NextFin News - The Manhattan Project spent $2.2 billion to build three atomic bombs. America's hyperscalers will spend roughly $800 billion on AI infrastructure in 2026 alone — an outlay that, as a share of GDP, is about five times larger than the nuclear program at its wartime peak. The comparison has become the default metaphor of the AI age: OpenAI's Sam Altman invoked it in 2019, Energy Secretary Chris Wright now calls AI "Manhattan Project 2," and Washington policy papers routinely call for "a Manhattan Project-like program" for artificial intelligence. The arithmetic is correct. The analogy is structurally wrong. And the error is not cosmetic — it tells investors the wrong story about who is paying, who is in charge, and where the losses land when the buildout outruns demand.
The Scale Is Real, but the Comparison Is Not
The raw numbers do support the headline. The Manhattan Project cost $1.89 billion through the end of 1945 — roughly $26 billion in today's dollars — and employed 130,000 workers at its peak, according to the Department of Energy. As a share of the economy, it consumed about 0.4% of GDP at its 1944 peak and represented barely 1% of total US wartime spending. One historian of the program described it as "a budgetary rounding error in the US war effort."
AI infrastructure is a different order of magnitude. Sell-side analysts expect the five largest US hyperscalers — Amazon, Alphabet, Microsoft, Meta and Oracle — to spend about $697 billion on capital expenditure in 2026. Goldman Sachs puts the group's AI-related outlay near $800 billion, with global AI investment approaching $1 trillion and US AI investment at roughly $581 billion. That is approximately 2% of US GDP, or about 20 Manhattan Projects per year in inflation-adjusted terms. The four largest hyperscalers alone plan roughly $725 billion of 2026 capex, up about 77% from approximately $410 billion in 2025. Capital intensity has reached 45-57% of revenue — levels described as "historically unthinkable" for companies of this size.
Against that, the historical benchmarks look small. The Apollo program cost about $189 billion in today's dollars over 1960-1973, peaking at roughly $31 billion a year and 0.7% of GDP in 1966. The Marshall Plan ran at about $34 billion a year and 1.1% of contemporaneous GDP. The Manhattan Project, at $9 billion a year and 0.4% of GDP, is the smallest of the three. By the GDP-share metric, AI capex exceeds each individual historical program — though it remains dwarfed by total wartime mobilization, which peaked at roughly 40% of GDP in 1945.
So the soundbite — "AI spending dwarfs the Manhattan Project" — is arithmetically true. It is also almost entirely beside the point, because scale is the least interesting thing the two endeavors share.
Where the Analogy Actually Breaks
The Manhattan Project was a single, government-run, top-secret program with one verifiable objective: produce a working atomic bomb before Nazi Germany did. Success was binary and could be tested in the New Mexico desert in July 1945. Funding came from wartime appropriations; the taxpayer bore the risk; and the entire effort answered to one chain of command under the Army.
The AI buildout is the opposite on every dimension that matters for capital allocation.
First, it is private, not public. The technology is being built outside government, without meaningful public investment or coordination — the first time in history that a technology of this consequence has unfolded this way. The US-China Economic and Security Review Commission recommended "a Manhattan Project-like program dedicated to racing to and acquiring advanced AI," and the Energy Department's Genesis Mission, launched by executive order in November 2025, borrows the same language. But the actual infrastructure is being financed and built by private boards answering to shareholders, not by a program director reporting to the President.
Second, there is no single objective. A bomb either works or it does not. AI has no such moment of truth — which is precisely why spending can keep compounding long after marginal returns have turned negative. A government program stops when the objective is met. A public company chasing a perceived platform shift stops only when capital markets force it to.
Third, there is no single chain of command. Instead of one Los Alamos, there are several competing private labs, each racing the others and duplicating infrastructure. Microsoft, Google, Amazon and Meta are all building sovereign compute clusters, buying identical chips and competing for the same enterprise workloads — the digital equivalent of Parliament authorizing separate, redundant railway tracks between the same industrial towns during Britain's Railway Mania.
Fourth, the risk-bearing is private. Taxpayers funded the Manhattan Project. AI is funded by operating cash flow, debt, and off-balance-sheet lease vehicles. If the technology disappoints, the losses fall on equity holders, corporate bondholders, and the private lenders holding the leases — not on the public balance sheet.
"There are these moments in the history of science where you have a group of scientists look at their creation and just say, you know, what have we done?" Altman said on a podcast, acknowledging that even he cannot map out where AI is going. "I think there is a race to get somewhere, but people don't agree on where it's to."
That is not the voice of a program director reporting on a known deliverable. It is the voice of a CEO in a competitive market, uncertain of the finish line — which is exactly the point.
The Financing: Private Cash, Private Debt, and the Off-Balance-Sheet Channel
How the buildout is paid for is where the Manhattan Project framing becomes genuinely misleading for investors. The hyperscalers are funding AI largely from operating cash flow and debt. The group raised $108 billion in debt during 2025 alone, and projections suggest roughly $1.5 trillion in debt issuance over the coming years to fund the buildout. Goldman Sachs strategists expect about one-third of hyperscaler capex to be debt-financed in 2026, with project finance and data-center transactions adding another $300 billion of supply in 2027 above and beyond direct issuance.
Much of the financing does not even appear on the companies' balance sheets. Goldman Sachs analysts counted approximately $1.5 trillion of lease commitments in the footnotes of hyperscaler regulatory filings, of which roughly $1 trillion had not yet started and therefore did not appear in the accounts as conventional liabilities. The bank warned that "this treatment can understate leverage and future liquidity needs as these obligations are recognised and contractual payments come due." Separate tallies put total off-balance-sheet lease and purchase commitments near $2.5 trillion as of the second quarter of 2026. These structures concentrate the downside in private credit and lease lenders rather than in the tech giants' reported debt.
And the return requirement is steep. Goldman Sachs calculated that the five hyperscalers need roughly $300 billion in annual AI revenue to break even on their combined 2026 capital expenditure. Current AI cloud revenue is running only about $70 billion above the group's pre-AI trend line as of the second quarter of 2026 — leaving a gap of roughly $230 billion a year. To put that in perspective, OpenAI's roughly $20 billion annual recurring revenue and Anthropic's roughly $9 billion run-rate, impressive as they are, together amount to only about 3-4% of the projected 2026 capex total.
In other words, the revenue needed just to justify this year's spending is more than ten times what the two most prominent AI developers currently earn. That is not a national mission with a defined endpoint. It is a private capex cycle whose continuation depends on demand materializing at a pace with few historical precedents.
The Capital Cycle: Cyclical Boom, Possibly Structural Shift
The most important distinction the Manhattan Project analogy erases is between the technology and the capital cycle. They can point in opposite directions, and confusing them is how investors lose money.
The AI capex cycle is cyclical. It is a concentrated investment boom financed by debt, and investment booms mean-revert. The late-1990s telecom and fiber buildout peaked at about 1.2% of GDP in 2000 before collapsing and tipping the economy into the mildest post-war recession. A cycle that builds at 0.85 percentage points of GDP per year can unwind at a similar pace — and that speed of reversal, rather than the buildout itself, is the macro risk if AI demand disappoints. The US shale oil boom of the 2010s offers a private-sector analog: excitement around a new extraction technology spurred massive capital investment, investors rewarded the outlays with soaring equity prices, and when supply outran demand, prices cratered and left firms with years of excess capacity and debt distress.
The AI technology, by contrast, may well be a structural shift — a general-purpose technology in the lineage of electricity and the internet. The two claims are not the same. Britain's Railway Mania of the 1840s ended in a stock-market collapse and a generation of consolidation, yet every claim about the transformative power of railways was ultimately correct. The investors who bought railway shares at the peak often lost money anyway. The same pattern repeated with the internet: the technology reshaped the global economy, but most of the capital committed during the 1999-2000 buildout was destroyed.
This is the trap embedded in the Manhattan Project analogy. It frames the entire endeavor as a single bet on a single outcome, when in reality investors are making two separate bets: one on the technology, which may be transformative, and one on the capital cycle, which is already priced for perfection. You can be right about the first and still lose on the second — which is precisely what happened to fiber investors in 2001 and railway investors in 1847.
What the Analogy Gets Wrong About Risk
The Manhattan Project framing carries an implicit promise: this is a national mission, and the nation will see it through. That framing subtly suggests a public backstop — that if the technology is strategically vital, society will absorb the cost. In reality, there is no backstop. The "Manhattan Project" label confers national-mission legitimacy on what is, financially, a private capex cycle. It may even encourage more spending by making each incremental data center feel like a contribution to a collective endeavor rather than a competitive gamble.
The analogy also misdirects attention on the transmission of stress. Because the financing runs through off-balance-sheet leases, private credit, and project-finance vehicles, the first sign of trouble is unlikely to appear in the hyperscalers' reported bond spreads or headline leverage ratios. Stress would show up first in the private lenders holding the lease-backed paper and in the terms on which new project finance can be raised. By the time it reaches the investment-grade bonds that dominate institutional portfolios, the repricing may already be advanced.
The Strongest Counter-Argument
Defenders of the analogy argue that it is not about financing at all — it is about existential risk. If advanced AI poses a threat comparable to nuclear weapons, then the appropriate response really is a Manhattan Project-style mobilization: centralized control, government oversight, a single chain of command. Some policy analysts argue the nuclear model offers "the most robust structural precedent for global AI governance," and that the alternative — a complete absence of international oversight — is too risky. On that view, the analogy is a policy argument, not an investment thesis, and criticizing it on financial grounds misses the point.
That counter-thesis has real force on the safety question. The diffusion of AI capability across borders and private labs does raise governance challenges with no clean precedent, and the nuclear non-proliferation regime is the closest thing history offers to a template. But it does not rescue the investment framing. Even if advanced AI warrants tighter government control in the future, the infrastructure being built today is not being built under government control. The capital is being committed now, by private boards, on private return assumptions. A regulatory regime that arrives after the buildout is financed does not change who owns the risk — and it does not make the $300 billion revenue gap any easier to close.
There is also a harder objection from the other side: that the analogy understates the risk rather than overstating the coordination. Garrett Graff, arguing that "AI isn't the Manhattan Project — it's Jurassic Park," notes that the nuclear program was a budgetary rounding error in a vast war economy, built under government control with a clear chain of accountability. AI is being built by competing private actors with shareholder pressure to move fast and no equivalent safety architecture. On that reading, the analogy is dangerous not because it exaggerates the state's role but because it borrows the Manhattan Project's aura of disciplined purpose for an endeavor that has none.
What to Watch
The single number that matters is the revenue gap. If hyperscaler AI revenue closes to within roughly 10% of the approximately $300 billion annual break-even requirement — reaching about $270 billion or more — within the next 12 to 18 months, while capex growth slows, the "misframed boom" thesis weakens materially. If instead the gap widens while debt issuance accelerates, the private-funding fragility argument strengthens.
Three secondary signals deserve attention. First, the off-balance-sheet lease channel: a tightening in private credit or a repricing of lease-backed financing would transmit stress to the hyperscalers faster than deterioration in their reported bond spreads. Second, capital intensity: if spending stays above 45% of revenue for an extended period without revenue converging toward the break-even requirement, the equity risk premium for the group should rise. Third, the equipment vendors: the capex boom's first-order beneficiaries — chipmakers, data-center builders, power providers — will show margin pressure before the hyperscalers do, because their revenue is the hyperscalers' spending.
Scenarios and Time Horizons
Short term (6-12 months): the AI capex cycle remains a momentum trade. The hyperscalers report that their markets are supply-constrained rather than demand-constrained, and earnings momentum at the equipment vendors supports the complex. Volatility will be driven by capex guidance and any sign of financing stress in the lease channel.
Medium term (1-3 years): the gap between roughly $800 billion of annual spending and roughly $70 billion of incremental revenue is a cliff that must be climbed. This is the window in which the cycle either proves self-funding or begins to unwind. The base case is a gradual convergence: revenue grows strongly but falls short of the break-even requirement, forcing a slowdown in capex growth rather than an abrupt stop. The downside case is a telecom-style reversal if demand disappoints and financing conditions tighten simultaneously.
Long term (5+ years): the technology may transform the economy even if this particular capital cycle ends in write-downs. The beneficiaries are likely to be the customers of the infrastructure — the firms that buy compute and embed AI into products — rather than the builders who financed it. That was the pattern with the railroads, the internet, and electricity: the infrastructure investors often lost money; the users of the infrastructure captured the surplus.
The Bottom Line
The Manhattan Project analogy is seductive because it makes the AI buildout feel inevitable and publicly sanctioned. It is neither. This is a private, debt-funded race with no finish line and no backstop, unfolding at a scale that exceeds any peacetime mobilization in American history. The technology may prove as transformative as its boosters claim, and the capital cycle may still end in the familiar way that concentrated, debt-financed buildouts do.
Investors who treat AI as a national mission may be the ones who pay for the difference between the two.
更多独家洞察尽在 nextfin.ai.
