ChinAI #274: After raising 3 billion, who else can large model startups get money from?

Greetings from a world where…

my book Technology and the Rise of Great Powers comes out in one month! If you’ve been a paid subscriber to ChinAI for more than 6 months and you’d like a free book, please let me know via email, and I’ll make sure to mail you a copy once it’s out. Offer applies to the first twenty people who express interest!

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Feature Translation: After a 20 billion (RMB) valuation, who can China's large model companies still get money from?

Context: Is 20 billion RMB (~3 billion USD) the ceiling? In China’s LLM investment and entrepreneurship circles, that’s become the question that everyone is mulling over. Two of China’s large model startups, Zhipu AI and Moonshot AI, are currently trying to raise financing, with a valuation at that 20 billion RMB threshold. They are joined by MiniMax and Baichuan Intelligence, also in the first tier of large model startups, as well as ModelBest, DeepLang, and 01.AI in the second tier. As U.S. dollar-denominated funds (US venture capital firms) pull back on Chinese investments, this Aitechtalk article, authored by Jin Zhang, (link to original Chinese) asks: Who will pay for large models?

Key Takeaways: Large model startups have struggled to attract funding from state-affiliated funds and the strategic investment funds of Internet giants, which are two crucial sources of domestic investment.

• Government-affiliated assets have already invested in one round, based on regional locations: “First, the RMB fund entities that can invest from Beijing, Shanghai and Shenzhen have already invested in one round. The Beijing Artificial Intelligence Industry Association [北京人工智能产业联盟] has invested in almost all large model-related companies in Beijing, Shanghai's state-owned assets have invested in MiniMax, Ziyou Liangji (自由量级) and other companies, and Shenzhen-based VCs have invested in Baichuan Intelligence.”

• Apart from the state funds from first-tier cities, another possible source is local capital from second- and third-tier cities (e.g., Hangzhou, Suzhou), but “the local capital investment (from these cities) is often purposeful, that is, it requires the industry to be local. However, with the first-tier large model companies, no specific industry can be established in the local area, so the probability of capital from second- and third-tier cities making a move in 2024 is not high.”

• What about the big tech giants that have their own strategic investment funds? Tencent, Alibaba, Meituan, Xiaohongshu, etc. have all invested in some initial rounds, but there is hesitation about continued investment because the marginal cost of large model applications has stayed high even with more users, which goes against the mindset familiar to these Internet giants.

• Here’s some interesting details about how state-affiliated funds find it hard to keep up with the speed of developments in the large model space: “Many large model companies are looking for investment institutions in a targeted manner in the early stage of entrepreneurship, first targeting the top funds, and some state-affiliated financing simply cannot squeeze in with large model companies. For example, many RMB funds did not know about Stepfun [阶跃星辰] (a Shanghai-based start-up founded by ex-Microsoft engineers) before it officially appeared this year.”

Financing trends for China’s previous wave of computer vision start-ups (“The Four Little AI Dragons”) can provide a good reference point for this generation of large model startups.

• Zhang writes: “Mingfei Wu, a senior industrial investor and a witness of the previous generation of AI, told AItechtalk that he believes that today's Chinese large model companies can refer to the rhythms and scale of financing received by the four AI dragons. Zhipu's latest round took money from the Middle East, and among the previous generation of AI companies, taking Megvii and SenseTime as examples, they also turned to state-owned and Middle Eastern money after US dollar-denominated VC funds.”

• If we count up the total financing raised by SenseTime, Megvii, Yitu, and Cloudwalk (the four dragons), that equals 6.93 billion USD and 5.3 billion RMB. Then, compare that to the financing raised by the four first-tier large model start-ups over the past year and a half, which gives 2.8 billion USD and 2.5 billion RMB. Zhang concludes, “If we refer to the capital performance of the previous generation of leading AI companies, there will be at least 4 billion USD + 2.5 billion RMB left in the market for the first tier of China's large models in the future.”

Let’s leave you with this comparison to the U.S. context. Among people I talk to in the AI governance/tech policy/nat sec space, there’s this common misconception that the Chinese government will just throw unlimited funds at building next-generation AI systems. We see from funding trends with large model startups that this is just not the case. In fact, as this article finds, Chinese startups are still searching for the “patient capital” that their Western competitors have been able to attract; moreover, the most likely source of this patient capital is not from government-backed funds:

• From the article: “‘Fast capital’ evaluates the business model of large model companies and lets these companies begin to monetize; ‘slow capital’ supports a large model company regardless of cost, chases OpenAI, and serves as patient capital. This is like how OpenAI and Microsoft are tied together, but only Internet strategic investment funds are likely to consider this path, and the chances of state-owned/affiliated capital taking such risks are low. And Chinese Internet companies have not fully shown the same investment and courage as overseas large companies in the pursuit of ‘innovation.’”

Note: All names attributed in quotes are pseudonyms.

ChinAI Links (Four to Forward)

Should-read: Previous ChinAI issues that profile startups mentioned in the feature translation

Avid readers should be familiar with some of the startups mentioned in the two tiers:

• SuperCLUE benchmark evaluations for LLM start-ups in the first tier: Tsinghua University and Zhipu AI’s GLM-4; Baichuan’s Baichuan3; Moonshot AI’s Moonshot; Minimax’s models (ChinAI #264)

• On Shengshu-AI’s efforts to keep up with Sora (ChinAI #257)

• On ModelBest’s quest to industrialize large models (ChinAI #199)

Should-read: The Promise and Pitfalls of Government Guidance Funds in China

I’ve previously recommended this China Quarterly article by Yifan Wei, Yuen Yuen Ang, and Nan Jia. Their conclusion — “It is impossible for the state to completely replace the role of private investors in technological innovation because market-oriented VCs/PEs are still the most effective means of financing for high-potential ventures” — appears to hold true for large language models.

This part of their piece also speaks to the constraints described in the feature translation when it comes to getting investment from local governments:

Moreover, VCs/PEs must comply with strict investment terms imposed by state investors in GGFs. Funds established by local governments are normally required to invest in designated locations or sectors. Local governments also expect the amount of private investment to be two to three times that of GGF investment. Such a high ratio does not attract participation in GGFs or GGF-invested projects by leading VCs/PEs.

Should-read: Elsevier’s Insights 2024 Attitudes toward AI Report

Elsevier surveyed 3,000 researchers and clinicians across 123 countries about their willingness to use AI in their daily work, with clear differences in attitude between US, China, and India. One finding: 31% of researchers and clinicians around the world have used AI for work purposes, this is higher in China (39%) than in the USA (30%) and India (22%).

Must-read: Why We Remain Fascinated With Twisters

For The Ringer, Hal Sundt pens a winding piece about the movie Twisters, the Midwest, the explosion of students with meteorology degrees, a Japanese scholar (“Mr. Tornado”), and randomness.

Thank you for reading and engaging.

These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99