A new Brookings conference paper puts a striking figure on a possible U.S. AI infrastructure cycle: $10.3 trillion of investment during 2025–2032. That number is a modeled scenario built from project capacity and cost assumptions, not money already spent, contracted or guaranteed to be spent. Its importance for investors lies less in the exact total than in a related question: if data centres and chips are financed through more layers of outside capital, who bears the loss when demand, power delivery or equipment values disappoint?
The Brookings summary highlights a shift from direct corporate funding toward joint ventures, leases, private credit and securitisation. Independent Axios reporting also focused on the appetite for external capital and the difficulty of tracking risk across structures. Neither account turns the scenario into a firm spending forecast or proves that a credit crisis is under way.
The headline is a modeled buildout, not an invoice
The paper starts with a project-level U.S. data-centre pipeline and assigns different completion probabilities to proposed capacity. Its central case has about 183 gigawatts of additional capacity operational by 2032, with further projects completed later and many proposed projects never completed. It then applies a representative campus cost, assumed annual cost growth and spending spread over construction years. The resulting $10.3 trillion includes money spent before 2032 on projects that would finish afterward. The paper explicitly calls this a scenario rather than a forecast.
That distinction changes the interpretation. A planned site can be delayed, scaled back or cancelled; a cost benchmark can change with chip prices, design and power infrastructure. Brookings assumes annual nominal GDP growth when comparing the modeled investment flow with the economy. Its headline scale is therefore useful for stress testing capital needs, not a tally of irrevocable corporate commitments. Treating it as a precise future invoice would conceal the model choices doing the work.
The author also compares the modeled annual investment share with earlier U.S. infrastructure waves. Such a comparison indicates the possible macroeconomic weight of the buildout under those assumptions. It does not show that every technology company will earn an attractive return, or that the spending schedule will materialise as drawn.
Hardware is the large, fast-aging slice
In the paper's example, a 200-megawatt AI campus costs about $8.2 billion to build and equip. Roughly two-thirds is computing hardware and other IT systems; the remaining third covers the facility and additional power infrastructure. This is an illustrative cost model, not an audited average for every campus. It matters because GPUs can lose economic value much faster than land or a shell building if technology improves or buyers favour different systems.
The physical constraint cannot be ignored either. The International Energy Agency notes that a data centre may be built in a few years while power systems often require longer planning and construction. Its global electricity-demand work presents several scenarios because adoption, efficiency and energy bottlenecks are uncertain. Those global power scenarios are separate from Brookings' U.S. investment model, but they reinforce the mechanism: a campus cannot earn from installed compute that cannot be connected and powered on schedule.
For chip suppliers, developers and utilities, a large buildout could create real demand. For the capital provider, however, value depends on delivered capacity, utilisation, pricing and the residual worth of equipment. Higher planned spending by itself is not evidence of corresponding profits for every participant.
The funding chain changes who carries the risk
Brookings estimates that combined capital expenditure by Oracle, Microsoft, Amazon, Meta and Alphabet exceeded $400 billion in 2025 and could exceed $800 billion in 2026, above their combined operating cash flow. The 2026 figure combines reported results with guidance and analyst estimates. Crucially, the series includes spending unrelated to AI and excludes data centres leased by those firms. It signals a financing pressure under the paper's assumptions, not a clean measure of AI-only cash burn.
A hyperscaler can lease a campus from a project company rather than own and finance the entire building. That project may have equity from an infrastructure investor and debt from banks or private lenders, underpinned by a long-term tenant contract. The arrangement can put productive capacity to work without the operating company supplying all the upfront capital. It also moves some debt and construction risk to the project. It does not remove the tenant's economic commitments or the creditor's dependence on the tenant's ability to pay.
The paper uses Meta's Hyperion financing as an example of high debt at project level alongside lease and residual-value obligations. It is one transaction, not proof that all AI campuses have that leverage. The broader mechanism is relevant: creditors may depend on the same few tenants and on collateral whose value changes quickly. If AI demand or power delivery falls short, cash flow, equipment resale values and refinancing terms can weaken together. Risk may be distributed among investors yet remain correlated.
A stress test needs visibility, not a bubble label
External funding has a sound counterargument: long-term contracts with creditworthy tenants can support useful infrastructure, while different investors can supply capital suited to buildings, power assets and shorter-lived chips. Weak demand is one scenario, not an established outcome. The Brookings author explicitly says it is premature to declare AI infrastructure a systemic risk on the scale of past credit booms.
The evidence that would sharpen the assessment is concrete: projects completed and energised versus announced; contracted capacity and actual utilisation; tenant concentration; maturity schedules; lease and guarantee obligations; and recovery values for older hardware. Company filings and project-credit disclosures need to make those exposures traceable across the operating firm and its financing vehicles. Until then, the $10.3 trillion case is a way to ask where capital and risk would accumulate if the ambitious pipeline were realised, not a prediction of either certain gains or an inevitable crash.