The Finite Capital Frontier

The debate about artificial intelligence infrastructure is conducted almost entirely in one number: capital expenditure. It is the wrong place to look. The figures that determine whether this build can be sustained sit one statement over, on the balance sheet and the cash-flow statement, and they are quieter and less flattering. This issue leads with that balance-sheet challenge to the hyperscalers. It then turns to the question it forces — namely where the capital to fund the build actually comes from — and previews a fuller treatment of the capital markets in a dedicated paper to follow.

The Balance Sheet Challenge to the Hyperscalers

The headline is genuinely vast. The four largest United States technology companies plan around 725 billion dollars of capital expenditure in 2026, up roughly 77 percent on the prior year, with capital intensity now running at 45 to 57 percent of revenue — a level the sector has never approached. That number deserves the scrutiny it receives. But it does not, on its own, tell us whether the build can be sustained, because it says nothing about the cash left to pay for it.

The starting position is already uncomfortable. For the five largest hyperscalers, aggregate capital expenditure, once dividends and buybacks are included, now sits above projected operating cash flow, which is precisely why they have begun to borrow. Around 100 billion dollars of bonds have been issued in 2026 to fund the build, and investors have demanded record levels of protection against default through the credit-default-swap market. Companies that until recently generated more cash than they could spend have become external borrowers.

A recent interview on the Wall Street Journal programme Take On the Week, with the accounting specialist Kevin Koharki, named the cost the capex debate tends to miss. It is not in the data centres at all. It is stock-based compensation — the equity used to attract and keep the engineers who design and run the compute. Under the accounting rules that pay is a non-cash charge, so it is added back when free cash flow is calculated, which flatters the figure. The real cash costs it creates appear elsewhere. When employee shares vest, the company pays withholding taxes in cash. To stop the share count rising, it buys stock back, also in cash. Both are genuine outflows, both are recorded outside operating cash flow, and so reported free cash flow looks stronger than the economics support.

The scale is material. Koharki argues that for several of the largest companies in the world, true free cash flow is 30 to 40 percent lower than the reported figure. In one worked example, roughly 90 percent of a 26.3 billion dollar buyback did nothing more than offset dilution, and once the cash costs of stock pay are netted out, free cash flow falls from about 46 billion to roughly 4 billion. Meta is the cleanest illustration: a business widely regarded as a money printer has been taking on debt to build data centres, in part because the free cash flow it reports is, once the cash cost of stock pay is removed, considerably thinner than it appears.

This is where the capital-expenditure debate and the compensation debate meet, and the meeting point is the talent war. The contest to hire scarce AI talent is funded largely in equity. Issuing that equity dilutes existing shareholders. Holding the share count steady then requires cash buybacks. Those buybacks drain the very free cash flow that is meant to fund the data centres. The silicon and the buildings on one side, and the people on the other, are competing claims on the same constrained cash, and one of those claims is partly concealed by the way it is reported.

Two further pressures sit on top. The revenue that justifies the spend is concentrated: the four largest cloud providers hold around 2.1 trillion dollars of contracted backlog, and roughly half of that is owed by OpenAI and Anthropic, two companies whose own free cash flow is deeply negative. And the assets are short-lived. Much of the spend — the chips especially — has a useful life of only three to five years, so the returns must arrive before the end of the decade. Short-lived assets, concentrated counterparties, and a cash position weaker than the headline, all funded increasingly by longer-dated debt.

None of this means the spending is irrational. The builders are pre-selling capacity before they pour concrete, the core returns are arguably already positive, and demand for compute genuinely exceeds supply. The Monard view is narrower, and we think more durable. The funding model is fragile. It depends on equity as a currency and on uninterrupted cash generation, and neither can be relied upon across a full cycle, least of all in a downturn, when issuing equity is most painful and refinancing is hardest. Even the strongest balance sheets in history are thinner than they look, and the cash that appears available to fund the next decade of compute is, in large part, already spoken for.

Which Market Can Fund It, If Any?

If the corporate balance sheet cannot carry the build alone, the question becomes where the rest of the capital comes from, and it is worth stating plainly that capital is finite. The claim now being made on it is larger than any single sector has made before. Estimates of total AI and data-centre investment to 2030 range from about 3 trillion dollars to as much as 7 trillion.

The most instructive framing comes from JPMorgan: the funding need for 2026, around 700 billion dollars, can plausibly be met from hyperscaler cash flow and the high-grade bond market, but by 2030 the annual need exceeds 1.4 trillion and will require contributions from every capital-providing market, leaving a funding gap of roughly 1.4 trillion. For context, the build would need around 650 billion dollars of new revenue every year, in perpetuity, simply to earn a 10 percent return.

Add the pools together and the dollars probably exist in aggregate. The difficulty is threefold. The capital is concentrated in a few issuers and two loss-making counterparties. The structures built to bridge the gap push risk onto the most opaque and least resilient balance sheets. And every pool can fund a good year without being able to fund a full cycle. Which market can fund the build, taken one at a time, has no satisfying answer. That is the point, and it is why the answer has to be structural.