Page 30 - The Annual AI Governance Report 2025 Steering the Future of AI

The Annual AI Governance Report 2025: Steering the Future of AI

Theme 5: AI Infrastructure and Compute

5.1 Global Compute Distribution Infrastructure

Compute Needs: Access to computing power and storage (“compute” for short) —alongside Theme 5: AI
data and skilled talent—is a crucial ingredient in building artificial intelligence systems. Scaling
laws in AI, which describe how the performance of AI systems improves as the size of training
data and computational resources improve, has been a dominant force in the field (although
critics caution that diminishing returns may have been reached); these scaling laws have driven
further development of compute resources. This access is heavily concentrated, with one or a
few firms dominating critical points in the supply chain. Control over compute has become the
central factor concentrating power in AI and raising barriers to entry. 106

Compute Distribution: According to an analysis of researchers at the University of Oxford ,
107
only 32 countries host data centers that provide the compute power essential for advanced
AI development. The United States, China , and the European Union account for over half of
108
the world’s most powerful data centers. American and Chinese companies operate more than
90 percent of the data centers used globally by other organizations for AI work. Africa and South
America have almost no computing hubs. India has at least five; Japan at least four. More than
150 countries have none at all.

Implication for AI Development Equity: AI development is highly spatially uneven, with a small
number of global hubs capturing most of the economic gains while other regions face growing
disparities. Effective policies must ensure broader distribution of AI's benefits to prevent
deepening regional inequality and economic polarization. 109

Advances in both hardware (e.g., GPUs doubling in price-performance every ~2 years) and
algorithms (e.g., image classification compute needs halving every 9 months) have drastically
reduced the cost of training powerful AI models. As compute becomes more efficient, more
actors can train models to a given performance level, and those with existing resources can
train even more powerful models, maintaining a frontier advantage. Large compute investors
110
(like major tech firms) discover new AI capabilities first and maintain a lead, especially in high-
performance. Smaller actors benefit from diffusion but face barriers due to economies of scale,
proprietary technologies, and strategic integration. 111

5.2 Energy and Sustainability

Calculations on energy use of generative AI vary on two axes of energy demand: (1) training,
the amount of energy necessary to host large training runs for models, and (2) inference,

106 Vipra, J. (2025, April 23). Computational power and AI. AI Now Institute.
107 Hawkins, Zoe and Lehdonvirta, Vili and Wu, Boxi, AI Compute Sovereignty: Infrastructure Control Across
Territories, Cloud Providers, and Accelerators (June 20, 2025). Available at SSRN: https:// ssrn .com/ abstract
= 5312977
108 Lewis, J. A. (2024). An overview of global cloud competition. Centre for Strategy and International Studies.
109 https:// www .researchgate .net/ profile/ Uchechukwu -Ajuzieogu/ publication/ 391430918 _The _Economic
_Geography _of _AI _Spatial _Distribution _of _Benefits _and _Costs/ links/ 68170 04960241d5 140226eed/ The
-Economic -Geography -of -AI -Spatial -Distribution -of -Benefits -and -Costs .pdf
110 Pilz, K. F., Heim, L., & Brown, N. (2025). Increased compute efficiency and the diffusion of AI capabilities.
Proceedings of the AAAI Conference on Artificial Intelligence, 39(26), 27582–27590.
111 Ahmed, N., & Wahed, M. (2020, October 22). The de-democratization of AI: deep learning and the compute
divide in artificial intelligence research. arXiv.org.

21