AI datacentres: $132bn of debt this year, $1.5tn owed on contracts

The Guardian set the AI buildout against its own bill this week, and the useful part is the gap between the two numbers in it. The debt everyone watches is $132bn, issued this year by the five companies building the datacentres. The obligation nobody books is $1.5tn, and it comes due inside two years.

The mechanism is the contract type. Datacentres are being built on take-or-pay terms under which not a dollar is owed until the deadline arrives and the building switches on, often two to three years out. Until then the company constructing it records the contract as expected future revenue and the shareholders take the win, while the lab that will use the capacity carries no cost at all. Both sides of the same contract look good at once, which is the part that should draw attention.

Three numbers from the piece

• $132bn of debt issued this year by Google, Amazon, Microsoft, Meta and Oracle, on the estimate cited.
• $1.5tn of compute contracts coming due over the next two years: about $700bn in the first, over $800bn in the second.
• Under $1 per million tokens on the Silicon Data index, down by more than half since June.

Run the two figures against each other and the visible borrowing is 8.8% of what has actually been promised. The rest sits outside the debt line. The scale is easier to feel against a market our readers already price daily: $1.5tn is 54.7% of the entire $2.74tn crypto market and 92.9% of bitcoin's market cap on its own. The $132bn of real debt, at the roughly 5% the article puts on ten-year treasuries, costs about $6.6bn a year in interest before a single server is switched on.

Why the wall points at crypto

What makes the wall a problem rather than a schedule is which way the revenue line is moving. Bloomberg's summary, quoted in the piece, is that the price of AI is collapsing while the cost of building it is not: the Silicon Data index has the price of a million tokens down by more than half since June, to under $1, while demand for chips keeps the build cost up. Miners know this trade from the other side. MARA is converting capacity to AI hosting while buying back bitcoin above the price it sold at, and hosting contracts are written on the same take-or-pay logic.

“These companies are claiming that they are so cool that their profitability can only be measured using a novel, secret form of mathematics.”

— Cory Doctorow, Quoted by The Guardian

Cory Doctorow, digital rights campaigner, on how the labs report profitability

The comparison the analysis draws is to 2007, when teaser mortgage rates expired and payments reset upward on borrowers who could not absorb them. The analogy holds only if the cash does not arrive, which is the open question rather than the conclusion. Check the arithmetic yourself: $700bn plus $800bn is the $1.5tn wall, so the two annual figures and the headline number describe the same obligation rather than three separate ones.

We have written before that twelve years of warnings about AI produced two labs and no pause. The financial structure under those labs is the part that can fail on a date, and the date is already in the contracts. Watch the token price index and the first take-or-pay deadline, not the safety statements.

Informational material, not investment advice. The debt and contract figures are third-party estimates cited by The Guardian; crypto capitalisation is CoinGecko data for 20 September 2026.