The AI Bubble, Argued With Numbers

In December 2025 the consensus figure for what AI companies would spend on infrastructure in 2026 was 527 billion dollars. Seven months later the tracked number was already above 690 billion for the fiscal year, and closer to 800 billion for the calendar year once finance leases are counted. The forecast was not slightly wrong. It was wrong in the same direction it had been wrong for two consecutive years.

That is the most reliable fact in this entire argument, and it belongs at the top because it cuts against both camps. The people saying the spending is unsustainable have been saying so while it accelerated. The people saying the forecasts are fine have watched consensus underestimate the number three years running.

So this piece does not ask whether the bubble is popping. It sets out the five quantities that decide the question, states where each one stands and when it was measured, and names what would have to change before any of it reached a business with one or two people in it.

Indicator one: how much is actually committed?

Goldman Sachs wrote in December 2025 that consensus for 2026 sat at 527 billion dollars, up from 465 billion at the start of the third quarter earnings season, with as much as 200 billion of further upside. It also noted third quarter 2025 hyperscaler capex of 106 billion, growing 75 percent year over year, and made the point that matters most for anyone reading forecasts: consensus estimates had been too low for two years running, with actual growth above 50 percent against roughly 20 percent projected.

By July 2026 the picture had firmed up. FactSet reported aggregate FY26 capex above 690 billion dollars, growth above 80 percent, and a trailing twelve month figure to May 2026 of about 490 billion. For scale, the same aggregate in FY20 was 95 billion. The projection for FY28 is above 900 billion.

Nothing about those numbers proves a bubble. Spending is not evidence of anything on its own. It is the denominator for everything that follows.

Indicator two: how much revenue has actually been recognised?

Far less, and the honest statement is that the gap is real but poorly measured in public. The cleanest documented case is OpenAI. The assembled record has it committed to 1.4 trillion dollars of datacentre spending across eight years against 13 billion dollars of reported revenue, with projected annual losses running through 2028, including 74 billion of operating losses in 2028 alone. Deutsche Bank's Jim Reid put cumulative losses at 140 billion between 2024 and 2029.

A caveat that matters and gets skipped: a committed spend across eight years is not comparable to one year of revenue, and treating them as a ratio produces a scary number that means very little. The meaningful version is the trajectory. Revenue growing faster than the commitment schedule closes the gap; revenue growing slower widens it. Anyone quoting the two figures side by side without saying which is happening is producing a headline, not an analysis.

Indicator three: who is paying for it?

This is the indicator that changed most in 2026, and it is the one we would watch above the others.

Through 2024 the build was funded from operating cash flow, which is why the spending was defensible almost regardless of return. The FactSet analysis shows that changing: incremental debt as a share of capex rose from 9 percent in FY24 to 32 percent in the twelve months to mid 2026, aggregate debt reached about 700 billion dollars, and free cash flow is expected to move close to zero or negative for all of the group except two names. Add lease related commitments of roughly 820 billion and an 84.75 billion dollar equity raise by Alphabet in June 2026.

Debt funded infrastructure is not automatically a bubble. Railways, telecoms and datacentres were all built this way. What debt does is remove the option to stop. A company funding capex from cash can slow down in a bad quarter. A company funding it from bonds has a payment schedule that does not care about the quarter, and that is the mechanism by which an investment cycle becomes a financial event. We looked at how central bankers have been framing this exposure in the piece on the warnings coming from JPMorgan and the BIS.

Indicator four: how fast is the hardware being written off?

The most technical indicator, the least discussed, and the one with the most direct route into reported profit.

Depreciation decides how much of a chip's cost hits earnings each year. The accounting analysis notes that Nvidia points to customers using four to six year depreciable lives, and that between 2020 and 2024 large technology companies extended those lives, which reduces annual depreciation and raises reported profit without anything changing in the world. In 2025 the trend split: Amazon shortened useful life on a subset of servers while Meta extended its estimates further.

The size of the exposure comes from Michael Burry's estimate that moving from four to six year schedules to a two to three year cycle would cost more than 176 billion dollars of reported earnings across 2026 to 2028. The supporting observation is more persuasive than the estimate: H100 systems trade at less than half the price of new ones by year three, which is a market signal about economic life that a straight line schedule does not reflect. There is also more than 100 billion dollars of construction in progress not yet being depreciated at all, which is a bill that arrives later by construction.

Indicator five: how much of the revenue is circular?

Circular here means a vendor investing in a customer who then buys the vendor's product, so that the same dollar is counted as investment on one side and revenue on the other. The documented examples are not obscure. Nvidia announced a 100 billion dollar investment into OpenAI in September, with the expectation of further GPU purchases. OpenAI bought billions of dollars of graphics cards from AMD and became a major shareholder in it. Microsoft holds a large stake in OpenAI. Oracle entered a 300 billion dollar agreement with the same company.

None of these is improper and all of them are disclosed. The problem is measurement. When the buyer's money came from the seller, revenue growth stops being an independent signal of demand, and revenue growth is the exact number the entire capex case rests on. Nobody outside these companies can currently strip the circular portion out, which means the most important denominator in the argument is unavailable to the public.

How reliable are the forecasts themselves?

Worth its own answer, because most arguments about this topic are arguments about projections rather than about outturns.

The record is one directional. Goldman's own account has actual capex growth exceeding 50 percent in both 2024 and 2025 against roughly 20 percent projected, and the 2026 consensus itself moved from 465 to 527 billion inside a single earnings season. The FY26 outturn tracked in July 2026 then came in above 690 billion. Three separate estimates, each too low, each revised up rather than down.

Two readings follow and they are both defensible. The optimistic one is that demand keeps outrunning the analysts, which is what an underestimated boom looks like from inside. The cautious one is that a series of upward revisions is also what an unconstrained commitment cycle looks like, and that the constraint arrives all at once rather than gradually. What is not defensible is quoting any single forecast as though the series had a good track record.

Goldman also states the limiting factor plainly, and it is not money: supply bottlenecks rather than cash flow have been constraining the build, against strong balance sheets and a new willingness to use debt. That sentence is from December 2025. By July 2026 the debt share had tripled, which suggests the balance sheet half of it aged faster than the supply half.

Then and now, as a table

The dot com comparison is usually made as a mood. Here it is with the figures that exist, each with its source and date. Where a like for like historical number is not available, the row says so rather than estimating one.

Two of those rows say the situation is milder than 1999 and two say it is larger. That is not fence sitting, it is what the numbers do when you put them next to each other, and the reason confident answers on this topic are always produced by choosing a subset.

The one structural difference worth stating on its own: this build is being financed largely by a handful of profitable companies and by institutional debt, rather than by retail investors buying shares in loss making startups. Ordinary savers are exposed through index funds and pensions rather than directly, which changes who absorbs a correction and how visibly. The local exposure is different again, and mostly physical: the four separate water measures a data center is judged by are the numbers a host community inherits whether or not the build pays off.

What would actually have to happen before this reached a small business?

Four things, in rough order of how early they would appear. This is the watchlist, and each has a threshold worth writing down rather than a vibe.

Debt share of capex above 50 percent. It moved from 9 to 32 percent in under two years. Past half, the build is primarily creditor funded, and creditors enforce schedules that customers do not. Check it when the quarterly filings land rather than when a headline appears.

A second major provider shortening useful life, explicitly. One has already done it on a subset of servers. A second one, stated plainly in an earnings call rather than buried, would confirm that the assets are being consumed faster than the schedules assume.

Capex guidance revised down rather than up. Three years of upward revisions have conditioned everyone to expect them. A guided reduction, absent a supply constraint to blame it on, would be the first real change of direction in the series.

Free tiers and low cost plans withdrawn. This is the only indicator on the list a shop owner will feel directly, and it is the last to arrive. Inference has been sold below its cost to acquire users. When funding tightens, the generous tier goes first, then the price of the paid tier stops falling.

What should a shop actually do about it?

Less than the headlines imply, and one thing more than most people do.

The thing to do: know what you would pay if your AI tooling cost three times more, and know which of your workflows would survive that. If a workflow only works because inference is nearly free, it is a workflow with a financing assumption inside it. Most small businesses have one or two of these and have never named them.

The thing not to do: change your tooling now in anticipation. Prices have fallen consistently, the spending is still accelerating, and there is no date. Rearranging a business around a correction that has been predicted for two years is its own kind of expensive. We made the wider version of this argument in the piece on making honest decisions during an AI mania, and the conclusion has not changed.

One genuine hedge exists and it costs nothing: prefer arrangements you can leave. Portable data, models you can swap, code you own. If the cheap tier disappears, the business that can move providers in a week is inconvenienced and the business that cannot is repriced. The same logic applies to the chip layer, where the customers building their own silicon are buying exactly this kind of optionality, as we covered in the piece on why the big buyers are designing their own chips.

Why does nobody give a date?

Because a date requires knowing when revenue stops justifying the spend, and the revenue figure is contaminated by circular deals nobody outside these firms can decompose. Every published date is therefore a guess with a chart attached.

The honest form of the question is conditional rather than temporal. Not when, but what would have to be true. If capex growth continues above 50 percent while debt share rises past half and useful lives are cut, the correction mechanism is visible and mechanical. If revenue growth continues at its current pace and the debt share stabilises, the same numbers describe an expensive but ordinary infrastructure build, of the kind that has happened several times and left useful assets behind.

Both futures are consistent with everything measured today, which is exactly why this is being argued rather than settled.

What we will update here

This page is written to be revised rather than replaced. The five indicators do not change. The figures under them do, and each is dated in the text so a stale number is visible as a stale number. The next revisions worth making: the FY26 capex outturn against the 690 billion figure, the debt share at the next set of filings, any further change to useful life assumptions, and the first quarter in which guidance is revised downward.

If you are budgeting AI spend for a small business through this, the useful exercise is not forecasting the macro picture. It is knowing your own cost per unit of work and what it would take to move it. Our credit pricing is stated per generation for that reason, so the number you plan with is one you can multiply rather than one you have to model.