Cheap Chinese and open source LLMs could trigger trillion-dollar AI infrastructure 'reset'

A financial analyst has raised concerns that the advent of low-cost Chinese AI models and open-source alternatives could prompt a significant "reset" in the AI landscape, as major cloud providers grapple with recouping billions in capital expenditures. According to Jefferies research, approximately half of the $2.3 trillion contracted backlog at Microsoft, Alphabet, Amazon, and Oracle is tied to OpenAI and Anthropic's offerings.

However, Chinese models and open-source alternatives are poised to undercut these prices, potentially undermining the revenue streams of the US leaders. Research from OpenRouter indicates that Chinese models processed 76.12 trillion tokens in late September, more than quadrupling the 4.37 trillion tokens processed by the top four US providers in late April.

Moreover, Moonshot AI's open-source Kimi K3 model is reportedly performing at or near the performance of Anthropic's Claude Fable 5. Jefferies questions whether these market dynamics have been fully incorporated in current valuations. The hyperscalers collectively plan to spend between $600 and $800 billion this year on datacenter construction, often financed through debt rather than cash.

The analyst queries the future revenue sources for these companies, given the ongoing decline in token pricing and the reduced profitability of large language models. While the long-term prospects for AI remain promising, the medium term presents considerable uncertainty, according to the report. As the industry continues to invest heavily in AI infrastructure, questions arise about the sustainability of these investments and the potential for new revenue streams to offset the significant expenditures.