A Jefferies research note argues that the rapid adoption of cheap Chinese and open-source LLMs could force a 'reset' in the AI infrastructure boom. The analyst points to OpenRouter data showing Chinese models processed 76.12 trillion tokens in the week ending 27 September, up from 4.37 trillion in late April, while the top four US providers processed 14.2 trillion. The arrival of Moonshot AI's open-source Kimi K3, estimated to deliver about 95% of Anthropic's Claude Fable 5 performance, is cited as a frontier competitor.
The concern is that hyperscalers have built datacenter capacity on the assumption that OpenAI and Anthropic will buy compute at current prices. With token pricing collapsing, Jefferies questions whether LLMs will ever be profitable. The four largest hyperscalers plan to spend $600-800 billion this year on datacenters, increasingly funded by debt, and Jefferies questions whether they will really spend $993 billion on capex in 2027.
Jefferies also notes that the hyperscalers' recent earnings growth has come from investment gains, not operations. Annualized earnings rose by $222 billion to $581 billion in the four quarters to 2Q26, but 'other non-operating income/loss' increased by $224 billion to $237 billion, accounting for all of the increase. Separately, Bain & Company estimates the AI industry needs $6 trillion in annual revenue by 2031 to sustain infrastructure buildout, leaving a $2.7 trillion gap that would require entirely new product categories.