The single most important argument in technology right now is not about what AI can do. It is about whether the staggering amount of money being spent to build it can ever be earned back. And unlike most tech debates, this one has a body count waiting on the other side if it goes wrong: not just tech shareholders, but pension funds, private-credit lenders, construction firms, and the broader economy that has come to lean on a handful of companies’ spending. So it is worth setting down, as plainly as possible, what is actually known — and what is genuinely still an open question.
The number that broke the scale
Start with the figure, because it is hard to overstate. According to totals compiled by the Financial Times, the four biggest spenders — Microsoft, Amazon, Alphabet, and Meta — are on track for roughly $725 billion in combined capital expenditure in 2026, up about 77 percent from $410 billion in 2025, making it the largest single-year concentrated infrastructure cycle in the history of technology. To put that in perspective the way the energy world does: the International Energy Agency noted that the capital expenditure of just five technology companies now exceeds global investment in oil and natural gas production. A handful of software firms are now out-investing the entire global fossil-fuel extraction industry. That is not a normal corporate spending cycle. It is something closer to a private-sector industrial revolution, financed on a compressed timeline.
The bear case, stated fairly
The skeptics are not cranks, and their argument has gotten sharper as the numbers have grown. It rests on a few load-bearing facts, each independently verifiable. The first is cash-flow strain: the spending has grown so fast that it is swallowing the companies’ own earnings. PIMCO estimates that combined hyperscaler capex will consume roughly 94 percent of their operating cash flow across 2026 and 2027 — a level that leaves almost nothing for anything else and increasingly forces the spending onto debt.
The second is what sits off the books. Moody’s reported that hyperscalers hold approximately $662 billion in data-center lease commitments that have been signed but not yet commenced, sitting off balance sheet under current accounting rules — a figure larger than the combined on-balance-sheet debt of the same companies. In other words, the visible capex number, gigantic as it is, understates the true committed obligation. The bull-case reassurance that this is all “conservatively self-funded” gets harder to hold when a liability that large is parked just out of frame.
The third, and most unsettling, is the financing structure itself. The Bank for International Settlements — effectively the central bank for central banks, not a bearish hedge fund — devoted part of its flagship annual report to the risk. It flagged the danger of “circular financing,” in which hyperscalers take equity stakes in AI labs, which in turn commit to multi-year purchases of computing power from those same hyperscalers — an arrangement that can make end-demand look larger and more independent than it really is. The BIS warned that a disappointment in returns “could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust”. When the institution whose entire job is spotting systemic financial risk writes that sentence about your industry, it deserves to be quoted exactly, not paraphrased away.
The bull case, also stated fairly
And yet the optimists are not fools either, because the thing the last great bubble lacked — actual revenue — is genuinely showing up this time. The pure-play AI companies are growing at rates that are difficult to dismiss as hype: OpenAI ended 2025 at roughly $20 billion in annual recurring revenue, about triple the prior year, while Anthropic’s revenue run rate surpassed $9 billion in January 2026, up from roughly $1 billion at the end of 2024. Those are not the numbers of a product nobody wants. Demand is real and, by the hyperscalers’ own repeated account, they are supply-constrained rather than demand-constrained — they would deploy even more compute if they could build it fast enough.
The strongest version of the bull case is that this is what financing a genuine platform shift looks like from the inside, and that it always looks reckless in the middle. Railroads, electricity, and telecom all required enormous, front-loaded, partly-wasteful buildouts that terrified contemporaries and looked insane right up until the demand caught up to the infrastructure. If AI is a general-purpose technology on that scale, then underbuilding is the real risk, and the companies pulling back to protect quarterly cash flow will be the ones that lose the decade.
Where the truth probably sits
Here is the thing both camps tend to miss by treating this as a single yes-or-no question: “is it a bubble?” is the wrong frame, because bubbles and real revolutions are not mutually exclusive. The telecom buildout of the late 1990s was both a genuine infrastructure revolution — the fiber it laid still carries the modern internet — and a financial bubble that wiped out a generation of investors when the revenue timeline slipped a few years behind the spending timeline. The infrastructure was real. The valuations and the debt schedules were fantasy. Both things were true, and the gap between them is where the wreckage happened.
That is the most likely shape of the AI story too, and it reframes the question productively. The demand is probably real. The technology is probably a lasting platform. And the financing may still be structured in a way that cannot survive even a modest delay between when the money is spent and when it is earned back. The danger is not that AI is fake. The danger is a timing mismatch — hundreds of billions in near-term, debt-and-lease-backed obligations betting on revenue that is real but may arrive on a slower clock than the payment schedule demands. A revolution can be completely genuine and still bankrupt the people who financed the first leg of it.
So the honest answer to “is this an AI bubble?” is the unsatisfying one: it can be a bubble and a revolution at the same time, and the two questions have different answers on different timelines. Whether the technology matters is nearly settled. Whether the current financing structure can survive the wait for the returns is not — and that, not the capability of the models, is the actual thing to watch over the next two years.
Related reading: Everyone’s Adopting AI. Almost Nobody’s Profiting. That’s the Real 2026 Story
