A collapse of the AI investment bubble could pose a more immediate threat to markets than fears over uncontrolled artificial intelligence itself. Hyperscalers have issued $132bn in debt this year to fund datacentre construction, while AI firms' unit economics remain strained and a wall of deferred compute costs looms over the next two years.
While AI safety warnings have dominated headlines, the financial structure underpinning the AI boom carries risks that could hit markets sooner. Debt-funded datacentre construction, weak unit economics and looming deferred costs together form a fragile foundation for the sector.
Hyperscalers pile on debt to build datacentres
Google, Amazon, Microsoft, Meta and Oracle have issued $132bn (£99bn) in debt this year alone to fund the rapid rollout of AI datacentres, on one estimate. That borrowing is happening in a market where yields on 10-year US treasuries are hovering at about 5%, a global benchmark for borrowing costs. The scale of the debt piles could itself become a trigger for a market rethink if bond markets stay fragile.
Token prices fall while chip costs stay high
An index tracking what customers pay for a million tokens processed by large language models has more than halved since June, to less than $1, according to research company Silicon Data. Yet demand for the physical components of datacentres, such as semiconductors, is keeping build costs elevated. Anthropic recently told investors its "adjusted operating income" turned positive, but that measure excludes many of the company's costs. Digital rights campaigner Cory Doctorow said these firms claim that "profitability can only be measured using a novel, secret form of mathematics".
A $1.5tn wall of deferred compute costs
Financial analyst Groundbreaker has identified a $1.5tn "compute commencement wall" facing AI labs over the next couple of years, drawing a comparison with the "teaser" mortgage rates that expired in 2007 and 2008. Many datacentres are built under "take or pay" contracts, with no payment due until a deadline, often two to three years out. Hyperscalers book that expected revenue now, while the AI labs buying the capacity don't yet account for the costs. Groundbreaker's analysis suggests the resulting jump in costs could reach $700bn next year and more than $800bn in 2027.
That could prove manageable if AI revenue keeps rising fast enough. But if end users balk at the price once cheaper alternatives emerge, the interlinked financial structures propping up the AI boom would be tested directly.
Source: The Guardian
Trading involves risk.
Risk Disclaimer
