The $10.3 trillion AI buildout: what the biggest infrastructure bet in generations would actually require.

Start with the most important caveat: this is a scenario, not a forecast.

The number that grabbed the market's attention came from Stijn Van Nieuwerburgh's Fall 2026 BPEA conference draft, Financing the AI Buildout. The paper models a U.S. data-center expansion large enough to require $10.279 trillion of investment between 2025 and 2032.

That estimate is not a claim that $10.3 trillion is already committed. It is the output of a capacity-realization scenario, a cost model, and a construction-spending schedule. The author explicitly calls it a scenario rather than a forecast and highlights sensitivity to large multiphase projects and developments without reported completion dates.

That distinction matters because this is exactly where data-center intelligence often goes wrong. Announced megawatts become treated as load. Proposed campuses become treated as inevitable construction. Capital estimates become treated as signed checks. Good analysis keeps the chain visible: announcement → evidence → realization assumption → infrastructure requirement → capital requirement → operating outcome.

On this measure, the buildout is larger than the great U.S. infrastructure booms.

Van Nieuwerburgh's historical comparison puts the AI buildout at 3.63% of GDP per year. The comparable averages in the paper are 2.24% for railroads, 1.13% for highways, 1.10% for telecom and fiber, 0.66% for canals, and 0.50% for electrification.

The paper also gives the comparison an important brake. Gross investment is not the same thing as durable capital formation. Rail lines, highways, fiber, transmission and buildings can remain productive for decades. GPUs may become economically obsolete in three to six years. The same dollar of AI infrastructure therefore may have a much shorter useful life than a dollar of railroad or highway investment.

There is a second caveat in the opposite direction: the $10.3 trillion estimate covers the initial construction and equipping of new capacity. It does not include later hardware refreshes. If early cohorts replace GPUs before 2032, cumulative gross spending would rise.

How the paper gets from projects to $10.3 trillion

The model begins with project-level Cleanview data accessed July 23, 2026. The database contained 2,972 projects; after excluding 39 canceled projects, the analysis retained 2,933 operational or planned projects. Capacity was reported for 2,501 projects and imputed for 432 using project characteristics including acreage, building area, investment, completion year, county and developer effects.

The underlying database represented roughly 566 GW of operating and planned U.S. data-center projects, compared with about 39 GW of installed capacity at the end of 2024. But the paper does not assume that the full announced pipeline appears on schedule.

Instead, it divides the 526.9 GW gross post-2024 pipeline into three outcomes:

• 182.8 GW operational by the end of 2032;
• 117.2 GW completed after 2032; and
• 226.9 GW never completed.

The “never” category is especially important. It is a capacity-weighted residual that captures cancellation and downsizing. It should not be read as a 43% probability that an individual project dies.

The biggest uncertainty sits in projects with no completion date.

The paper identifies 353.6 GW across 662 planned projects without a reported completion date. That bucket dominates the uncertainty. Under the central assumptions, 59.7 GW of that capacity reaches operation by 2032, 100.6 GW completes later, and 193.3 GW never materializes.

That is a major intelligence lesson for developers, utilities and governments: pipeline size is not enough. A useful project ledger needs status, evidence, timing, infrastructure dependencies, commercial commitments and confidence level. “Planned” is too broad a category for infrastructure planning.

The spending wave is uneven — and peaks above $2 trillion in 2032.

To translate capacity into money, the paper starts with a 2025 benchmark of approximately $41 million per MW, grows cost per MW by 4% annually, and spreads project expenditure across the completion year and the three preceding years.

The result is not a smooth annual build. Scenario-implied investment is $1.153 trillion in 2025, rises to $1.432 trillion in 2026, falls to $791 billion in 2029, then rises sharply to $2.058 trillion in 2032. Because construction precedes energization, spending through 2032 also includes work on projects that become operational after 2032.

Why one 200 MW AI campus can cost more than $8 billion

The paper's illustrative 200 MW AI training campus costs $8.216 billion. The physical building is only part of that number.

• $5.616 billion — 68.4% — installed IT equipment, including rack-scale systems, network fabric, shared storage, integration and commissioning;
• $2.2 billion — 26.8% — the data center facility; and
• $400 million — 4.9% — an allowance for incremental campus-related power infrastructure.

At that scale, the campus supports roughly 166.7 MW of critical IT load at a PUE of 1.2. The paper uses a representative rack configuration of roughly 142 kW per rack and estimates about 1,170 high-density AI racks.

This is one of the most useful findings in the paper. The value is concentrated in rapidly depreciating compute; the feasibility is concentrated in long-lived physical infrastructure. A project can have the capital for GPUs and still fail to energize on time. It can have a site and building shell and still face hardware obsolescence. It can have contracted demand and still face grid, permitting, water, equipment or community constraints.

The financing model changes when internal cash stops covering the pace.

The paper tracks aggregate capital expenditures and operating cash flow for Oracle, Microsoft, Amazon, Meta and Alphabet. Aggregate CapEx rises from $96.8 billion in 2020 to $415.8 billion in 2025 and a projected $800.5 billion in 2026. The same 2026 series puts combined operating cash flow at $707.1 billion.

That creates a crossover: CapEx reaches about 113% of operating cash flow. The 2026 estimate combines reported results, company guidance and Wall Street estimates; the paper also notes that hyperscaler CapEx includes non-AI spending and excludes data centers leased from third parties.

That is why the next phase of the buildout is not simply “tech companies spending more.” It is a broader capital-market story. Specialized developers, REITs, infrastructure funds, private equity, banks, private credit funds and securitization vehicles increasingly fund parts of the physical stack.

In the paper's framing, external financing expands the pool of available capital by separating the users of compute from the owners of the assets and, increasingly, from the investors who bear the risk.

Hyperion shows what “off balance sheet” can mean in practice.

The paper uses Meta's Hyperion project as an example of the emerging architecture. It describes roughly 2 GW of capacity and about $30 billion of investment, excluding the IT equipment that Meta finances separately as the tenant.

According to the paper, Meta sold an 80% equity stake to Blue Owl for approximately $2.5 billion. The resulting joint venture then raised $27 billion of external debt, implying debt equal to roughly 90% of the project's asset value. The financing uses bankruptcy-remote entities, long-duration contractual cash flows and a residual-value guarantee tied to Meta's lease options.

The point is not that this structure is automatically fragile. The debt was investment grade. The tenant is one of the strongest corporate credits in the world. The structure deliberately isolates project assets and creates creditor protections.

The intelligence issue is visibility. A hyperscaler's corporate balance sheet can remain lightly levered while underlying infrastructure carries much more project-level debt, lease exposure and contingent obligations.

The paper's real warning is correlation, not imminent collapse.

The paper is careful here. It says it would be premature to conclude that AI infrastructure already poses systemic risk comparable with earlier credit booms. The more useful questions are where exposure sits, how concentrated it is in the same tenants, how sensitive it is to AI demand assumptions, and how losses could propagate across banks, insurers, private credit, securitization vehicles and other investors.

The underlying risks are familiar to anyone developing large infrastructure:

• tenant concentration: large campuses may depend heavily on a single hyperscaler;
• technology obsolescence: chips, cooling and rack design can change faster than debt matures;
• execution risk: power, transmission, permits and specialized hardware may arrive late;
• residual-value risk: specialized assets may be worth less than assumed if demand or technology changes; and
• correlated exposure: the same AI demand assumptions can support cloud revenue, leases, asset values, chip sales and project debt simultaneously.

A second stress test: what revenue would be needed to support the investment?

Appendix D asks a different question. It is explicitly not a revenue forecast. It calculates the mature revenue required for the modeled buildout to recover investment, replace depreciating capital and compensate investors.

Under a 10% required unlevered return and a 50% operating cash-flow margin, the paper estimates that the 2025–2032 completion cohorts would need about $3.725 trillion in mature annual revenue by 2032 — roughly 9.2% of the paper's assumed 2032 GDP. The estimate assumes a six-year economic life for the 68% IT share and a 20-year life for other assets.

The sensitivity is the story. If the IT share has a three-year economic life instead of six years, the modeled mature annual revenue requirement rises to about $6.0 trillion. That does not mean AI companies will need exactly $6 trillion of revenue. It shows how strongly the economics depend on utilization, pricing power, equipment life and refresh cycles.

Good Data Center view: this is a coordination problem disguised as a capital boom.

The biggest development implication is not simply that “a lot of money is coming.” It is that money, infrastructure, project schedules and host systems have to line up at the same time.

1. Megawatts need a confidence level.

Utilities, governments and investors should distinguish announced, optioned, permitted, contracted, under construction, energized and operating capacity. One pipeline number cannot carry all of those meanings.

2. Infrastructure readiness is now part of financial readiness.

A project cannot monetize AI demand without power, transmission, equipment, water where required, permits, roads, public systems and a workable local development path. Delay in one system can delay revenue across the capital stack.

3. The responsibility for infrastructure costs matters more as projects scale.

If trillions of dollars of private investment depend on new generation, substations, transmission and other public or regulated systems, cost allocation becomes a core project question. Who pays, when, under what contract, and what happens if the project downsizes or arrives late?

4. Material change needs to be designed into commitments.

The paper itself assumes substantial pipeline non-realization. Host agreements, infrastructure commitments, community investments and public decisions should anticipate changes in size, phasing, timing, ownership and technology rather than treating the original announcement as static.

5. Better transparency is practical risk management.

Van Nieuwerburgh concludes that better measurement and transparency may be the most important policy contribution at this stage. That principle applies at the project level too: decision-makers need a visible record of assumptions, evidence status, dependencies, commitments and changes.

What better development looks like

• Maintain a confidence-weighted capacity ledger. Separate early concepts from contracted and deliverable load.
• Connect financial milestones to infrastructure milestones. Track power, water, land, permits, equipment and construction dependencies in the same project record.
• Make cost responsibility explicit. Record who funds grid upgrades, substations, roads, water systems, public-safety capacity and other project-driven needs.
• Stress-test scale and timing. Ask what happens if capacity arrives later, smaller, larger, or under a different technology configuration.
• Track commitments through material change. Define how community and infrastructure commitments adjust if phases, ownership, tenant mix or demand changes.
• Preserve evidence status. Keep verified facts, company claims, models, scenarios and Good Data Center analysis visibly separate.

Questions to carry into projects

• What capacity state are we actually discussing: announced, contracted, permitted, under construction, energized, or operating?
• Which infrastructure assumptions determine the project's critical path, and which are controlled by third parties?
• Who bears the cost if power, transmission, hardware or public infrastructure arrives late?
• What obligations remain if the project is delayed, downsized, rephased, sold or never reaches full buildout?
• How reusable are the buildings, power assets and public infrastructure if compute technology changes faster than expected?
• Where does leverage sit — corporate balance sheet, project vehicle, landlord, utility, private credit, equipment finance — and what cash flow ultimately supports it?
• Which facts should be public, which are legitimately confidential, and which need independent verification?