AI does not make money in the clean way investors want to believe. The revenue is real, but much of it is money moving among the same companies that fund, build and sell the computing infrastructure. It is a subsidy loop with very expensive running costs.
One deal shows the loop. Amazon will invest up to $25 billion in Anthropic, an AI model maker, which will spend over $100 billion on Amazon cloud technology over ten years. That creates reported demand for Amazon’s cloud business, with the seller of the infrastructure helping to fund the buyer. Google is moving the same way and plans to invest up to $40 billion in Anthropic.
The cloud company books that demand while it still pays for data centres, chips, power and cooling. So reported revenue can grow before customers have paid enough to cover those costs.
Every request carries a cost, because an AI model has to run each time it answers, writes, searches, codes or makes video and other media. Research on AI energy use found that the power used varies with the size of the model, the hardware and the length of the answer. Larger models can cost more every time they are used.
The pressure is visible at OpenAI, the company behind ChatGPT. It says it has discontinued the website and app for Sora, its video generator. Video sits at the expensive end of AI because making it takes heavy processing. Shutting a heavy product before the numbers work is a cost decision.
Costs reach down to chipmaking. Helium prices rose after a halt in liquefied natural gas production in Qatar, which produces nearly one third of the world’s helium. Chipmakers use helium for cooling, leak detection and precision manufacturing. Industry executives have warned that manufacturers may need to slow or shut down chip production if the shortage continues.
Google, Amazon, Microsoft, Meta and Nvidia can absorb losses for longer than most firms. Higher input costs still shorten their runway, because the same AI request costs more to answer when chips, power and helium get more expensive.
The long-term hope behind those losses is AI that completes multi-step work and AI with human-level ability. Those could create a stronger revenue base if customers pay for finished work instead of paying for each use.
