Warnings are emerging about the financial stability of the AI industry, with concerns centred on the scale of debt being used to fund datacentre expansion and the economics of AI services. Hyperscalers like Google, Amazon, Microsoft, Meta, and Oracle are estimated to have issued $132bn (£99bn) in debt this year alone for datacentre rollouts.
Simultaneously, the "unit economics" of AI are reportedly moving in the wrong direction. An index by Silicon Data indicates that the price customers pay for a million tokens, a unit of data processed by large language models, has more than halved since June to less than $1. This contrasts with the elevated costs of real-world datacentre components, such as semiconductors, driven by high demand.
Financial analyst Groundbreaker has highlighted a potential $1.5tn "compute commencement wall" for AI labs over the next couple of years. This refers to future obligations from "take or pay" contracts for datacentres, where costs are not due until the facilities become operational, often two to three years after construction begins. The analysis suggests an abrupt jump in costs of $700bn next year and over $800bn in 2027 as these contracts mature.