Article
AI News Cybersecurity

The AI boom is being built on borrowed money. Debt funding and why it spark the next financial crisis

Data centres and chips are increasingly debt-financed, and the demand to fill them is unproven.

by TechDefused Newsroom
The image shows a wall with graffiti that reads 'DEBT WILL TEAR US APART'. The stark black text is contrasted against a weathered, beige background, suggesting social commentary on financial issues.

Strip away the talk of intelligence too cheap to meter and the AI buildout starts to look like something older and more familiar. It looks like a giant property deal, financed the way property deals always are, with other people's money.

The numbers are eye-watering and the plumbing is opaque. Cloud providers borrow to buy the graphics chips their customers rent. Developers borrow to raise the data centres those chips sit in, secured against a customer's promise to keep paying. Startups now raise a billion dollars at a time simply to reserve compute they have not yet used.

Because the whole edifice is leveraged, the terms have hardened. Providers increasingly demand long contracts and hefty payments up front, which work rather like the deposit on a house, shrinking the mortgage that follows. Creditworthiness sets the price of the money. A blue-chip such as Microsoft might finance its chips at around 6%. A shakier name pays closer to 9%, and borrowing costs across the sector are drifting upward.

Where it could bite

The clever structures are multiplying to match. Deals are increasingly routed through separate joint ventures that raise their own debt, occasionally with a giant such as Google standing behind them as guarantor so that lenders will touch an unproven counterparty. It is ingenious. It also spreads the risk around in ways that are hard to see from the outside.

And the risk is real enough. The central worry is not clever financing but plain demand: whether all this capacity will actually be wanted once it is switched on. Underneath sit the familiar hazards, customers who cannot pay, power that cannot be secured, construction that runs late, and the great unknown of what a two-year-old graphics chip will fetch when its contract expires.

For now the sums work because compute is scarce and the specialist "neocloud" providers can name their price. The hottest capacity is being sold at a premium on three to six month terms, snapped up by buyers who need it now and will not wait. CoreWeave, the biggest of the neoclouds, has just reiterated as much.

That premium is not the whole story, and it does not mean an entire book of contracts reprices overnight. The long deals already signed are locked in. The question is what happens when they roll off, into a market where the cost of borrowing has quietly climbed and nobody can promise the demand will still be there. Cheap money built this. Dearer money will test it.

by TechDefused Newsroom