AI hyperscalers' $132bn datacentre debt raises bubble-collapse risk

Token prices halve since June while datacentre hardware costs stay elevated

The hyperscaler companies leading the current AI infrastructure buildout — Google, Amazon, Microsoft, Meta and Oracle — have issued an estimated $132bn (£99bn) in debt this year to fund datacentre expansion, according to Heather Stewart’s economics column published in The Guardian on 20 September. Stewart warned that the financial structures underpinning the AI industry could prove a more immediate threat than the existential risks dominating the week’s headlines. “It will hardly be our most pressing concern if the bots are poised to take over the world, but a collapse of the AI bubble would have repercussions far beyond the US,” Stewart writes.

The column arrives amid heightened public attention to AI safety, with industry leaders this week calling for limits on the pace of frontier AI development and on how the technology can be used. Stewart argues those concerns are warranted but should not crowd out scrutiny of the industry’s underlying financial architecture.

The scale of borrowing reflects what it takes to construct the physical infrastructure powering large AI models. Against 10-year US Treasury yields standing at 4.94%, a global benchmark for borrowing costs, the size of those debt piles could itself become a trigger for a market rethink, Stewart argues.

Stewart points to a second concern: the unit economics of AI are not improving. As Bloomberg reported, “The price of AI is collapsing, while the cost of building it is not.” Stewart notes that OpenAI has repeatedly cut its fees to retain customers, and an index from the research firm Silicon Data tracking what customers pay per million tokens — the units of data processed by large language models — shows the figure has more than halved since June to less than $1.

At the same time, frenzied demand for the real-world components of datacentres — semiconductors and other hardware — has kept construction costs elevated. “The maths only works, it seems, on the assumption of epic revenue growth,” Stewart writes. Anthropic apparently told investors recently that its “adjusted operating income” was positive, but Stewart notes that this measure effectively excludes many of its costs. Digital rights campaigner Cory Doctorow put it more sharply: “These companies are claiming that they are so cool that their profitability can only be measured using a novel, secret form of mathematics.”

Beyond the headline debt figures, Stewart cites a research note from the financial analyst Groundbreaker setting out what it calls a $1.5tn “compute commencement wall” facing AI labs over the next couple of years, drawing a parallel with the moment in 2007 and 2008 when cut-price “teaser” mortgage rates began to run out. When those teaser rates expired, low-income homeowners were flipped onto much higher rates and began defaulting in large numbers, lighting the touchpaper on the global financial crisis.

Groundbreaker’s analysis finds that datacentres are often being built and fitted out on “take or pay” contracts, with not a dollar due until a deadline — often two to three years out — is hit and the facility comes online. In the interim, the hyperscaler building the datacentre books the contract as expected future revenue and shareholders cheer; meanwhile the buyer, typically a frontier lab such as OpenAI or Anthropic, does not yet have to account for the costs it will eventually owe.

The abrupt jump in costs as those contracts mature could reach $700bn next year and more than $800bn in 2027, the analysis suggests. That could be fine if revenue continues to rocket, or “not so much,” Stewart writes, if AI’s end users are not prepared to pay enough to cover the costs — perhaps because cheaper options arise. Though the commitments are not technically debt, “the impact, if obligations can’t be met in full, would shake the foundations of the entire edifice,” she argues.

Stewart situates the financial anxieties alongside the AI safety concerns that have animated the week’s headlines. Some recent proposals, including independent analysis of AI models, appear to be genuine improvements on the unregulated status quo, she writes, citing as evidence the ability of Meta’s smart glasses to film people without consent and the swarms of chatbots that went on a hacking spree. “There is ample evidence that AI urgently needs regulating,” Stewart writes.

But Stewart also flags a more parochial concern. “Perhaps a government-backed AI ‘pause’ could prevent cheaper Chinese options from encroaching on Silicon Valley’s market dominance, for example,” she writes. “We should also be alert to the risk that a small number of intricately linked megafirms that have racked up multibillion-dollar debts are hoping the state will throw a regulatory moat around them.”

Both sets of concerns matter, Stewart concludes. “Given the spate of recent revelations about revolting bots, the focus on AI safety and the risks to humanity may well be justified, and should be tackled,” she writes. “But that should not prevent us from fretting about the delicate, interlinked financial structures that underpin the AI boom, and the risks for us all if they crumble.”