Cheaper Chinese and open-weight AI models are putting pressure on the business case for the enormous data-centre build-out, according to analysis from Jefferies. The question is not whether AI needs computing power, but whether providers can earn enough from it at falling prices.
Watch Desk analysis
What happened
The Register reports that Jefferies estimates roughly half of the $2.3 trillion contracted backlog at Microsoft, Alphabet, Amazon and Oracle is tied to OpenAI and Anthropic. The analyst argues that cheaper models could put pressure on what those major AI developers can charge, and therefore on their ability to buy cloud computing capacity.
Jefferies points to OpenRouter figures showing Chinese models processed 76.12 trillion tokens in the week ending 27 September, compared with 14.2 trillion for the four largest US providers. It also notes that hyperscalers are expected to spend $600 billion to $800 billion on data centres this year, and questions whether their planned spending can be sustained if AI revenue falls short.
The Register also reports Bain’s estimate that the AI industry may need $6 trillion in annual revenue by 2031 to fund expected infrastructure demand. Bain sees a possible $2.7 trillion gap even after accounting for existing and emerging uses. These are analyst forecasts and arguments, not a settled reckoning of future revenues.
Why it matters
The AI infrastructure race rests on two things happening together: demand for computing keeps growing, and customers can pay enough to justify building it. If cheaper models increase usage while squeezing prices, those two sides may not move in step. That is a more useful question than simply counting new servers or tokens.
Our read
This is a serious challenge to the assumption that rising AI use automatically means rising returns for every company building the infrastructure. The figures make the case worth examining, not a verdict that the boom is about to collapse. Jefferies’ argument is that the price of intelligence may fall faster than the industry can find customers for all that capacity. A rather awkward mismatch, if it holds.
What to watch
- Whether falling model prices translate into lower cloud spending by AI developers.
- Whether hyperscalers disclose stronger returns from AI infrastructure investment.
- Whether new AI products close the revenue gap estimated by Bain.
Discussion spark: If cheaper models drive much higher AI usage but reduce what providers can charge, which matters more for the data-centre boom: token growth or revenue per token?
Sources and evidence
- Cheap Chinese and open source LLMs could trigger trillion-dollar AI infrastructure 'reset' – The Register (13 October 2026, 16:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.