Cheap Chinese, open-source LLMs may force AI infrastructure reset

What happened

Financial services firm Jefferies has raised doubts about the long-term profitability of large language models, pointing to what it describes as an ongoing collapse in token pricing. According to reporting by The Register, the bank's analysts suggest that the rapid emergence of cheap Chinese and open-source LLMs is putting sustained downward pressure on the cost of running AI inference, raising questions about whether current levels of capital spending on AI infrastructure can be justified.

The note reportedly frames this as a potential trigger for a broader "reset" in AI infrastructure investment, given the scale of capital — running into the trillions of dollars globally — that has been committed to data centres, chips and compute capacity on the assumption that demand for proprietary, premium-priced AI models would remain strong and high-margin.

Why it matters

The story matters because it strikes at the economic assumptions underpinning the current AI infrastructure build-out. If token pricing keeps falling as open-source and lower-cost Chinese models close the capability gap with Western frontier models, the commercial case for some of the enormous capital expenditure committed by hyperscalers and chipmakers becomes harder to sustain at current trajectories. That has direct implications for how quickly, and on what terms, enterprises and public-sector bodies can expect access to affordable AI capacity for digital transformation programmes.

For leaders planning AI-enabled services, the signal is that the underlying unit economics of AI are shifting fast, and cost assumptions baked into multi-year transformation business cases may need revisiting well before those programmes conclude.

The Renascence take

Much of the commentary around this story will focus on winners and losers in the infrastructure race — who overbuilt, whose margins compress, whose capex gets stranded. The more useful lens for operators is what falling token prices actually unlock on the demand side.

A collapse in token pricing is not just a supply-chain or investor story — it is a behavioural one. When the marginal cost of an AI interaction falls towards zero, organisations stop rationing AI use and start embedding it everywhere: every customer query, every internal workflow, every edge case. That changes the experience design brief entirely, from "where can we justify AI" to "where should a human still be the default." Operators who treat this purely as an infrastructure or procurement question will miss the real work, which is redesigning service journeys and governance for a world where AI assistance is cheap, constant and easy to overuse.