Microsoft's AI ambitions may be hindered by a shortage of chips, despite the company's claims of rapid expansion. The Guardian's investigation reveals a discrepancy between Microsoft's public statements and the actual number of advanced AI chips in operation. The company targeted 1.8 million AI chips by 2024, but internal documents show only 2.2 million installed, a shortfall that raises questions about the true state of its AI infrastructure.
This issue is significant because the AI arms race demands a massive build-out of datacentres, which rely on expensive chips from Nvidia, one of the world's most valuable companies. The supply chain of these chips is a closely guarded secret, with Nvidia not disclosing sales figures or client information. This lack of transparency makes it challenging to assess the true scale of AI development.
Microsoft's own claims of adding 5GW of datacentre capacity over two years and having hundreds of datacentres on five continents seem to be at odds with the number of AI chips in operation. Professor Shaolei Ren suggests that Microsoft's sustainability reports provide a more accurate picture, indicating a lower AI capacity of 1.2GW. This would imply a need for around 4 million AI chips, which is significantly less than the targeted 1.8 million.
The discrepancy highlights the complexity of assessing AI progress. Microsoft's build-out of AI capacity may be slower than its annual reports suggest, and some of its large projects, like the Fairwater datacentres, are far from operational. This could be due to challenges in securing electrical power and building datacentres close to power sources.
Nvidia's balance sheets show significant chip sales, but it's unclear if Microsoft has purchased and installed them. Microsoft's CEO, Satya Nadella, acknowledged the issue of having chips without the necessary infrastructure to use them. The company's spokesperson refutes the Guardian's calculations, but the true number of AI chips in Microsoft's possession remains uncertain.
Calculating the number of chips from AI capacity figures is complex. Dividing power usage by chip power consumption provides a broad approximation, but factors like cooling systems and non-AI chips on server racks must be considered. Microsoft's own efficiency estimates suggest a lower usable capacity, further complicating the calculation.
