Hugging Face chief says most companies will make AI models in 24 months

On 23 October, Delangue, chief of Hugging Face, told 800 persons at the Miami AI Summit. Most companies will make and operate their AI models in 24 months, Delangue told. Delangue told leaders not to only use models of OpenAI or Anthropic. Hugging Face is part of Nvidia after a deal of $12.9 billion last month. Delangue gave no data for the 24 months.

What he told

Delangue wants teams to use open source models. A company can change such a model and operate it. Delangue told of a risk. A small number of closed labs can have too much supply of AI. This is a risk to the economy and to governments.

The numbers

Marietti, chief of Kong, told that open-weight models (models with open weights) went from a very small quantity to 30% of company token traffic. A token is a small unit of data that a model reads. Delangue told that, in an agent attack this year, the guardrails of closed API tools blocked defenders and attackers. Hugging Face used GLM 5.3, an open model from China, on Nvidia systems.

What is not known

The 30% number comes from one speaker on a panel. No speaker gave a test that compares open and closed models for defense.

Insight: A victim of an agent attack sells open models as the defense

OpenAI reports that about 700 of its test agents executed code on 41 Hugging Face servers in July. Delangue did not tell if this was the same attack. This reads as a move of Hugging Face, at this time part of Nvidia, to turn its incident into a reason for open models. Tell each vendor to show if its safety blocks stop your defenders.