A recent study by Columbia Business School professor Stijn Van Nieuwerburgh released through the Brookings Institution estimates that cumulative US investment in data centers, power facilities, networking equipment, GPUs and other AI infrastructure could reach about $10.3 trillion between 2025 and 2032, averaging the equivalent of 3.63% of US GDP per year.
The report notes that by this measure of investment intensity, the current US AI buildout cycle has already surpassed several major historical infrastructure expansions. Railway construction from 1870 to 1890 averaged about 2.24% of GDP annually, highway construction from 1956 to 1973 about 1.13%, and communications and fiber buildout from 1996 to 2003 about 1.10%.
Behind the enormous investment is the rising cost of AI data centers. Using a 200-megawatt AI training campus as an example, the report estimates an initial construction cost of about $8.2 billion, including about $2.2 billion for the data center building, about $400 million for new power facilities, and about $5.6 billion for IT equipment such as GPUs, networking and storage, with roughly two-thirds of the money going into faster-depreciating computing equipment.
By this standard, the cost of a 1-gigawatt AI campus has already reached about $41 billion. The report estimates that by 2032, the US could add about 182.8 gigawatts of data center capacity, with another 117.2 gigawatts of projects coming online after 2032. Construction spending from 2025 to 2032 could reach $10.3 trillion.
As data centers become more expensive to build, the combined operating cash flow of several tech giants is beginning to fall short of capital expenditure. Combined capex at Oracle, Microsoft, Amazon, Meta and Alphabet rose from $96.8 billion in 2020 to $415.8 billion in 2025, while operating cash flow over the same period was $603.2 billion.
The report projects that in 2026 the five companies' capex will rise further to $800.5 billion, while operating cash flow will be about $707.1 billion, meaning capex will exceed operating cash flow for the first time, equal to about 113% of operating cash flow. Goldman Sachs also expects US AI-related investment as a share of GDP to rise from 1.8% in 2026 to 2.5% in 2027 and further to 2.8% in 2028.
US AI giants are increasingly relying on borrowing and external financing for funding. The report cites Morgan Stanley estimates that from 2025 to 2028, about $2.9 trillion will be needed to meet large cloud providers' additional computing demand, with more than half of that relying on external capital. Across all financing, the equity-to-debt ratio is about 60-40, with private credit alone potentially reaching about $800 billion, corporate bonds about $200 billion, and structured financing about $150 billion.
Van Nieuwerburgh uses Meta's Hyperion AI data center in Louisiana as an example. The project analyzed in the report is about 2 gigawatts in scale, corresponding to about $30 billion in assets. After Meta sold 80% of the project equity to US alternative asset manager Blue Owl, the joint venture entity raised another $27 billion in debt, bringing the project-level debt-to-asset ratio to nearly 90%.
The Hyperion debt was issued at a yield of 6.58%, at least 100 basis points higher than similar corporate bonds issued directly by Meta at the same time. The report estimates that this spread alone will add more than $5 billion in interest costs over the entire financing cycle, ultimately reflected in the rent Meta pays. Similar long-term payment obligations have already appeared at more large tech companies.
US credit rating agency Moody's estimates that large cloud providers have already taken on about $970 billion in lease commitments, of which about $660 billion are future leases not yet on the balance sheet. The report cites Wall Street Journal statistics showing that for Microsoft, Alphabet, Amazon and Meta alone, future off-balance-sheet obligations such as leases and purchases have already reached about $2.4 trillion, including $904 billion in future lease commitments and $1.52 trillion in future purchase commitments, the latter mainly involving chips.
However, AI data centers themselves are a type of asset whose core equipment depreciates quickly. Traditional infrastructure such as railways can be used for decades, while most of the money in today's data centers goes into GPUs, servers, storage and high-speed networking equipment. The report estimates that of the roughly $8.2 billion investment in a 200-megawatt AI data center, about 68% is for IT assets such as GPUs, servers and networking equipment, which are calculated on a 6-year economic life; the remaining assets such as buildings, power and cooling facilities are calculated on a 20-year basis.
If the economic life of IT assets shortens from 6 years to 3 years, all else equal, the annual capital recovery ratio would rise from 19.37% to 31.10%, and the corresponding mature-stage annual revenue requirement would increase from about $3.7 trillion to about $6 trillion, equivalent to 14.8% of projected 2032 US GDP.
Hyperion also reflects the contradiction between long-term financing and rapid technological iteration. Its project debt maturity extends to 2049, while Meta will use the campus starting in 2029 through a series of four-year leases and bear corresponding residual value guarantees. At the same time, the GPUs, cooling systems and computing architecture supporting data center demand may change significantly within just a few years.
Creditors look at the project's future long-term rent and residual asset value, but those revenues ultimately still depend on whether Meta continues to need such massive computing power. Once technological iteration or changes in AI demand cause data center utilization and asset values to fall below expectations, pressure could still be transmitted back to tech giants through rent, guarantees and refinancing costs. The report notes that this mismatch between long-term financing maturities and rapid underlying technological iteration is one of the important risks facing AI infrastructure.
Electricity is another layer of obstruction. The report estimates that a 200-megawatt AI data center running continuously at full load consumes roughly as much electricity in a year as 170,000 US households. Based on the projected 2032 construction scale in the report, the electricity consumption of new AI data centers could be equivalent to the current electricity use of the entire US residential sector. Many projects also need to simultaneously secure substations, high-voltage transmission lines, grid connection capacity and new generation facilities.
The report notes that grid connection may take years, while projects also face rising power costs, local opposition and insufficient supply of GPUs, high-bandwidth memory and networking equipment. The report estimates that of about 509 gigawatts of planned data center projects in the US, about 226.9 gigawatts will ultimately not be built, while another 117.2 gigawatts will not come into operation until after 2032.
Beyond construction obstacles, the bigger question is whether these facilities can earn back such enormous investment once built. Based on 182.7 gigawatts of new capacity and about $9.61 trillion in capital put into operation, the report estimates that if investors require a 10% unlevered return and the operating cash flow rate is 50%, these data centers will need to generate about $3.725 trillion in annual revenue once mature, equivalent to 9.2% of projected 2032 US GDP.
Converted into computing prices, at 100% utilization, each GB300 GPU would need to generate about $5.5 per hour in revenue; if equipment utilization is only 80%, that figure rises to about $6.9, and if utilization falls to 70%, it would require about $7.9. The report notes that current on-demand or short-term rental prices for high-end Nvidia GPUs are usually about $6 to more than $10 per GPU hour, meaning that at current market prices, the above revenue requirements are still within a achievable range.
The real question is that by 2032 the market may add about 180 gigawatts of computing power, and whether demand can grow at a similar pace. As cloud providers, model companies and computing suppliers all expand at the same time, once new data center capacity grows faster than actual demand, equipment utilization and GPU rental prices may both decline, and actual revenue will also fall below initial investment estimates. For data centers built heavily on debt, high leverage could also translate into greater equity losses and debt repayment pressure.
Absorbing such a massive increase in computing power itself requires AI industry revenue to maintain extremely high growth. The report cites current estimates that OpenAI and Anthropic together generate about $100 billion in annual revenue. If that revenue scale is used as a reference, reaching $3.725 trillion by 2032 would require maintaining a compound annual growth rate of about 80% over the coming years.
The report also notes that as the scale of AI infrastructure investment expands, chipmakers, cloud providers, data center operators and financial institutions are forming a tighter capital loop. Take Nvidia as an example: it sells GPUs to cloud providers and data centers while also helping customers obtain financing through residual value guarantees and other means, lowering the funding threshold for large-scale chip purchases. Cloud providers continue building data centers and selling computing power to model companies, while data center investors rely on long-term leases and purchase commitments from large tech companies to support debt.
This structure can continue to operate only if new capital keeps entering while AI demand, computing utilization and rental prices also keep growing. But the future may not always go as these companies wish. Once demand growth cannot keep up with the release of data center and GPU capacity, chip prices, computing rental prices and equipment utilization may all decline, while already signed debt, rent and purchase commitments still must continue to be paid.
Comments