a snag, as internal documents reveal only 2.2 million chips in operation. This discrepancy raises questions about the true state of Microsoft's AI infrastructure and the potential impact on its ambitious expansion plans.
The issue lies in the closely guarded nature of Nvidia's chip supply chain, which is a critical component in the AI arms race. Nvidia, one of the world's most valuable companies, does not publicly disclose its chip sales or clients, making it challenging to assess the true scale of AI development. This lack of transparency extends to Microsoft, which has been investing heavily in AI infrastructure.
Microsoft's public statements and annual reports suggest an impressive growth trajectory, with claims of adding 5GW of datacentre capacity over two years and hundreds of datacentres on five continents. However, these figures may be misleading. Professor Shaolei Ren highlights a discrepancy between Microsoft's sustainability reports and its financial statements, indicating a potential underestimation of its AI capacity.
The company's largest AI development, the Fairwater project, is also shrouded in uncertainty. Despite CEO Satya Nadella's claims, satellite footage and internal documents suggest only partial operation, with a significant delay in bringing the project online.
The shortage of chips is further exacerbated by Microsoft's partnership with OpenAI, which may account for some datacentre deployments not reflected in the available documents. The exact terms of this partnership are unknown, adding another layer of complexity to the situation.
Nvidia's CEO, Jensen Huang, revealed orders for Blackwell GPUs from top customers, including Microsoft, amounted to 3.6 million. Given Microsoft's historical status as a major Nvidia customer, this suggests a potential shortage of Blackwell chips, which are crucial for AI development.
Microsoft's spokesperson acknowledges the challenge of matching power capacity with chip availability, stating that the company's datacentres use custom silicon, AMD, Intel, and Nvidia chips. The spokesperson refutes the Guardian's calculations, but the underlying issue remains: the company's AI capacity expansion may be slower than projected.
The calculation of chip numbers from AI capacity is a complex process. Dividing power usage by AI chip power consumption provides a broad approximation, but factors like cooling systems and server configurations must be considered. Ren estimates that only 80% of electricity in a datacentre goes to computer chips, and oversubscription of power capacity is common.
In conclusion, Microsoft's AI plans may be hindered by a shortage of chips, as indicated by the discrepancy in chip installations. The company's reliance on Nvidia's supply chain and the lack of transparency in chip sales make it challenging to assess the true state of AI development. As Microsoft continues to invest in AI infrastructure, addressing this chip shortage is crucial to achieving its ambitious goals.
