AI Infrastructure Sovereignty Analyzed Through a Three-Layer Framework

The paper 'AI Compute Sovereignty', published by researchers from the University of Oxford and Aalto University, does not view AI infrastructure sovereignty as a simple binary, but rather divides it into three layers.

• Where: Where is the compute physically located? (Territorial control)
• Who: Who operates the cloud/provider? (Ownership)
• What: Which accelerators are supplied? (Hardware control)

A survey of nine major public cloud providers found that there are 225 cloud regions across 43 countries, of which 33 countries have accelerator-enabled regions. However, only 24 countries have training-related compute.

In the case of India, 3 out of 5 accelerator-enabled regions had training-related compute, but 4 of these were operated by US providers and 1 by a Chinese provider. This demonstrates a 'Hedging' phenomenon where domestic infrastructure relies on foreign providers.

Additionally, 95.5% of the surveyed accelerator-enabled regions use US-owned accelerators, revealing that even if the physical location is domestic, dependency on the underlying hardware stack remains high. The paper does not argue that all countries must build a fully self-contained AI stack, but recommends strategically selecting 'the parts of the AI stack that actually need to be controlled' while considering energy, water, land costs, and infrastructure construction costs.