Jefferies: cheap Chinese and open source LLMs could trigger trillion-dollar AI infrastructure 'reset'

Jefferies questions whether large language models will ever be profitable amid a collapse in token pricing. Chinese and open source models processed 76.12 trillion tokens in a week versus 14.2 trillion for the top four US providers, while about half of the $2.3 trillion contracted backlog at Microsoft, Alphabet, Amazon and Oracle is owed by OpenAI and Anthropic.

• Chinese models hit 76.12T tokens in a week vs 14.2T for top 4 US providers
• About half of hyperscalers' $2.3T backlog is owed by OpenAI and Anthropic
• Hyperscalers to spend $600–800B on datacenters this year, increasingly debt-funded
• Bain: industry needs $6T in annual revenue by 2031, a $4.2T shortfall