--- title: "Kai-Fu Lee bet big on open models — then rebuilt 01.AI around enterprise AI" date: 2026-10-03 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-kai-fu-lee-01ai language: en --- # Kai-Fu Lee bet big on open models — then rebuilt 01.AI around enterprise AI > How Kai-Fu Lee's 01.AI went from open-weight Yi models to a lean enterprise-AI play, and what the pivot says about China's startup economics. The boardroom fell quiet when the most famous name in Chinese AI admitted his company would stop chasing the biggest models. Outside, Beijing's startup scene was already muttering that the open-model dream had hit a wall. Few founders can say "I was wrong about the plan" and keep the room's respect — Kai-Fu Lee is one of them. **The man who named the era** Long before he ran a model lab, Lee (李开复) was the face of AI in China's public conversation. He ran Microsoft Research Asia in the late 1990s, then became Google's president for Greater China. In 2009 he founded Innovation Works (创新工场), the incubator that backed a generation of mobile apps. A Carnegie Mellon PhD, he had spent decades arguing that China would lead the applied-AI wave. So when he came out of semi-retirement in 2023 to start 01.AI (零一万物), the market treated it as a verdict, not just a startup. The name itself — "zero one," from binary — signaled a fresh start from first principles. **Yi and the open-model gamble** 01.AI's first headline came in November 2023 with Yi-34B, an open-weight, bilingual large model (大模型) that briefly sat at the top of Hugging Face's open-model leaderboard. For a company barely eight months old, it was a statement: China's startups could still ship frontier-class weights. Investors bought the story. Within months 01.AI had raised roughly **US$300 million** at a **US$1 billion** valuation, led by Alibaba Cloud. Lee talked about building the "Microsoft of the AGI era" — a company that would train its own flagship models and sell them to the world. **The January 2025 reckoning** Then, in the first week of 2025, Lee sat down with Chinese media and walked back the grand plan. 01.AI had formed an "industrial large-model joint lab" with Alibaba Cloud. Most of its pre-training and infrastructure teams would move to Alibaba. The company, Lee said, would no longer try to train super-large models from scratch. His reasoning was blunt. In interviews with LatePost and KrASIA, he argued that only technology giants can afford the bills for frontier training, that the scaling law was hitting diminishing returns sooner than expected, and that a startup burning cash to out-spend giants was a losing game. "A startup going head-to-head with big companies on who can burn more to build a bigger model will not succeed," he said. **What 01.AI kept** The pivot was not a shutdown. 01.AI kept a smaller team to build faster, cheaper models — Yi-Lightning, which Lee said cost a fraction of GPT-4o per token — and leaned into applications. Its global-facing product, PopAI (later branded 万知, Wanzhi), is a document and presentation assistant aimed at overseas users. Domestically, the company pushed tailored AI into finance, gaming, and legal verticals. Lee has described the repositioning as building "the Palantir of China" — selling bespoke AI to enterprises rather than racing OpenAI on raw capability. The economics are the point: he told Chinese media the company booked around 100 million yuan (≈ US$14M) in revenue in 2024 and expected to grow from there. **The DeepSeek quote that reframed the debate** Lee's most-cited line of 2025 came in a Reuters interview in Hong Kong. Asked how far China had closed the gap with the United States, he said: "Previously, I think it was a six to nine-month gap and behind in everything. And now I think that's probably three months behind in some of the core technologies, but ahead in some specific areas." He called Washington's semiconductor sanctions a "double-edged sword" — painful in the short term, but a forcing function for Chinese teams to innovate under constraint. The remark landed because it came from a founder who had just shrunk his own ambitions to fit that reality. ## The pivot in plain terms Lee's retreat from frontier training is not a story about one company. It is a live test of the question every Chinese AI founder now faces: when a single training run can cost more than a startup's entire war chest, where does a small lab create value? 01.AI's answer — sell bespoke AI to enterprises, ship applications, and stop measuring itself against the largest model — has become a template others study. The open-weight Yi line still proves Chinese startups can produce competitive research, but the business logic has shifted from "build the biggest" to "find the paying use case." For readers watching the sector, the marker to track is recurring revenue from real deployments, not parameter counts. ## Honest limitations Dates, funding figures, and the pivot narrative are drawn from reporting by Reuters, KrASIA, LatePost (via Sina), and company-adjacent trackers such as China AI Index, cross-checked against each other. The 2024 revenue figure is reported by secondary outlets citing Lee's interviews, not an audited statement. Quotes attributed to Lee come from published Reuters and Chinese-media interviews; translations are our own. We did not independently verify 01.AI's current headcount or unreleased product roadmap. The "Palantir of China" framing is Lee's own analogy, repeated in coverage, not our characterization of the business. ## What readers can do now - Follow 01.AI's product direction directly at 01.AI and the Wanzhi (万知) app to see how a Chinese lab monetizes applications instead of frontier training. - Read Lee's Reuters and LatePost interviews for his unfiltered reasoning on why Chinese startups are retreating from super-large models. - Track the Yi model family on Hugging Face to judge whether the open-weight line still evolves now that pre-training has moved to the Alibaba joint lab. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-kai-fu-lee-01ai Free to quote with attribution and a link to the original.