Meta Chief AI Scientist Yann LeCun has weighed in on the growing dispute over AI model distillation, arguing that the economics point in one direction: frontier foundation models will eventually become free or open source.
In a post on X, LeCun said training a frontier model is expensive, but distilling one is cheap. 「Just this market force tells us that frontier foundation models will eventually become free or open source,」 he wrote, adding that there are 「other reasons」 too. The post received more than 500 likes and about 46,000 views.
Why distillation matters for closed-model economics
Distillation refers to training one model on the outputs of a more capable model, allowing the second model to learn similar capabilities at a much lower cost. LeCun’s argument is straightforward: if distillation remains far cheaper than training from scratch, then expensive closed frontier models may struggle to preserve both a performance lead and pricing power over time.
Under that view, the moat does not sit permanently in the frontier model itself. Value would shift toward other layers, including data, applications, and services.
Tapestry and the open-source path LeCun is backing
LeCun also included a link to Tapestry, a project under the AI Alliance where he serves as chief scientific adviser. According to the project website, Tapestry is an open-source platform meant to let organizations around the world jointly train frontier AI models while keeping control of local data. Participants share model weights, while sensitive data stays on-site, and each participant can build fully owned derivative models.
The website’s timeline says Tapestry completed its proof-of-concept stage on Sept. 1, training three models across India, Australia, and the United States. The project aims to train its first base model from scratch by the end of this year, move into industry and government deployments in the first half of next year, and begin pursuing frontier-scale training in summer 2027. The site also says that the infrastructure, data pipelines, and design decisions behind frontier models remain concentrated in a small number of companies and regions.
Competition or theft?
LeCun’s comments arrived at a moment when the distillation debate is especially heated. According to CNBC, U.S. Treasury Secretary Bessent described distillation as 「theft」 in July and threatened sanctions against overseas companies that use distillation to extract capabilities from U.S. models. This month, the Cybersecurity and Infrastructure Security Agency, or CISA, also accused Chinese AI companies of carrying out 「industrial-scale knowledge distillation operations.」
Anthropic said it had found Alibaba and DeepSeek engaged in 「illegal distillation,」 an allegation denied by the Chinese side. Chain News had also previously reported that OpenAI disclosed a distillation attack and pointed to Moonshot AI, the developer behind Kimi.
Nvidia CEO Jensen Huang has taken a different line. Speaking to CNBC, he said: 「That’s called competition.」 He added that companies are free to test other products as much as they want, and that if a provider does not want others using its product, it should 「know your customers, and then stop service.」
Market pressure versus regulation
LeCun has long backed open-source AI, and this latest comment fits that position. But his argument rests on an important condition: distillation must remain unrestricted. If the U.S. government treats distillation as theft and limits it through sanctions, export controls, or service terms, then the market pressure pushing frontier models toward openness could be slowed.
What comes out of this dispute could shape AI business models. If LeCun is right, model capabilities may spread quickly and competition may move toward applications, data, and distribution. If regulation succeeds in building higher walls, frontier models may remain concentrated in the hands of a small number of companies. With Chinese open-source models still catching up and the U.S. tightening controls, how distillation is classified may become one of the factors that shape the industry’s next map.