AI Capex in 2026: $725 Billion Spent, and Nobody Has Published the Payback Math

Four companies plan to spend roughly $725 billion on capital expenditure in 2026. That is up about 77% on the $410 billion they spent in 2025. No participant in AI capex at this scale has published a payback schedule you can check. The spending is real, the revenue is real, and the distance between the two is the whole argument.

What "AI capex" actually means in 2026

Capital expenditure is money spent on assets a company expects to use for years, rather than costs consumed this quarter. AI capex is the slice of that pointed at data centres, accelerators, networking and power.

The distinction matters for one reason. Capex does not hit the income statement all at once. It is depreciated across an assumed useful life, so a company can spend enormously this year and report a modest earnings hit.

That is not a trick. It is standard accounting. It does mean the reported profit picture and the cash picture tell you different things, and right now they are diverging.

Where the $725 billion goes

Four companies have each committed a sum close to a national infrastructure budget, aimed at the same three inputs. Chips, power, and floor space.

Figures are company guidance as reported through Q2 2026 earnings coverage. Ranges are the company's own, not estimates. Oracle is included for the ratio, which is why the four-company total quoted elsewhere in this post excludes it.

Chips are the majority of it

Accelerator procurement is the first line in every one of these plans. That concentration is why a single supplier's order book has become a proxy for the health of the entire category.

Power and land are the constraint

Chips can be bought faster than substations can be built. Grid connection queues, turbine lead times and local permitting now set the pace of deployment more than silicon supply does.

Some of the increase is price, not volume

Part of the 2026 rise reflects component cost inflation rather than more capacity. Microsoft's higher chip prices alone were estimated at around $25 billion of its 2026 figure, per Statista's tracking of the reported numbers. A rising capex line does not always mean more computers.

The revenue side of the ledger is much smaller

The build is justified by demand for AI products. So it is fair to ask what those products currently earn.

By the middle of 2026 the two largest AI labs were reporting annualised run rates in the tens of billions. Anthropic reported a run rate of about $47 billion in late May 2026 alongside its Series H. OpenAI has confirmed roughly $2 billion per month, near $25 billion annualised.

Those are real businesses growing at rates the software industry has not seen before. They are also, combined, a fraction of one year of the capital being deployed to serve them.

Run rate is not revenue

A run rate annualises a recent period. It is a forward-looking snapshot, not an audited result. Reporting conventions differ too. Anthropic reports on a gross basis that counts end-customer spend through cloud resellers, which makes direct comparison with OpenAI's figures unsound.

Treat every number in this section as directional and dated. That is a limitation of the public record, not a criticism of the companies.

The last row is the honest gap in this table. The hyperscalers do not disclose AI revenue as a clean line, so the comparison above understates what the build already earns. Do not read the first three rows as a complete picture.

Free cash flow is the real tell

Reported earnings can absorb a lot of capex through depreciation. Cash cannot.

The four largest US internet companies generated a combined $200 billion of free cash flow in 2025, down from $237 billion in 2024, according to CNBC's analysis of the reported results. That decline happened while revenue and profit were both growing. Capex now runs at between a quarter and most of annual sales, depending on the company.

Watch this line rather than the headline capex number. Capex tells you ambition. Free cash flow tells you what the ambition costs, and how long it can continue without new financing.

Alphabet's decision to raise $80 billion in equity in June 2026 to fund infrastructure commitments is the clearest signal so far. When companies of that size stop funding a build from operations, the funding source has become the story.

The accounting choices that flatter the picture

Depreciation schedules

The useful life assigned to an accelerator decides how much of its cost lands in this year's profit. Stretch the assumed life and reported earnings improve without anything changing in the business.

Nobody outside these companies knows the true economic life of a 2026 accelerator under continuous load. The assumption is a judgement, and it is a judgement that moves billions.

Leases signed but not commenced

Data centre leases that are signed but have not started do not appear on the balance sheet under current lease accounting. Moody's Ratings analysed the year-end 2025 disclosures of Amazon, Meta, Alphabet, Microsoft and Oracle and found $969 billion in total undiscounted future lease commitments.

Of that total, $662 billion covers leases that have not yet commenced. Moody's analysts David Gonzales and Alastair Drake calculated that the unrecorded portion equals 113% of the same five companies' adjusted on-balance-sheet debt.

Gonzales was careful about the framing. He told Fortune the money is not a liability avoided through structuring, it is a liability not yet triggered, because the services have not yet been received. That distinction matters and it does not change the size of the number.

You can watch it move in a single company's own filings. Alphabet disclosed uncommenced data centre lease payments of $23.9 billion in Q2 2025, rising to $42.6 billion in Q3. That is one quarter, one company, and an obligation that a reader of the balance sheet alone would never see.

Where this argument is weakest

Every post arguing the spend is irrational skips this section. Here it is.

The case that overbuilding is rational

Compute has been supply-constrained for three years. In a constrained market, the cost of underbuilding is losing a generational platform position. The cost of overbuilding is idle assets and a bad two years.

Those risks are not symmetric. A rational board facing that asymmetry overbuilds on purpose. Framing it as mania misses that it may be a deliberate, defensible bet.

What the sceptics keep getting wrong

The comparison to 2000 is used loosely. Around 14% of dot-com companies were profitable at the peak. The companies funding this build are among the most profitable in history and are spending from operating cash flow, not only from raised capital.

The revenue is also growing fast enough that a static spend-to-revenue ratio misleads. A ratio calculated against a business tripling annually looks very different 18 months later.

The part nobody can settle

Both cases rest on a forecast of demand that does not exist yet. Anyone claiming certainty in either direction is selling something. The useful position is to name the indicators that would change your mind.

Three numbers to watch instead of the headline

The $725 billion figure is the least useful number in this story. It is guidance, it moves every quarter, and it tells you nothing about return.

The third row is not hypothetical. Nvidia's announced framework to invest up to $100 billion in OpenAI was a letter of intent, and the finalised investment came in at $30 billion. Jensen Huang said publicly in March 2026 that the larger figure was probably not in the cards. Announcements and agreements are different objects, and the gap between them is covered in more detail in the breakdown of how the circular deals actually work.

What this means if you run a services business

You are not a hyperscaler CFO, so the relevant question is narrower. What does this do to your costs and your buyers?

Two things, both practical. Pricing power is migrating upstream toward chip, power and silicon suppliers, which means the compute inside the software you buy is not getting cheaper as fast as headline token prices suggest. And your own buyers are being asked harder questions about AI spend, which changes how you have to justify a purchase.

The discipline that follows is unglamorous. Measure what a tool changed before you renew it. That is the same test being applied to a $725 billion build, scaled down to a line item you actually control, and it is the subject of the piece on where measurable AI return has actually shown up.

Apply the same test to your own vendors. Ask any supplier for one number with a stated sample, a stated period and a stated exclusion list. A vendor who can produce that has measured something. A vendor who answers with a multiplier has not.

Frequently asked questions

Where to start this week

Two things, both doable in an afternoon.

First, pick your three largest software and AI line items and write down what each one was supposed to change, with the number you had before you bought it. If you cannot find that baseline, that is the finding.

Second, set a calendar reminder for the next quarterly earnings from the four companies above and read one line only, which is free cash flow. It will tell you more about the next 18 months of AI pricing than any capex headline will.

Related reading