Kai-Fu Lee is reshaping 01.AI around a simple strategic choice: stop competing in pre-trained large language models, and build services on top of Deepseek instead.
The former Google China head says Deepseek's open-source models have changed the market in China and created a direct challenge to OpenAI. For 01.AI, the shift is not a side experiment. Lee says Deepseek has become the company's primary bet.
01.AI moves away from proprietary model training
Lee's 200-employee startup previously trained its own proprietary large language models. That strategy has now been abandoned, according to Lee, who explained that 01.AI now relies entirely on Deepseek's open-source offerings.
The change followed a surge in demand for Deepseek models from Chinese CEOs in late January. Lee described Deepseek's release as a "ChatGPT moment" in China, with strong enthusiasm for AI applications and many Chinese hardware and software providers moving to support Deepseek models.
Lee put the decision plainly: "It became imperative for us to … embrace [DeepSeek] as our primary bet."
The new direction turns 01.AI from a company trying to own the full model stack into one focused on making Deepseek useful for business customers. That is a different kind of competition. Instead of spending heavily to build a pre-trained model from the ground up, 01.AI is concentrating on customization, deployment and performance for specific industries.
Why Lee sees Deepseek as a threat to OpenAI
Lee's argument is not only that Deepseek is popular. He says its free, open-source approach undercuts the business logic behind paid AI subscriptions and proprietary model access.
He frames the pressure on OpenAI in blunt terms. "The biggest nightmare for Sam Altman is that his competitor is free," Lee says. "I've already met many people who have canceled their ChatGPT subscriptions because Deepseek is free."
For Lee, the point is not a narrow comparison over which model performs slightly better. He argues that the deeper question is whether the OpenAI model can remain sustainable when an open-source competitor is available at no cost to users.
In a separate Bloomberg interview, Lee said pre-trained model development is already consolidating in both the US and China. He expects open-source approaches to win over time, while large language model pre-training remains in the hands of only a few major companies.
That view explains why 01.AI is no longer trying to stay in the pre-training race. Lee said: "A pre-trained model can only be justified by amassing hundreds of millions of users." He added: "So Alibaba can justify it, Google can justify it, DeepSeek can justify it and ByteDance can just justify it, but the rest of us can't."
Customization becomes the business
01.AI now plans to focus on adapting Deepseek models for corporate clients. The company is targeting the financial, gaming and legal sectors.
That does not mean Lee sees 01.AI's technical knowledge as obsolete. He argues that experience with large language models still matters, especially when companies need models to run well in real-world settings.
Lee described the relevant capabilities this way: "How do you train it, how do you tune it, how do you do reinforcement learning and how do you do fast inference? This last part can only be done by companies with LLM capabilities."
In practical terms, the company is shifting its value from owning a base model to helping customers get reliable results from an open-source one. The model may be free, but the work of tuning, adapting and delivering it for specific business uses remains a service opportunity.
Lee also projects stronger revenue. He expects 01.AI to reach 100 million yuan ($13 million) in revenue for the first quarter of 2025, matching the company's entire 2024 revenue. Even so, the company has not yet reached profitability.
The cost debate behind the strategy
Lee also points to operating costs as a major reason Deepseek matters. He said OpenAI reportedly spent $7 billion in operating costs during 2024, while Deepseek requires only about two percent of that amount.
That contrast is central to his criticism of proprietary AI economics. If one company needs far more money to operate while another can offer strong open-source models at much lower cost, the pressure shifts from model quality alone to business durability.
Lee said: "So the issue really isn't whose model is 1% better? I think they're all very good. But the issue is is OpenAI's model even sustainable."
He also described Deepseek as "infinitely lasting" because its founder has enough funding to maintain current operations and has reduced computing costs by a factor of 5 to 10. In Lee's view, that makes Deepseek a difficult rival for companies built around closed, costly AI systems.
OpenAI and Anthropic recently urged the US government to ban Deepseek models, calling the Chinese startup "state-controlled." Lee called that a "sign of paranoia."
His broader message is that the AI market is moving into a phase where only a small group can justify pre-training at scale, while many other companies may need to build on open-source foundations. For 01.AI, that future has already arrived.