Warnings about uncontrolled AI have dominated headlines, but the AI-industrial complex faces equally urgent financial risks that could reverberate far beyond US markets.
The five largest hyperscalers — Google, Amazon, Microsoft, Meta and Oracle — have issued an estimated $132bn this year to fund datacentre construction, according to one estimate. In a fragile bond-market environment with 10-year US Treasury yields hovering near 5%, the sheer volume of borrowing represents a potential trigger for a broader market reassessment.
Compounding the debt concern are deteriorating unit economics. The price customers pay for AI has been falling sharply while the cost of building it remains high. OpenAI has cut its fees repeatedly to retain subscribers, and an index by research firm Silicon Data tracking the price per million tokens processed by large language models shows the figure has more than halved since June, to below $1.
Meanwhile, intense demand for datacentre components — particularly semiconductors — continues to push costs upward. Some companies are reporting profitability under adjusted measures that exclude significant expenses. Anthropic recently told investors its adjusted operating income was positive, though the metric effectively omits many of its costs.
The most significant risk may be what financial analyst Groundbreaker calls the "$1.5tn compute commencement wall" — a surge in take-or-pay datacentre obligations expected over the next two years. Under these contracts, payments are deferred until a datacentre comes online, often two to three years after signing. During the interim, the builder books the contract value as future revenue while the buyer, typically a frontier AI lab, does not yet record the associated costs.
Groundbreaker's analysis suggests costs jump by approximately $700bn next year and exceed $800bn in 2027 as contracts mature and facilities begin operating. These are not technically debt, but failure to meet the obligations would transmit shockwaves through the financial system.
The dynamic bears comparison to the expiration of subprime "teaser" mortgage rates in 2007 and 2008, when borrowers suddenly faced much higher payments and defaults accelerated into the global financial crisis.
Whether the model holds depends largely on sustained revenue growth outpacing costs — or on the absence of cheaper alternatives from competitors, including Chinese providers. If end users cannot or will not pay enough to cover the bills, the interlinked financial structure underpinning the AI boom could face severe stress.
