According to Columbia Business School, building infrastructure for artificial intelligence will cost the U.S. approximately $10.3 trillion from 2025 to 2032. For the investments to pay off, the AI sector needs to generate about $3.7 trillion in annual revenue by the end of this period. These figures are presented in an analysis by Professor Stein van Nieuwerburgh of Columbia Business School, shared at the Brookings Papers on Economic Activity conference.
The forecast suggests adding approximately 182.7 GW of data center capacity by 2032. This corresponds to about 3.63% of U.S. GDP in the form of annual investments. In scale, the project will surpass historical infrastructure builds, including the railroad network and the interstate highway system. Profitability requires about $5.5 in revenue per installed GPU-hour at full load. If the load drops to 80%, the figure rises to $6.9 per GPU-hour. The model assumes a 10% unleveraged return with a 50% cash flow margin. To achieve this, the sector needs to maintain about 80% cumulative annual revenue growth until 2032.
This growth is measured from the current combined revenue of OpenAI and Anthropic – around $100 billion. The transition from $100 billion to $3.7 trillion in annual revenue in about seven years is a trajectory that impresses in presentations and raises concerns in risk councils. Hyperscalers, including Microsoft, Amazon, and Google, are increasingly resorting to external debt and complex financial schemes. The report highlights this interconnectedness as a potential source of systemic risk, especially if demand does not grow as quickly as capacity.
Related articles
