AI boom faces scrutiny over debt, falling prices and rising infrastructure costs

The Guardian’s Heather Stewart says concern about AI should extend beyond safety to the financial foundations of the industry. She points to heavy borrowing for datacentres, falling prices for AI services and large future contractual commitments that could expose companies and markets if expected revenue growth fails to materialise.

Warnings about powerful and uncontrolled artificial intelligence have renewed debate over safety and regulation. But The Guardian’s Heather Stewart argues that the industry’s financial structure also deserves attention, saying a collapse of the AI bubble could have consequences well beyond the United States.

Stewart highlights the borrowing used to finance rapid datacentre construction by major technology companies including Google, Amazon, Microsoft, Meta and Oracle. One estimate cited in the article puts their combined debt issuance for this purpose at $132bn this year. Stewart says the scale of borrowing is a potential concern while 10-year US Treasury yields, a benchmark for global borrowing costs, are about 5%.

The article also questions whether the economics of AI are improving. Citing a Bloomberg report, Stewart says the price of AI services is falling while construction costs remain high. OpenAI has repeatedly reduced fees to retain customers, while an index tracking the cost of processing a million tokens has more than halved since June to below $1. At the same time, demand for datacentre components such as semiconductors is keeping some costs elevated.

Stewart says the industry’s financial projections depend on very large increases in revenue. She cites Anthropic’s claim to investors that its “adjusted operating income” was positive, while noting that the measure excludes many costs. She also quotes digital rights campaigner Cory Doctorow, who criticises technology companies for using unusual measures of profitability.

A further risk, according to analysis from financial research company Groundbreaker cited by Stewart, comes from future obligations that are not fully reflected by headline debt figures. The report identifies a $1.5tn “compute commencement wall” over the next several years. Datacentres may be built under “take or pay” contracts, with payments delayed until facilities become operational, allowing expected revenue to be recorded before customers face the full cost.

Groundbreaker estimates that payments could rise sharply, reaching $700bn next year and more than $800bn in 2027. Stewart says those obligations could be manageable if AI revenues grow rapidly, but warns that cheaper alternatives or insufficient customer demand could leave companies unable to meet them. She concludes that AI safety risks require action, but says they should not obscure the interconnected financial vulnerabilities supporting the boom.