Big Tech Hides Three Trillion Dollars of AI Commitments

Big Tech hides three trillion dollars of AI commitments from investors.

It is true that the companies are building infrastructure for a market with genuine demand, in the narrow sense that cloud customers are buying more computing capacity and that the current generation of large language models requires extraordinary quantities of chips, electricity, cooling, and data-centre space. The trouble is that demand is being treated as a revenue forecast, while the obligations required to serve that forecast are treated as footnotes.

A Wall Street Journal analysis of recent securities filings found that nine technology companies have accumulated roughly $3 trillion in commitments tied to artificial-intelligence infrastructure: about $1.9 trillion in purchase obligations and $1.2 trillion in leases that have not yet begun. The same companies reported roughly $600 billion in traditional AI-related capital expenditure over the trailing twelve months.

The ratio is the important part. The companies have committed to roughly five times as much future infrastructure as they have spent openly on AI over the past year. That comparison understates the gap, because the $3 trillion is a multi-year pipeline of commitments, while the $600 billion is a single trailing twelve-month slice of cash actually leaving the building. The uncommenced leases alone are four times the level disclosed a year earlier.

The commitments also total roughly three times the companies’ combined reported borrowings and outstanding leases, which amount to about $1 trillion. Even Morgan Stanley’s accounting analysts wrote in April that it was “becoming increasingly difficult for investors to assess companies’ total potential leverage.” That is a diplomatic way of saying that the people paid to read the footnotes no longer trust the bindings.

Purchase commitments generally remain off the balance sheet until the product or service is delivered. Lease obligations do not appear until rent payments begin. A company can therefore sign the contract, reserve the chips, guarantee the energy, and secure the building while the headline balance sheet continues to present a cleaner version of the future.

This is not an accounting curiosity.

It is a financing architecture.

Meta’s Hyperion data centre in Louisiana makes the structure visible. The campus covers the equivalent of roughly 1,700 football fields, which is the sort of unit that ensures we will never take American football fields away from each other. Funds managed by Blue Owl Capital own most of a joint venture that owns the campus. A holding company called Beignet Investor raised $27 billion in construction financing through a bond sale. Meta is the minority partner and tenant.

Meta initially agreed to lease Hyperion for four years beginning in 2029, with options to renew for up to twenty years. It also guaranteed that bondholders would be made whole if Meta did not remain for the full two decades. Meta says payments under that guarantee are not probable, so it has not recorded a related liability. It has nevertheless disclosed an initial lease commitment of about $12.3 billion for Hyperion, alongside roughly $347 billion in total obligations for leases that have not yet begun.

Read that sentence twice.

The guarantee does not appear on the balance sheet because the loss is not probable. That is not a finding of risk. It is a finding of accounting. The bill exists, the company helped make the bill possible, and the company may have to pay it, but the bill is not yet sitting in the room where investors are looking.

The future is the bag.

Meta’s structure is not illegal merely because it is complicated. The technical term is off-balance-sheet exposure. The plain-language translation is that the company has arranged for someone else’s balance sheet to carry infrastructure built around Meta’s promise to use it. The bondholders receive a more legible stream of payments. Meta receives capacity before rivals can take it. Investors receive a growth story built around revenue that has not yet arrived.

Alphabet’s contractual obligations rose from $332 billion to $811 billion in three months. The company says the commitments are primarily for technical infrastructure, inventory, and energy agreements for data-centre use, with some energy obligations extending to 2054.

Thirty years.

The filing does not explain the full reason for the quarterly increase. It does not need to. The language is sufficient to establish that the company is committing itself to technical infrastructure and energy contracts far beyond the time horizon in which most people can confidently predict demand, electricity prices, regulation, or the useful life of the present generation of large language models.

Nvidia committed to $27 billion in equity investments between April 26 and the end of its fiscal year in January 2027. These are not abstract promises made at a conference. They are contracts, leases, guarantees, and equity commitments with a long tail.

The chips are real. The power is real. Nvidia’s revenue is not a fiction, and the financing deals it is signing with Wall Street firms are not a hallucination. That is precisely the trap. Every real input is being laundered through a financial structure whose leverage no one can read.

Real demand in a real asset class is what makes an off-balance-sheet guarantee dangerous, because the thing being guaranteed is the one item no one wants to question. The gap between what the balance sheet claims and what the enterprise is worth has been filled with something that is not a belief but a guess about the future. In accounting, a guess about the future has a technical name: an estimate. It is the only part of the financial statements that is ever actually the product. Everything else is raw material.

The companies are not required to tell investors that they are borrowing against the weather. They are required to disclose where the weather is, in a note filed with the Securities and Exchange Commission, in a typeface that renders it invisible to anyone not being paid by the hour to read it.

The proper frame is not that the companies are lying. It is that they do not need to.

The entire structure of modern financial accounting — disclosure-based and footnote-reliant — was designed for a world in which the footnotes were read and the readers were sober. Instead, the market treats a footnote the way a drunk treats a lamppost: for support, not illumination. The companies did not need to hide anything, because no one in the casino is looking at the table stakes. They are all looking at the dealer.

This is where the bezzle becomes useful. John Kenneth Galbraith coined the term in The Great Crash, 1929 for the interval in which the embezzler has received the gain while the victim has not yet noticed the loss. Cory Doctorow has spent the last decade applying the idea to technology: value looks present because the cost has not yet arrived at the person who must bear it.

The AI bezzle is not necessarily fraud. It is the distance between the promise of future utilization and the obligation already fixed in a contract. During that distance, the data centre is an asset, the chip reservation is strategic foresight, and the guarantee is merely a disclosure item. If demand slows, the same objects become excess capacity, stranded capital, and debt service.

Alphabet and Amazon have recently posted negative free cash flow, meaning capital spending exceeded operating cash, before these off-balance-sheet obligations are counted. That does not prove either company is insolvent. It does prove that visible spending already exceeds internally generated cash in at least part of the buildout while less visible commitments continue accumulating behind it.

The current AI debate has spent too much time asking whether the models work and too little asking who remains obligated to pay when the models do not generate enough revenue. This is the criti-hype problem described by Lee Vinsel: a critique can repeat the industry’s capability claims so faithfully that it inflates the thing it means to question.

The models can be useful in defined tasks. That fact does not establish that every data centre now under construction will earn an adequate return, that every power contract will become a profitable computing business, or that the revenue forecast attached to a thirty-year energy obligation will survive contact with the next technical cycle.

The companies are also creating a wicked problem. Each firm has an incentive to build because not building may leave it dependent on a rival’s cloud, chips, cables, or energy supply. Every firm can therefore describe its own commitment as defensive. Collectively, the companies can produce an infrastructure glut that none of them can safely abandon.

The individual decision is rational.

The combined result may be a very expensive way to discover that rational decisions can still add up to a bad balance sheet.

My father spent thirty years on the bar mill at Selkirk, and he had a tradesman’s understanding of two things at once: the difference between a schedule and a forecast, and the fact that financing documents live somewhere the floor never sees. When Gerdau bought Manitoba Rolling Mills in 1995, the line did not stop being useful. The question was who would control the asset, which jobs would remain attached to it, and who would carry the cost of making the operation more efficient.

My father kept his job. Several of my uncles did not.

The acquisition did not need to destroy the mill to extract more from it.

That is the inheritance of extraction: an existing productive system is treated as a platform for commitments made elsewhere, and the people who depend on it discover the new arrangement only after the important decisions have already been signed. In the AI buildout, the exposed parties include shareholders, ratepayers, workers, cloud customers, chip suppliers, and communities expected to host the electricity and physical infrastructure.

The obligation is distributed widely even when the decision is concentrated in a handful of corporate offices.

The obvious answer is not to prohibit data centres or pretend that all AI investment is a scam. It is to make the full commitment legible before the capital is sunk. Securities filings should present purchase obligations, uncommenced leases, guarantees, energy contracts, and special-purpose-vehicle debt in one consolidated exposure schedule, with cancellation terms, duration, counterparties, and minimum payments stated plainly. Independent auditors should test whether the claimed separation from the balance sheet reflects economic separation or merely legal packaging.

Regulators should treat guarantees and long-term capacity contracts as financial commitments when evaluating leverage, rather than waiting for rent to begin before acknowledging that the tenant helped finance the building. If governments absorb the downside while private firms retain the upside, the public should receive enforceable terms in return.

The pattern is the same as the financing deals built around Nvidia: future AI revenue is used to justify present financing, and present financing is used to make future AI revenue appear inevitable.

You can call this pessimism if you like. I prefer to call it the difference between a budget and a forecast. A budget is what you tell shareholders you will do. A forecast is what you tell creditors you might do. The binding document is the one that says who is responsible when the forecast fails.

The companies have not hidden the commitments in the sense of concealing them from the filings. They have hidden them in the more consequential sense: they have placed the obligations where the public must already know the question to find them.

There is $3 trillion in the footnotes.

The footnotes are where the bill begins.