Yin Qi Becomes Chairman of StepFun: Neither 2B Nor Pure 2C Works for a Foundation Model Company
A foundation model burns at least 3 billion RMB a year and has to be sustained for three to five years, so neither 2B nor pure consumer software adds up; StepFun's bet is foundation model plus devices, trading software-hardware integration for a seat at the table.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Neither 2B nor pure 2C works for a foundation model company
Yin Qi lays out the business model by elimination: doing foundation models means at least 3 billion RMB or more of investment a year, sustained for three to five years, with at least 100 billion RMB of basic R&D investment. On that premise, doing 2B at the same time does not hold, because 2B monetizes slowly and has a limited ceiling on revenue and profit; doing pure consumer software applications at the same time does not hold either, because that is the domain existing internet giants are good at — they have users, data and a data flywheel. So StepFun chose a consumer scenario leaning toward software-hardware integration, doing some 2B services in between, and ultimately landing on AI plus devices.
— Yin QiStepFun's bet is foundation model plus devices
Yin Qi splits the divergence among large model companies into two layers: the first layer is whether to do foundation models at all, which is the biggest bet, and staying the course over the next three years will be very hard; only the second layer is the commercialization path. StepFun's path is to let devices pull the model, because bigger is not always better — Gemini Flash proved that too — and efficiency, cost and specialized capability matter just as much. So StepFun does not obsess over topping a single benchmark, but over whether the model architecture fits future device scenarios.
— Yin QiCoding is the best reinforcement learning environment, but it earns nothing
Yin Qi says that Chinese large model companies are basically all grinding on coding right now. Coding is valuable for raising intelligence, because it is a complex scenario with outcomes and right-or-wrong judgments — the best reinforcement learning environment. But coding itself has very limited value from a commercialization standpoint: the giants do well here, open source has DeepSeek pulling the water level very high, aggressive competition among the giants will crush the price, and coding is closely tied to its downstream environment. So this track, like Chatbot, means nothing for a new company.
— Yin QiChatbot is not a long-term product form
Yin Qi sees Chatbot as a stage of results, and possibly not even something at the level of search. The reason is that its interaction is very unnatural: putting AIGC capability inside a dialog box has, factually, not been very successful. He judges that replacing search with large models is a certainty, but what the product form will be is still unclear. Chatbot may be a stage product overseas, but its moat is not solid — Gemini's DAU has risen extremely fast, and right now nobody has a particularly large barrier.
— Yin QiAutonomous driving will be highly concentrated; the cockpit will return to in-house development
Yin Qi judges that autonomous driving and the cockpit are different: the cockpit is highly personalized, tied to the human-machine interaction experience, and has a lot of room for value-added services in the future, so carmakers will develop a high proportion of it in-house; autonomous driving is a safety component and will return to the hands of Tier 1s, as safety components did with Bosch back then. Because margins are not high and R&D investment is very high, the amount that can be charged per car is limited, so scale is mandatory, and in the end there will be around three core suppliers — Huawei is certainly one, there may be two other comprehensive suppliers, plus suppliers that small and mid-sized carmakers need; it is not necessarily two plus one or two plus two.
— Yin QiSmart people are not scarce; those who persist long-term are
Yin Qi says that among the 40 people Megvii benchmarked against back then there were a pile of OI gold medals, IPHO gold medals and math gold medals, but smartness comes in two kinds: one kind can collaborate in a team, the other is strong on its own but hard to collaborate with. What he values more is smart people who can persist long-term, because smart people have many choices, and being smart gives you many shortcuts — unless there is a mission to complete, why not take the shortcut. The second point is to use the correct dumb method to do one thing over the long term; smart people tend to want the quick way, which actually makes it easy to go wrong. So he thinks being smart is not that important — you also need business sensitivity and wisdom.
— Yin QiAI organizations must fuse top down and bottom up
Yin Qi believes every generation of enterprise is the most leading organizational form of its stage, and stands out by fitting the technology and business of the time. From manufacturing to ICT it is broadly top down, internet companies are broadly bottom up, and the AI system needs a fusion of the two: because any big project is a multi project, consuming enormous resources and requiring concentrated effort, yet at the same time the organizational culture and talent tolerance need bottom up vitality. He also judges that future AI leaders may have to be ten or twenty thousand people — but ten or twenty thousand people with extremely high talent density.
— Yin QiThe deciding factor is technology that holds up plus an application that closes the loop
Yin Qi poses a math problem: technical R&D starts at 3 billion a year — he predicts 3 to 5 billion — and over five years that is a floor of 15 billion RMB of investment; conversely, three to five years from now, can there be a core application that gives you the possibility of 3 to 5 billion RMB of profit a year. Both things must be satisfied to have a chance of reaching that stage. He also says China's 2B market is essentially a 2G market, covered outward from government and central and state-owned enterprises, while small and mid-sized B is more like C, so in China you cannot be 2B.
— Yin QiIn their own words · checked verbatim
If you run a foundation model company and at the same time do 2B, I think that definitely does not hold.
如果做基模公司，然后同事做2B，我觉得就肯定不成立
Yin Qi17:13
Smart people actually have many choices; it is the smart people who can persist long-term that matter more.
聪明的人其实有很多选择，就是长期能坚持的聪明人，才是比较重要的
Yin Qi1:15:40
Smart people first have to persist long-term, and second have to use the correct dumb method — both are quite hard, so actually I think being smart is not that important.
聪明人第一要长期坚持，第二要用正确的笨办法来做，其实都挺难的，所以其实聪明我觉得没有那么重要
Yin Qi1:16:42
Figures
| Floor for annual foundation model investment | 3 billion RMB or more | 18:16 |
| Basic R&D investment in foundation models | At least 10 billion RMB | 18:16 |
| Predicted annual technical R&D investment | 3 to 5 billion RMB | 1:39:59 |
| Predicted number of core autonomous driving suppliers | Around three | 49:30 |
| Time horizon for embodied AI | Around five years | 12:11 |
| Medium-to-long-term time for collecting physical space data | Five to seven years | 35:27 |
| Salary increase for AI talent | Five to ten times | 28:22 |
Glossary
- VLA / vision-language-action model
- An architecture that unifies vision, language and action execution into a single model.
- foundation model
- A general model pretrained on large-scale data that can be adapted to many downstream tasks.
- edge / long board
- Yin Qi's term for a genuinely differentiated leading position in technology.
- MVP / minimum viable product
- The practice of validating a product hypothesis with the smallest possible feature set.
How to listen
Founders and investors watching the commercialization path of large models, AI devices and the autonomous driving landscape, especially anyone who wants to hear how a foundation model company does its math.
After 1:42:27, the part about Yao Class, the cold bench and personal temperament can be fast-forwarded.