Brookings Institution report estimates $10.3 trillion in US AI infrastructure investment over 8 years

[[**U.S. AI infrastructure investment may reach $10.3 trillion over the next 8 years, far exceeding the scale of historical infrastructure, but demand growth is unlikely to match capacity expansion; combined with high leverage and rapid technological iteration, there is a risk of a "left foot stepping on right foot" style bubble burst.**

**Key points**

**1. Investment intensity hits a record high**

Annual investment accounts for 3.63% of GDP, far exceeding railways (2.24%), highways (1.13%), and fiber-optic construction (1.10%); a single 1-gigawatt AI campus costs about $41 billion, with IT equipment such as GPUs accounting for two-thirds of investment.

**2. Capital expenditure has exceeded operating cash flow**

In 2026, capital expenditure by the five major tech companies is expected to reach $800.5 billion, exceeding operating cash flow ($707.1 billion) for the first time; the gap relies on borrowing and external financing, and private credit may reach $800 billion.

**3. High leverage and long-term leases amplify risk**

The debt-to-asset ratio of the Meta Hyperion project is close to 90%, and financing costs are 100 basis points higher than corporate bonds; off-balance-sheet leases and purchase commitments of large cloud vendors have reached $2.4 trillion, but the economic life of core assets such as GPUs is only 6 years, and technological iteration may cause rapid asset depreciation.

**4. Power and grid bottlenecks are significant**

A 200-megawatt data center consumes as much electricity annually as 170,000 households; by 2032, the additional electricity consumption of new AI data centers may equal that of all U.S. residential electricity use, and about 227 gigawatts of planned projects may ultimately fail to be built.

**5. A huge demand gap may lead to a vicious cycle**

This article is from the WeChat public account: Daju Caijing, author: Tang Puji

Recently, the Brookings Institution released the latest research by Columbia Business School professor Stijn Van Nieuwerburgh. In the report, he estimates that from 2025 to 2032, cumulative investment in U.S. data centers, power facilities, network equipment, GPUs, and other AI infrastructure may reach about $10.3 trillion (about 69.1 trillion yuan), averaging 3.63% of U.S. GDP per year.

The report points out that at this investment intensity, the current U.S. AI construction cycle has exceeded multiple rounds of large-scale infrastructure expansion in history. Railway construction from 1870 to 1890 averaged about 2.24% of GDP per year, highway construction from 1956 to 1973 was about 1.13%, and communications and fiber-optic construction from 1996 to 2003 was about 1.10%.

The report takes a 200-megawatt AI training campus as an example. Its initial construction cost is about $8.2 billion (about 55.1 billion yuan), of which data center buildings are about $2.2 billion (about 14.8 billion yuan), new power facilities are about $400 million (about 2.7 billion yuan), and IT equipment such as GPUs, networking, and storage is about $5.6 billion (about 37.7 billion yuan). About two-thirds of the funds are invested in computing equipment that is updated faster.

By this standard, the cost of a 1-gigawatt AI campus has reached about $41 billion (about 275.2 billion yuan). The report estimates that by 2032, the United States may add about 182.8 gigawatts of data center capacity, with another 117.2 gigawatts of projects put into use after 2032. Construction spending from 2025 to 2032 may reach $10.3 trillion (about 69.15 trillion yuan).

Data centers are becoming more expensive to build, and the combined operating cash flow of several tech giants is beginning to fall short of capital expenditure.

The combined capital expenditure of Oracle, Microsoft, Amazon, Meta, and Alphabet rose from $96.8 billion (about 650 billion yuan) in 2020 to $415.8 billion (about 2.79 trillion yuan) in 2025, while operating cash flow in the same period was $603.2 billion (about 4.05 trillion yuan). The report expects that in 2026, the five companies' capital expenditure will further increase to $800.5 billion (about 5.37 trillion yuan), while operating cash flow will be about $707.1 billion (about 4.75 trillion yuan). Capital expenditure will exceed operating cash flow for the first time, equivalent to about 113% of operating cash flow.

Changes in capital expenditure and operating cash flow of the five major U.S. tech companies Brookings

Goldman Sachs also expects that the share of U.S. AI-related investment in GDP will rise from 1.8% in 2026 to 2.5% in 2027 and further to 2.8% in 2028.

The report cites Morgan Stanley's estimate that from 2025 to 2028, about $2.9 trillion (about 19.5 trillion yuan) will be needed to meet the additional computing power demand of large cloud vendors, of which more than half will need to rely on external capital. Of all financing, the ratio of equity to debt is about 60:40; private credit alone may reach about $800 billion (about 5.37 trillion yuan), corporate bonds about $200 billion (about 1.34 trillion yuan), and structured financing about $150 billion (about 1.01 trillion yuan).

Van Nieuwerburgh takes the Hyperion AI data center built by Meta in Louisiana, USA, as an example. The project analyzed in the report is about 2 gigawatts, corresponding to about $30 billion (about 201.4 billion yuan) in assets. After Meta sold 80% of the project equity to U.S. alternative asset manager Blue Owl, the joint venture entity raised another $27 billion (about 181.3 billion yuan) in debt, bringing the project-level debt-to-asset ratio to nearly 90%.

Meta Hyperion data center financing structure Brookings

The yield on Hyperion's debt issuance reached 6.58%, at least 100 basis points higher than Meta's direct issuance of similar corporate bonds over the same period. The report estimates that this spread alone will add more than $5 billion (about 33.6 billion yuan) in interest costs over the entire financing cycle, and will ultimately be reflected in the rent Meta pays.

Similar long-term payment obligations have already appeared at more large technology companies. U.S. credit rating agency Moody's estimates that large cloud providers have already taken on about $970 billion (about 6.51 trillion yuan) in lease commitments, of which about $660 billion (about 4.43 trillion yuan) consists of future leases that have not yet entered the balance sheet.

The report cites Wall Street Journal statistics as saying that for just four companies—Microsoft, Alphabet, Amazon and Meta—future off-balance-sheet obligations such as leases and purchases have already reached about $2.4 trillion (about 16.1 trillion yuan), including $904 billion (about 6.07 trillion yuan) in future lease commitments and $1.52 trillion (about 10.2 trillion yuan) 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, whereas most of the money in today's data centers goes into GPUs, servers, storage and high-speed networking equipment.

The report calculates that of the roughly $8.2 billion investment in a 200-megawatt AI data center, about 68% goes to 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.

68% of investment goes to IT equipment such as GPUs, networking and storage Brookings

If the economic life of IT assets is shortened from 6 years to 3 years, all else being equal, the proportion of capital that needs to be recovered each year will rise from 19.37% to 31.10%, and the corresponding mature-stage annual revenue requirement will also increase from about $3.7 trillion (about 24.8 trillion yuan) to about $6 trillion (about 40.3 trillion yuan), equivalent to 14.8% of U.S. GDP in 2032.

Hyperion also embodies the contradiction between long-term financing and rapid technological iteration. Its project debt maturity extends all the way to 2049, while Meta will begin using the campus in 2029 through a series of four-year leases and bears the corresponding asset residual value guarantee. At the same time, the GPUs, cooling systems and computing architectures that support data center demand may change significantly within a few years.

Creditors look at the project's future long-term rent and asset residual value, but these 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, the pressure may still be transmitted back to the technology giants through rent, guarantees and refinancing costs.

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 U.S. households. According to the construction scale projected in the report for 2032, the electricity consumption of new AI data centers could be equivalent to the current electricity consumption of the entire U.S. residential sector.

A large number of projects also need to simultaneously obtain substations, high-voltage transmission lines, grid connection capacity and new power generation facilities. The report points out that grid connection may take years, while projects also face constraints such as rising electricity costs, local opposition and insufficient supply of GPUs, high-bandwidth memory and networking equipment.

The report calculates that of the roughly 509 gigawatts of data center projects currently planned in the United States, about 226.9 gigawatts will ultimately not be built, and another 117.2 gigawatts will not be able to enter operation until after 2032.

Beyond construction obstacles, the bigger question is whether these facilities, once built, can earn back such massive investment.

Based on 182.7 gigawatts of new capacity and about $9.61 trillion (about 64.5 trillion yuan) of capital put into operation, the report calculates that if investors require a 10% unlevered return while operating cash flow reaches 50%, these data centers will need to generate about $3.725 trillion (about 25 trillion yuan) in revenue each year once mature, equivalent to 9.2% of U.S. GDP in 2032 as projected in the report.

Converted into computing power prices, at 100% utilization, each GB300 GPU needs to generate about $5.5 (about 37 yuan) in revenue per hour; if equipment utilization is only 80%, this figure rises to about $6.9 (about 46 yuan), and when utilization falls to 70%, about $7.9 (about 53 yuan) is needed.

The report points out 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 (about 40 to 67 yuan), and at current market prices, the above revenue requirements are still within an achievable range. The real problem 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 power suppliers expand simultaneously, 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 the original investment calculations. For those data centers built heavily with debt, high leverage may also translate into larger equity losses and debt repayment pressure.

Absorbing such massive new computing power itself requires AI industry revenue to maintain extremely high growth. The report cites current estimates as saying that OpenAI and Anthropic together have annual revenue of about $100 billion (about 671.3 billion yuan). If this revenue scale is used as a reference, reaching $3.725 trillion in 2032 would require maintaining a compound annual growth rate of about 80% over the coming years.

Take Nvidia as an example: it sells GPUs to cloud providers and data centers on the one hand, and on the other hand helps customers obtain financing through residual value guarantees and other means, lowering the capital threshold for large-scale chip procurement. Cloud providers continue to build data centers and sell computing power to model companies, while data center investors rely on long-term leases and purchase commitments from large technology companies to support debt.

Once demand growth cannot keep up with the release of data center and GPU capacity, chip prices, computing power rental prices and equipment utilization may all decline, while the debt, rent and purchase commitments already signed still need to be paid. AI industry signal channel: Frontier Technology This content comes from the internet and the original link; the views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]