Lee Kai-Fu (李开复) spent decades telling founders how to win before trying to win as one himself. In May 2023, at 71, he launched 01.AI (零一万物) and promised a model built "world's No.1." Eighteen months later he moved most of his training team into Alibaba (阿里巴巴) and began selling software to chief executives.
Key takeaways
- Lee, a former Microsoft (微软) and Google (谷歌) vice president who founded Innovation Works (创新工场) in 2009, started 01.AI in May 2023; it reached unicorn status above US$1B within eight months.
- 01.AI released Yi-34B on 6 November 2023, which topped both the Hugging Face open-model leaderboard and the C-Eval Chinese benchmark; Yi-Lightning (October 2024) reached 6th worldwide on the LMSYS arena, first among Chinese models.
- In early 2025 Lee moved most of 01.AI's pretraining and AI-infrastructure teams into Alibaba Cloud through a joint "industry large-model lab," and stopped training trillion-parameter models, citing costs beyond a startup's reach.
- The new target is enterprise decision AI: in July 2026 01.AI launched the "No.1 decision-maker AI (一号位AI)" suite — boss AI, sales-champion AI and investor AI.
- Lee's reframed thesis: only the largest companies can sustain frontier pretraining, so startups should build applications on top of it.
The investor who became the founder
Before 01.AI, Lee's name meant venture capital. He built Microsoft Research Asia in the late 1990s, led Google's China business, and in 2009 founded Innovation Works, an incubator that backed AI 1.0 winners such as Megvii (旷视), Momenta and Horizon (地平线). That history is why 01.AI assembled fast: a team drawn from Google, Microsoft, Baidu and BAAI, with a mandate to chase the global front tier from the first line of code.
The company was incorporated on 16 May 2023 in Beijing. Lee's framing at the time was unusually absolute — he described the goal as becoming the world's number one, not a credible domestic alternative. For a man better known as a communicator than a builder, the stance raised the stakes considerably.
The Yi moment
Yi-34B arrived on 6 November 2023 with a genuinely strong claim. At 34 billion parameters — what Lee called the "golden ratio" size for open models — it led both the Hugging Face English open-model ranking and the C-Eval Chinese benchmark, a dual crown no domestic model had held before. Its 200K context window could process roughly 400,000 Chinese characters, and the smaller Yi-6B gave researchers a free, commercial-friendly entry point.
The run continued. Yi-Lightning, a 2024 flagship, reached 6th place worldwide on the LMSYS Chatbot Arena blind test in October 2024 — the best result any Chinese model had posted there. For a brief window, 01.AI looked like the proof that a Chinese startup could play the frontier game on cost-efficient terms.
The retreat
The frontier game turned out to be unaffordable. In early 2025 01.AI formed an "industry large-model joint lab" with Alibaba Cloud and moved most of its pretraining and AI-infra teams into it. Lee told local media that only the largest companies can sustain ultra-large model training, and that 01.AI would stop building trillion-parameter models. Reporting from The Paper described it as the first model company to abandon super-large-model training; the National Business Daily quoted Lee calling the preceding period one of the three darkest stretches of his career.
The cause was arithmetic, not doubt about the technology. Training a trillion-parameter model once can cost tens of millions of dollars, and open-source rivals were closing the gap faster than any startup could price in. DeepSeek's open reasoning model, released weeks later, made the point for him.
Selling to the corner office
Rather than chase parameters, 01.AI spent 2025 reorienting toward paying customers. In July 2026 it launched the "No.1 decision-maker AI (一号位AI)" suite: a boss AI for chief executives, a sales-champion AI for revenue leads, and an investor AI for capital-allocation decisions. The product lead said internally the tools lifted order value roughly fivefold and doubled conversion — a vendor's claim, not an audited one.
The repositioning mirrors Lee's own shift in language, from AGI idealism in 2023 to "industrial AI" and profitability by 2026. In a three-year internal letter he even renamed the firm's identity away from the "little tigers" label applied to top model startups, toward a profit-first posture.
What it signals
Lee's arc is the cleanest available evidence that China's foundation-model business split into two tiers: the hyperscalers who can fund pretraining indefinitely, and the application firms who must ride on top. 01.AI did not fail — it changed the question it answers, from "can we build the biggest model?" to "can we sell the smartest seat in the room?" For observers, the lesson is that a unicorn valuation and a frontier model are no longer the same achievement.
Honest limitations
This profile relies on 01.AI's own site and timeline, plus reporting from The Paper, the National Business Daily and China Daily. The unicorn valuation, Yi leaderboard positions and the July 2026 product metrics are drawn from company statements repeated by those outlets; 01.AI is private and files no public financials, so revenue, headcount and the "fivefold order value" claim are unverified. The LMSYS 6th-place and C-Eval/Hugging Face dual-crown results are reported leaderboard outcomes, but rankings shift and I did not re-run them. Lee's quoted remarks about training costs and the "darkest period" come from media interviews, not primary transcripts I retrieved. Where only a single source existed, I omitted the number rather than assert it.
What readers can do now
- Separate "frontier model builder" from "model application vendor" when judging Chinese AI firms; the labels no longer mean the same thing.
- Evaluate 01.AI on named enterprise deployments and renewal rates, not on parameter counts or past leaderboard wins.
- Treat the 01.AI story as a case study in strategic retreat, not failure — useful context for where smaller model labs go next.