How much revenue does AI capex need?

What the calculator answers

The question behind every argument about an AI bubble is a simple one: how much revenue does all this spending need to generate? Analysts have answered it many ways, usually with assumptions buried in a spreadsheet. This calculator puts the assumptions on the surface. You choose how long the equipment lasts, what return the money should earn and what margin the new revenue carries, and it shows the revenue that would justify one year of spending.

It is deliberately a back-of-the-envelope tool. It doesn’t forecast anything, and it treats one year of capex in isolation. Its value is in seeing how the answer moves when you change an assumption, and in comparing the answer with the cloud revenue the companies actually report.

How the math works

A dollar of capex has to earn back two things. First, its own depreciation: equipment that lasts six years has to be paid for out of six years of earnings, so a sixth of its cost is an expense every year. Second, a return on the money tied up in it, because the company could have done something else with the cash, such as buying back stock. Add the two and you have the operating profit the spending must generate each year.

Revenue isn’t profit, so the last step divides by the cash margin: the share of each new revenue dollar left after running costs (power, staff, networking, sales) but before depreciation. The formula is: revenue needed each year = (capex ÷ useful life + capex × required return) ÷ cash margin.

Each input is a judgment call. Useful life starts at six years because that is the top of the server range most buyers state in their 10-Ks, but data-center buildings last much longer and the newest accelerators may be replaced sooner, so a blended figure could sit on either side. A 10% pre-tax return is a common hurdle for large companies. The cash margin is the least certain input; no company discloses the margin on AI revenue before depreciation, so the calculator starts at 50% and lets you try others.

A worked example

With the starting values, one year of spending at $566B creates $94.4B of depreciation a year for six years, and a 10% return on it is $56.6B a year. To cover both at a 50% cash margin, it needs about $302B of new revenue every year.

For scale: AWS, Microsoft Cloud, Google Cloud and Oracle’s cloud infrastructure took in $459B over their latest four reported quarters, $114B more than the four quarters before. On these assumptions, a single year of capex at today’s level needs about 2.7 years of the clouds’ current revenue growth to pay for itself, and the following year’s spending needs as much again.

That comparison overstates the gap in one way and understates it in another. It overstates it because not all capex is for AI, Meta earns its return through advertising rather than cloud revenue, and the clouds’ revenue growth isn’t the only payoff (Microsoft also sells AI through its software). It understates it because the clouds’ revenue also has to pay for every earlier year of capex, which is still being depreciated. Choose the “cloud renters only” preset to drop Meta from the spending side.

Company by company

The table applies the same assumptions (a 10% return and a 50% cash margin) to each company, using its own capex over its latest four quarters and the top of the server life it states. Next to it is the cloud business through which most of that spending is meant to be repaid.

Read it as a measure of how much each company is asking of its cloud business. Where the revenue needed is a modest share of the cloud line, the spending is in proportion to a business that already exists. Where it is close to or above the whole cloud line, the company is building for revenue it doesn’t have yet: Oracle and CoreWeave are the clearest cases, which is why they rely on their backlogs to justify the spending. Meta has no cloud business, so its return has to come through advertising, which no filing breaks out.

The precedent, and why it only half applies

The first cloud build-out is the obvious precedent. In the 2010s Amazon, Microsoft and Google spent heavily on data centers years before their cloud businesses were large, and the spending looked excessive until the revenue arrived. That is the bull case for today: capacity comes first, demand fills it, and the companies that built most end up earning most.

The differences are scale and speed. The current build-out is far larger relative to the companies’ cash flows, the equipment is a bigger share of the spending and has a shorter life than buildings, and much of the demand comes from a few AI labs funded by investors rather than from millions of businesses moving existing workloads. None of that means the revenue won’t come. It means the time it has to arrive in is shorter, because the equipment wears out sooner.

Reading the result

The sensitivity table shows why reasonable people disagree so sharply about AI spending. Moving the useful life by a year or the margin by ten points changes the answer by tens of billions of dollars, and none of these inputs is disclosed. Anyone quoting a single number for “the revenue AI needs” has made these choices for you.

One more limitation is worth keeping in mind: the calculator treats a year of spending as if all of it starts earning at once. In practice a data center takes one to two years to build and fill, so the revenue arrives later than the spending, and the early years of a build-out always look worse than the later ones. That argues for patience in reading the gap, but only as long as the capacity being built is actually being sold when it comes online.

What the filings do show is the direction of travel. If the clouds’ revenue growth keeps accelerating, the number of years of growth needed falls. If capex keeps growing faster than cloud revenue, it rises. The capex tracker and following the money follow both every quarter, and the useful-life calculator shows what the depreciation assumption alone is worth.

Starting capex is cash paid for property and equipment over each company’s latest four quarters (10-K and 10-Q XBRL; finance leases excluded). Cloud revenue is AWS, Microsoft Cloud, Google Cloud and Oracle cloud infrastructure as reported in segment and product disclosures, summed over the latest four quarters in which all four report (to Q2 2026). Stated server lives come from each company’s 10-K.