Three facts describe the current state of artificial intelligence infrastructure spending, and none of them are matters of opinion. The companies building that infrastructure have, over the past five years, repeatedly extended the accounting assumptions that determine how quickly their hardware loses value on paper — assumptions that flatter today’s earnings regardless of what eventually happens to the hardware itself. The revenue actually being generated by the AI industry falls far short, by any reasonable measure, of what would be required to justify the scale of capital currently being committed. And the financing structure built to bridge that gap has diffused the risk away from the technology companies themselves and onto pension funds, index-fund investors, and, in some cases, ordinary electricity ratepayers — parties who never chose to make a bet on artificial intelligence at all.

None of this proves that artificial intelligence lacks real value, or that every dollar spent on it will be lost. It proves something narrower and more useful: that the scale and structure of the current buildout does not yet match the revenue behind it, and that the mechanisms built to bridge that gap have quietly moved the downside onto people with no say in how the bet was placed. This essay examines the argument in three parts — the accounting, the revenue, and the risk — and closes with the market’s own verdict so far, delivered in a single volatile week in June 2026.

The Accounting: How Long Does a GPU Really Live?

Depreciation is not a prediction about the future. It is an estimate, made now, of how many years a piece of equipment will remain economically useful, spread evenly across a company’s income statement. A longer estimated life means a smaller annual expense and a larger reported profit in the near term — with no change whatsoever to the cash actually spent. Since 2020, every major hyperscaler has moved that estimate in the same direction: longer.

Microsoft’s Form 10-K for fiscal year 2023 disclosed that the company extended the estimated useful life of its server and network equipment from four years to six, effective that fiscal year — a change that, based on equipment already on the books, increased operating income by $3.7 billion and net income by $3.0 billion for the year (Microsoft Corporation, 2023). That followed an earlier extension, completed in 2020, that had already moved server life from three years to four and network equipment from two years to four.

Alphabet made a comparable move. In its fourth-quarter and full-year 2023 earnings release, the company reported that extending server useful life from four years to six, and certain network equipment from five years to six, reduced full-year depreciation expense by $3.9 billion and increased net income by $3.0 billion, or $0.24 per diluted share (Alphabet Inc., 2024). Meta followed in 2025, extending the useful life of servers and associated networking equipment to five and a half years from a prior four-to-five-year range, a change the company disclosed would reduce 2025 depreciation expense by $2.9 billion (Meta Platforms, Inc., 2025).

Only one hyperscaler has moved in the opposite direction. Effective January 1, 2025, Amazon shortened the estimated useful life of a subset of its servers and networking equipment from six years back to five, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” The reversal increased depreciation expense by $392 million in the third quarter of 2025 alone, reducing net income by $298 million for the quarter (Amazon.com, Inc., 2025). It is a small adjustment relative to the scale of the industry’s spending, but it is notable precisely because Amazon had every commercial incentive to extend its estimate further, as its competitors had, and chose instead to shorten it for at least part of its fleet.

The dispute over which estimate is closer to correct is not a rounding error. Nvidia’s own architecture cadence refreshes roughly every three years, with each generation delivering two-to-three-times the efficiency of the last per unit of compute and per watt of power (Usvyatsky, 2025). A chip does not need to fail physically to become uneconomic; it only needs to cost more per unit of useful work than the model on the market beneath it. Investor Michael Burry has argued publicly that hyperscalers are depreciating chips over five-to-six-year schedules when the real economic life is closer to two or three years, and has estimated the cumulative effect at roughly $176 billion in understated depreciation and overstated profits across the industry between 2026 and 2028 (Usvyatsky, 2025). Nvidia and its customers dispute that estimate, arguing that a chip’s value cascades downward — from frontier training, to inference, to general-purpose compute — even after it stops being state of the art, which is a genuine argument and not merely a rationalization. But it is also, notably, an argument that assumes the cascade of use cases keeps growing fast enough to absorb the supply of aging chips, which returns the question directly to revenue.

The Revenue: What Would Actually Justify This Spending?

In 2024, Sequoia Capital partner David Cahn built a simple method for estimating how much annual revenue the AI industry would need to generate to justify its infrastructure spending. Take Nvidia’s annualized data center revenue — the chips being the closest thing to a hard number in the entire supply chain — and double it, to account for the full cost of building and running a data center around those chips: power, buildings, and networking, not merely silicon. Then double the result again, so that whoever resells that computing capacity to end customers can still clear a normal gross margin of roughly 50 percent. Applied to Nvidia’s revenue at the time, the method produced what became known as “AI’s $600 billion question” — the annual revenue gap the industry needed to close just to break even on the infrastructure already committed (Cahn, 2024).

Rerunning that same method on today’s numbers makes the gap larger, not smaller. Nvidia reported data center revenue of $75.2 billion for its first fiscal quarter of 2027, ended in April 2026 — up 92 percent from a year earlier (NVIDIA Corporation, 2026). Annualized, that is roughly $300 billion, which under Cahn’s same logic implies an industry-wide revenue bar of approximately $1.2 trillion per year — double the 2024 figure, because Nvidia’s own run rate has roughly doubled.

Measured against that bar, the two leading AI-native companies are not close. OpenAI disclosed roughly $25 billion in annualized revenue in early 2026, but at a negative 122 percent non-GAAP operating margin in its first quarter — meaning it lost $1.22 for every dollar of revenue it brought in (Zitron, 2026). Anthropic, growing faster, reported annualized revenue in the range of $30 billion to $47 billion in 2026 depending on gross-versus-net accounting methodology, a dispute the two companies have not resolved between themselves (Epoch AI, 2026). Combined, the two companies most singularly identified with the commercial promise of generative AI are generating perhaps $55 billion to $70 billion a year in revenue — under 6 percent of the bar their own infrastructure supply chain now implies.

The gap does not close on a longer horizon, either. Bain & Company’s sixth annual Global Technology Report estimates that by 2030 the industry will need $2 trillion in combined annual AI revenue to fund projected global compute demand, and forecasts that even generous projections leave the industry roughly $800 billion short of that figure (Bain & Company, 2025). This is not a fringe or adversarial estimate; it is a management-consulting firm’s own base case, built for the executives making the capital allocation decisions in question.

The Risk: Who Actually Absorbs the Gap?

A revenue shortfall of this size has to land somewhere, and the financing structure built around this buildout has been unusually effective at moving that exposure away from the companies making the largest bets.

The clearest case is CoreWeave, a “neocloud” that rents GPU computing capacity and has financed its fleet almost entirely with debt collateralized by the chips themselves. Total debt has grown from under $8 billion in 2024 to roughly $24.9 billion by mid-2026, even as the rental rates on that same GPU fleet have fallen 50 to 70 percent — a direct, present-tense symptom of the depreciation argument above (Finterra, 2026). Analysts have described the company’s central vulnerability as a “GPU maturity wall”: if the pace of hardware improvement continues at its current rate, the resale or collateral value of an aging cluster can fall faster than the debt secured against it amortizes (Finterra, 2026).

That debt is not, for the most part, sitting on the books of traditional, regulated banks. It has been originated and held by private credit funds — chiefly Blackstone, Blue Owl Capital, Apollo, Pimco, and BlackRock — which deployed more than $15 billion into GPU and data-center debt in a single month early in 2026, part of a wave that Morgan Stanley projects will reach $250 billion to $300 billion in hyperscaler-adjacent debt issuance for the year (Abasiita, 2026). Private credit funds source much of their own capital from institutional investors seeking stable, long-duration yield — pension funds and insurers prominent among them. One July 2026 analysis put the connection plainly in its headline: this circular chain of financing “puts pension funds at risk” (Tech Times, 2026). It is a chain in which the ultimate capital sits several steps removed from any decision about whether a given data center, or a given generation of chips, was ever going to generate enough revenue to pay for itself.

A second, more diffuse channel runs through public equity markets rather than private debt. The seven companies commonly grouped as the “Magnificent Seven” — Nvidia, Apple, Microsoft, Alphabet, Amazon, Meta, and Tesla — accounted for roughly 34 percent of the entire S&P 500 by weight as of mid-2026, a concentration without precedent in the index’s modern history (Brock, 2025). In January 2026, JPMorgan strategists warned that this concentration had created what they termed a “fragile fifty percent”: a structure in which a single earnings disappointment or shift in sentiment toward the handful of companies most exposed to AI capital spending could trigger a broader de-leveraging event, dragging down an index regardless of the underlying health of its other 493 constituent companies (MarketMinute, 2026). Anyone holding a standard S&P 500 index fund in a retirement account — which is to say, a very large share of American households — carries that concentration by default, without having made any deliberate decision about artificial intelligence at all.

The Market’s Own Verdict, So Far

This is no longer a purely theoretical argument. On June 5, 2026, the thesis underwent its first real stress test. The Nasdaq Composite fell roughly 4 percent in a single session — its worst day since April 2025 — while semiconductor stocks lost more than $1.3 trillion in combined market value, and the Philadelphia Semiconductor Index fell more than 6 percent (TheStreet, 2026). Coverage at the time attributed the trigger less to any single catastrophic earnings report than to a broader “expectations reset”: stretched valuations following months of momentum finally colliding with investor demand for concrete evidence that AI infrastructure spending was translating into revenue, compounded by shifting Federal Reserve rate expectations and a disappointing revenue guidance from at least one major semiconductor company (TheStreet, 2026). It was not a collapse. It was a market beginning, for the first time at this scale, to ask the same question this essay has asked: where, precisely, is the revenue that justifies the spending?

Conclusion

None of the three arguments here — the accounting, the revenue, or the risk-diffusion structure — depends on artificial intelligence turning out to be commercially worthless. Real productivity gains from these tools are already measurable in specific domains, and it is entirely possible that AI-native revenue continues to grow at the extraordinary rates OpenAI and Anthropic have posted so far. But possibility is not the same as pricing, and the current scale of capital expenditure — $600 billion to $725 billion in 2026 alone across the largest hyperscalers — has been committed on a set of assumptions about depreciation schedules, revenue trajectories, and risk tolerance that the companies’ own disclosures do not yet support.

What should concern an ordinary investor, or an ordinary citizen with a retirement account, is not whether artificial intelligence has value. It is that the specific financial architecture built around this buildout — extended depreciation schedules that flatter today’s earnings, debt structures collateralized by rapidly depreciating hardware, and equity concentration that has folded AI-specific risk into the default retirement portfolio of tens of millions of people — has been engineered, whether by design or by accident, to make the downside difficult to see until it is already underway. The accounting can be restated. The debt can be refinanced or written down. The index concentration cannot be un-concentrated by anyone who does not actively choose to rebalance away from it. A prudent investor does not need to predict the exact date or severity of a correction to act on that fact. They need only recognize, as this essay has tried to show with the industry’s own numbers, that the gap between spending and revenue is real, that it has not closed, and that someone, eventually, pays for it.

Sources