The Compute Coalition: How to Build the Future of AI in the Free World
Summary
The paper uses a financial model to demonstrate that 'time to power'—the speed at which data centers become operational—is the primary driver of economic returns for AI infrastructure, outweighing energy costs and tax incentives. It identifies the United States and the UAE as the most competitive locations due to speed, while Germany is the least competitive. The authors argue that because no single democracy can build the necessary infrastructure alone, a coalition of allies should pool geography, energy, and capital to outcompete authoritarian regimes. Recommended domestic reforms include fast-tracking permitting and reforming grid connection queues, while international efforts should focus on securing the AI value chain from critical minerals to deployment.
Key insights
- The speed of getting a data center operational, termed 'time to power', is the most critical factor for economic returns, far exceeding the importance of electricity prices or tax incentives.
- A one-year delay in operation for a typical 100-megawatt U.S. data center is estimated to cost over $500 million in life cycle value, representing more than 5 percent of its total value.
- The United States and the United Arab Emirates are the most competitive countries for AI data centers due to faster project timelines, whereas Germany is the least competitive due to the longest timelines.
- Behind-the-meter (BTM) power, such as on-site natural gas turbines, has been essential for U.S. competitiveness by bypassing grid connection queues, though it is viewed as a temporary solution due to higher costs and environmental concerns.
- The United States currently holds a significant lead in AI compute, hosting approximately 75 percent of the world's high-performance AI compute as of 2025, compared to less than 5 percent for the European Union.
- Authoritarian regimes, particularly China, pose a strategic threat by mobilizing state resources to close the AI compute gap, which could lead to increased global repression and higher risks of catastrophic AI accidents due to a lack of transparency.
- The authors propose a 'democratic compute coalition' to pool resources across allies, leveraging Canada's hydroelectricity, Northern Europe's cool climates, and Australia's solar potential to reduce concentration risk and bypass U.S. domestic constraints.
- To improve competitiveness, governments should prioritize removing obstacles like permitting backlogs and grid connection queues rather than providing energy subsidies or tax cuts.
Cite the original document
- APA
- Winter-Levy, A. P.-R. T. T. S. (2026). The Compute Coalition: How to Build the Future of AI in the Free World. Carnegie Endowment for International Peace. https://carnegieendowment.org/research/2026/06/the-compute-coalition-how-to-build-the-future-of-ai-in-the-free-world
- Chicago
- Winter-Levy, Alasdair Phillips-Robins, Teddy Tawil, Sam. The Compute Coalition: How to Build the Future of AI in the Free World. Carnegie Endowment for International Peace, 2026. https://carnegieendowment.org/research/2026/06/the-compute-coalition-how-to-build-the-future-of-ai-in-the-free-world.
- Wikipedia
- {{cite report |last1=Winter-Levy |first1=Alasdair Phillips-Robins, Teddy Tawil, Sam |title=The Compute Coalition: How to Build the Future of AI in the Free World |publisher=Carnegie Endowment for International Peace |date=8 June 2026 |url=https://carnegieendowment.org/research/2026/06/the-compute-coalition-how-to-build-the-future-of-ai-in-the-free-world |access-date=5 November 2026 |via=Climate Insights Directory (document carnend-000261)}}
- BibTeX
- @techreport{winterlevy2026compute, author = {Winter-Levy, Alasdair Phillips-Robins, Teddy Tawil, Sam}, title = {{The Compute Coalition: How to Build the Future of AI in the Free World}}, institution = {Carnegie Endowment for International Peace}, year = {2026}, month = jun, url = {https://carnegieendowment.org/research/2026/06/the-compute-coalition-how-to-build-the-future-of-ai-in-the-free-world}, urldate = {2026-11-05}, note = {Indexed by Climate Insights Directory, document carnend-000261} }
Full text
Collected · Record updated