Building infrastructure for artificial intelligence is projected to cost roughly $10.3 trillion in the United States alone between 2025 and 2032. For that investment to pencil out, the AI sector would need to generate approximately $3.7 trillion in annual revenue by the end of that window.

Those figures come from Columbia Business School professor Stijn Van Nieuwerburgh, whose analysis was presented at the Brookings Papers on Economic Activity conference, a long-running forum where economists present new research on the US economy. The study attempts to answer a question that has been quietly nagging anyone paying attention to the AI buildout: does the math actually work?

The numbers behind the buildout

The core projection calls for adding approximately 182.7 gigawatts of data center capacity by 2032. For context, that represents a staggering amount of power infrastructure, with investment spending roughly equivalent to about 3.63% of US GDP annually.

To put the scale in perspective, the report notes that this buildout would surpass historical infrastructure investments such as the construction of the railroad network and the interstate highway system.

The per-unit economics tell an equally demanding story. The analysis estimates that profitability requires about $5.5 in revenue per installed GPU-hour at full utilization — a benchmark that ties the return on each installed chip directly to what customers actually pay for AI compute. utilization drops to a more realistic 80%, that figure climbs to $6.9 per GPU-hour.

The return threshold used in the model assumes a 10% unlevered return — a measure calculated before the effects of debt financing — with a 50% cash-flow margin. Meeting those benchmarks would require the AI sector to sustain approximately 80% compounded annual revenue growth from today through 2032.

That growth rate is benchmarked against a current combined revenue run-rate of about $100 billion from OpenAI and Anthropic, the developers behind ChatGPT and Claude, respectively. Going from $100 billion to $3.7 trillion in annual revenue over roughly seven years is a trajectory that looks impressive on a pitch deck and daunting on a risk committee's whiteboard.

How it's being financed

The report digs into financing structures, an area of particular relevance for anyone worried about systemic risk. Hyperscalers — the operators of the largest cloud and data center networks, such as Microsoft, Amazon, and Google — are increasingly turning to external debt and complex financial arrangements to fund their AI infrastructure ambitions.

The report flags this interconnectedness as a potential source of systemic risk, particularly if demand for AI services doesn't scale as quickly as capacity is being built. A shortfall in AI demand would then be felt not only by the technology companies themselves but by the lenders and investors on the other side of that debt. Overcapacity in data centers funded by external debt is a very different problem than overcapacity funded by retained earnings.

The $3.7 trillion question

The $3.7 trillion annual revenue target by 2032 would represent roughly 9.2% of projected US GDP for that year.

There are also physical constraints that could slow the buildout regardless of financing availability. Power generation and grid capacity remain bottlenecks in many parts of the country, and permitting timelines for new power plants and transmission lines often stretch far beyond what the AI investment timeline demands.

The report also raises questions about utilization rates. Building 182.7 GW of capacity only generates returns if that capacity is actually used. The difference between the $5.5 and $6.9 per GPU-hour revenue requirements at full versus 80% utilization illustrates how sensitive the economics are to demand fluctuations.

The study frames these figures as a benchmark for assessing whether the sector's revenue trajectory can support the pace of infrastructure investment now underway. Measured against that yardstick, the observable markers going forward are the revenue disclosures of the major AI labs, the utilization of installed GPU capacity, the terms on which hyperscalers raise external debt, and the pace at which new power generation and permitting clear the constraints the report identifies.