China's LLM Funding War: The Quick Death of Romance

By Lili Yu

Edited by Jing Liu

A Capital War in the Shadows This was a sudden shift in fundraising.

In China's prolonged large language model battle, Moonshot AI — founded by Zhilin Yang — was frequently mentioned but had not actually held the upper hand in financing.

Until late 2023, MiniMax (founded in 2021) and Zhipu AI (founded in 2019), by virtue of "starting earlier and having relatively mature LLM products," remained China's two most valuable LLM startups, while Baichuan Intelligence, Moonshot AI, and 01.AI followed behind.

The 2024 Lunar New Year marked a watershed. In mid-February, news suddenly broke that Moonshot AI had "completed a new funding round exceeding $1 billion, reaching a $2.5 billion valuation." Though this was quickly corrected to $800 million, An Yong Waves learned that $800 million was actually a conservative figure, "not including additional investments from existing shareholders or institutions whose investment processes remained incomplete due to the holiday."

While LLMs are inevitably an expensive game requiring continuous fundraising, most companies typically raise "two or three hundred million to get by." In October, Baichuan Intelligence's single round raised $300 million; Zhipu AI's cumulative 2023 funding exceeded 2.5 billion RMB (approximately $348 million). Now, barely a year old, Moonshot AI had not only pushed single-round funding to unprecedented levels but also become one of the highest-valued unicorns in the arena.

"This was clearly a deliberate, strategic move," one LLM investor told An Yong Waves.

In some investors' view, while Zhilin Yang's "strengths are very strong" — his research and technical talent are exceptional — his weaknesses are equally apparent, such as "lacking business and management experience." What ultimately enabled Alibaba to make such a heavy bet, beyond strategic alignment, came down to one person.

She is Zhang Yutong, managing partner at GSR Ventures. Multiple industry sources told An Yong Waves that Zhang was a crucial facilitator behind Moonshot AI's fundraising. Interestingly, GSR Ventures itself did not directly invest in Moonshot AI — though they had invested in Yang's previous company, Recurrent AI, and the firm's iconic figure, managing partner Allen Zhu, had recently told Tencent News in no uncertain terms: Chinese LLM companies have no future, and he won't invest in any of them.

Zhang and Yang's connection runs deep. Before Moonshot AI's founding, Zhang — also a Tsinghua graduate — had already invested in Yang's previous venture, Recurrent AI.

To convince a tech giant to shift strategy and back you heavily, and to devise an ideal transaction structure, are no easy feats — one LLM investor attributed much of this to Zhang's behind-the-scenes coordination.

In 2013, GSR Ventures hosted a Chinese student entrepreneurship competition at Stanford. Among over twenty participating teams, Zhang and her colleague invested in only one person: the future founder and CEO of Xiaohongshu, Mao Wenchao. An AI investor who participated in LLM startup fundraising cited this experience to emphasize that Zhang — the second person after Allen Zhu to advance to managing partner under a performance-based mechanism — "has seen the big leagues."

Another investor who participated in MiniMax's funding saw Moonshot AI's key strategy for extracting more money from Alibaba as "striking first" through "active, significant compromise."

After all, more funding means more dilution. In this investor's view, Moonshot AI's willingness to compromise — something companies prioritizing independent development typically reject — included accepting that "Alibaba would become Moonshot AI's single largest shareholder with 40% ownership," with much of the funding "paid in compute credits, ultimately becoming Alibaba Cloud consumption."

But this characterization was denied by a Moonshot AI investor. "The notion that Alibaba took control of Moonshot AI during fundraising simply didn't exist," they told An Yong Waves. "Moonshot AI negotiated away several harsh terms one by one," specifically including "AB share structures that preserve the team's absolute decision-making and control rights."

To this investor, the round represented Yang "deeply understanding that pursuing AGI requires enormous compute and capital" — essentially a "rational choice" to "acquire these resources at a reasonable cost."

Moonshot AI's new funding round clearly created urgency for other LLM companies, particularly MiniMax, which was conducting its own new round.

An Yong Waves learned that MiniMax's new Alibaba-led round of over $600 million has been substantially completed, with post-money valuation also around $2.5 billion. A MiniMax investor told us that the new round is ongoing, "and the two rounds combined will exceed $1 billion."

However, we understand that Alibaba's investment amount is significantly less than what it committed to Moonshot AI. Of course, one informed source revealed, "Alibaba's equity stake in MiniMax is also below 20%," though neither MiniMax nor Alibaba has officially announced this funding.

One dollar-fund AI investor believes MiniMax didn't take more from Alibaba due to different fundraising strategy — perhaps they wanted a "more balanced equity structure." After all, choosing one giant means forgoing others; moreover, for companies under heavy giant control, capital exit pathways are typically unclear, inevitably increasing difficulty for many financial investors to participate.

With both companies similarly focused on consumer applications and both using dollar-denominated structures, MiniMax has long been viewed as Moonshot AI's most direct rival.

Compared to Moonshot AI, MiniMax has always been an enigmatic company. Its founding team remains vague public figures. The only interview available online came from vice president and open platform head Wei Wei.

In fact, before founding MiniMax, founder Yan Junjie was already a prominent figure at SenseTime. To one investor who participated in MiniMax's funding, Yan was arguably SenseTime's most formidable intern ever: in four years, rising from intern in 2015 to SenseTime vice president and general AI technology head in 2019.

During this period, he not only built SenseTime's deep learning toolchain and foundational algorithms, and advanced general AI technology development, but also constructed SenseTime's facial recognition and smart city technology systems.

An LLM company's leader must have technical authority to command respect, and Yan fit this standard: in deep learning and computer vision, he published over 100 top-tier conference and journal papers, with Google Scholar citations exceeding 10,000. At the end of 2021, Yan founded MiniMax with his Chinese Academy of Sciences alumnus Yang Bin and others.

One LLM investor told An Yong Waves that MiniMax's early fundraising was exceptionally smooth. Particularly its first 2021 funding round: "from formally deciding to raise to finalizing allocations took about a week." In April 2023, MiniMax was reported to be raising a new round at a valuation exceeding $1 billion. After this round, MiniMax's investor list included not only the rarely seen miHoYo, but also capital giants Hillhouse Capital, Hongshan, and IDG (which had previously invested in SenseTime), plus Tencent — revealed for the first time to have made an LLM investment.

MiniMax was China's first LLM startup to reach a $1 billion valuation, and in 2023 was the only company besides Shein that Shanghai headhunters said could poach talent from ByteDance. Indeed, those joining MiniMax in 2023 included Zhang Qianchuan, former head of user products at Toutiao.

Coincidentally, multiple investors revealed that MiniMax's fundraising also closely involved a woman: MiniMax (Xiyu Technology) co-founder and COO Yun Yeyi. Publicly available information about her is virtually nonexistent, but An Yong Waves learned that before founding MiniMax, she worked in SenseTime's CEO office handling strategy and related business.

Multiple sources close to Yan Junjie indicated that "Yan is a typical tech geek, soft-spoken, probably not skilled at external-facing work," so the company's external fundraising, plus some management and external activities, were typically handled by Yun.

In one early MiniMax investor's eyes, Yun Yeyi, born in 1994, is "capable, has presence, strong execution," with "a maturity beyond her years." Like Yan, SenseTime was Yun's first job after graduation. During her years there, as assistant to SenseTime co-founder Xu Li, she witnessed SenseTime's journey from startup through 12 funding rounds and $5.2 billion raised, growing into the world's highest-valued $10 billion AI unicorn before IPO.

To one investor who participated in both AI waves, those who experienced the frequent fundraising era of the "AI Four Dragons" era "inevitably carry some optimism." However, even if Alibaba's bet was smaller than for Moonshot AI, if successfully completed, Alibaba would still become MiniMax's largest shareholder. Beyond Alibaba, MiniMax's equity structure also includes miHoYo and Hillhouse Capital, each having invested over $100 million. One informed source told us that Hongshan's investment in MiniMax is also its largest bet on any single LLM company.

The "close combat" between Moonshot AI and MiniMax in fundraising demonstrates at least this: barely a year apart, the few foundational LLM startups that made it to the table must now dramatically converge in fundraising due to tightening capital environments and shifting giant strategies.

After a year of shakeout, the remaining Chinese LLM startups at the funding table besides Moonshot AI and MiniMax are: Zhipu AI, Baichuan Intelligence, and 01.AI. Except for Zhipu AI with its distinctly RMB-denominated structure surrounded by typical RMB financial investors and state-owned/government industrial funds, the others largely compete in the same pool for limited capital.

"2024 will be a critical elimination year for several LLM startups," one dollar-fund AI investor told us.

Compared to catching up on model capabilities, China's LLM funding crisis has clearly arrived earlier.

Alibaba's Fear and Ambition

Convincing a giant to shift strategy is no easy task — unless the giant actively seeks change. In a sense, Moonshot AI happened to catch such a window.

In one dollar-fund AI investor's view, Alibaba's sudden aggression in LLM investment stemmed from the return of Joseph Tsai — the once-dominant king of strategic investment — and Eddie Wu, a VC with dual "technology" and "investment" perspectives. This meant "more decisive, faster decision-making" and "more willingness to bet big."

Tsai, who earlier orchestrated acquisitions including Yahoo China and later participated in acquiring Ele.me, was the architect of Alibaba's investment business. Under his leadership, Alibaba began abandoning financial investments in 2013, moving toward strategic investments emphasizing business integration.

Before returning to Alibaba, Wu was already a professional investor: in 2015, he founded Yuanjing Capital outside Alibaba. By September 2023, Yuanjing had invested in over 150 deals concentrated in Series A and B rounds, across hardcore technology, industrial manufacturing, industrial digitization, and cross-border expansion.

Wu is most widely known for this understanding of early-stage investment: "Those who bear uncertainty early on must recognize and understand strange ideas, ideas that seem impossible."

The return of these two professional investors also brought obvious stylistic shifts to Alibaba's LLM investments. If previously Alibaba preferred more seasoned entrepreneurs in LLM startups — such as 01.AI's Kai-Fu Lee, Baichuan Intelligence's Wang Xiaochuan, or even Zhipu AI's Tang Jie — a clear shift now is heavy betting on younger founders like Zhilin Yang and MiniMax's Yan Junjie. This fully aligns with two underlying logics of Alibaba's current internal business adjustment: betting on "technology" and "young people."

This situation inevitably invites comparison between today's Zhilin Yang and Jiang Fan of years past.

Earlier, in Taobao's migration from PC to mobile, the technically prescient Wu sensed the tide's direction earlier and pushed Alibaba to make many advance arrangements. This included participating in acquiring Jiang Fan's Youmeng from Innovation Works.

Events proved that after Jiang Fan joined Alibaba, he brought explosive growth to mobile Taobao. The Youmeng case wasn't unique — Gaode, another Wu investment and acquisition, later became one of Alibaba's rare traffic entry points.

To the investor above, this AI wave represents a new platform-level traffic entry point battle. Alibaba's strategy of "gang-fighting with five LLM companies" has an upper bound that is clearly another "new traffic entry point layout." As for the lower bound, it can accomplish making its cloud service the compute infrastructure for numerous LLM companies. "Each will give some face to their shareholder and consume Alibaba Cloud — then Alibaba is the biggest winner."

To many AI investors, Moonshot AI's binding with Alibaba is a very natural outcome. "It's like OpenAI binding with Microsoft, Anthropic binding with AWS and Google," but the problem is China only has Alibaba.

Actually, in early LLM startup development, many internet giants were not stingy about spending to see the cards.

Before heavily investing in Moonshot AI and MiniMax, Alibaba had already invested in 01.AI, Zhipu AI, and Baichuan Intelligence; Tencent had invested in Zhipu AI, Baichuan Intelligence, and MiniMax; Meituan invested in Zhipu AI and invested in and acquired Light Year Beyond; Lei Jun not only personally visited MiniMax and Zhipu AI, but his Shunwei Capital also invested in Baichuan Intelligence, Zhipu AI, and MiniMax.

But as LLM startups' fundraising appetite further inflated and the cost of seeing cards grew, various players fell into different hesitations.

Regarding Tencent's post-New Year silence, some investors attributed it to "perhaps Tencent wants to focus on its own Hunyuan LLM," and "Tencent Cloud has never been as important to Tencent as Alibaba Cloud is to Alibaba."

One AI investor who participated in multiple LLM startups' fundraising told An Yong Waves that late last year, after game license approvals tightened, Tencent held an internal review meeting. The gist was to tighten unnecessary investments: if new money went out, they needed to calculate whether it could outperform buying back Tencent stock. In his view, this may be why Tencent later didn't actively invest in LLMs. Additionally, following Tencent's historical pattern across gaming, video, music and other sectors, Tencent isn't the aggressively proactive type — but can still catch up later.

As for Meituan and Xiaohongshu, one dollar-fund AI investor revealed that in Moonshot AI's new round, they had intended to participate in strategic investment, but "because Alibaba raised Moonshot AI's valuation by 50% and took a heavy 40% stake, they ultimately gave up." Of course, another AI investor told us, "Xiaohongshu will still invest in Moonshot."

By comparison, ByteDance with "Doubao" and Baidu with "Wenxin Yiyan" have been essentially inactive in investing in LLM startups.

One investor who participated in MiniMax's fundraising told An Yong Waves that before June last year, ByteDance had intended to invest in MiniMax, even spending considerable time haggling over equity percentages with another major tech company "down to decimal points," but ultimately withdrew possibly due to regulatory pressure concerns. Beyond this, ByteDance also withdrew from StepFun, founded by former Microsoft global partner, VP, and former Microsoft Research Asia chief scientist Daxin Jiang, which it had also intended to invest in.

Only Alibaba struck hard. To many investors, this is also because Alibaba feels the most crisis.

A few years ago, when Alibaba was valued at $800 billion, Microsoft and Alibaba were roughly at the same level. Recently Microsoft reached $3.12 trillion, surpassing Apple to again top global rankings. Many industry figures believe that beyond maturity of software markets, the biggest difference behind Microsoft Cloud's rapid growth versus Alibaba Cloud's stagnation is that Microsoft has OpenAI. Microsoft once revealed in an earnings call that of its current 53,000 Azure AI customers, one-third joined in the past 12 months, with AI-related business (such as OpenAI API services, model training and fine-tuning, and other MaaS services) contributing six percentage points of growth.

In a sense, Alibaba's desire to find its own OpenAI has never been stronger — this is the root of its heavy bets on Moonshot AI and MiniMax. Moreover, these investments are simultaneously orders; the training and inference consumption brought by invested LLM companies can directly reflect in AI-related cloud business growth.

Of course, Alibaba's heavy equity investment in Moonshot AI also revives many memories of the mobile internet era. Under Tsai's push, Alibaba gradually increased its equity stakes in invested companies from 20% to 30% or more. Many of its important investments, including Gaode and UC, followed a trajectory from minority equity investment to commercial cooperation to full acquisition — inevitably prompting speculation about future possibilities between this AI wave's companies and Alibaba.

As for whether this wave of LLM startups can ultimately compete with giants, Qu Kai, founder of 42章经, offered an interesting formulation in a podcast dialogue with family office practitioner Mo Jielin. In Mo's view, essentially giants are playing Snake while global startups play Tetris — so for startups to win, they must fill spaces where giants won't go or execute poorly. The key is to be fast enough — in fact, if giants can learn to make themselves smaller and faster, they can similarly act like startups.

Romance Lasted Less Than a Year

Moonshot AI's new funding round, in a sense, also means Chinese LLM company investment is bidding farewell to the previous coexistence of financial investors and giants, formally entering a giant-dominated era.

But in this AI wave, traditional venture capital simply cannot compare to its role in the early internet rise.

Though they haven't been absent from this wave — indeed they were among the fastest to react. One FA who participated in several LLM startups' fundraising told An Yong Waves that in this AI wave, mainstream VC reaction was undoubtedly fastest, followed by internet strategic investment, then state-owned or government industrial funds.

Early dollar VCs participated considerably, but when LLM startups subsequently needed massive capital injection, they could hardly find those free-spending VC institutions of the past.

Rapidly inflated valuations are one reason. Even "quick-draw" Allen Zhu said in a recent interview that he didn't talk to MiniMax because "it was expensive from the start."

Second, VCs are no longer what they once were. One investor who participated in both AI waves told An Yong Waves that the direct reason for VC caution is exits dampened by weak secondary markets, and worsening fundraising difficulty. This capital predicament also relates to declining capital activity across the entire market. "If this were 2020, it would be a completely different situation — think how much autonomous driving raised."

But the core issue remains the persistent business model controversy around Chinese LLMs. Allen Zhu represents a typical skeptical camp: LLM companies lack scenarios, lack data, and will always face competition from open-source models — so the risk is enormous.

Looking at current commercial exploration by several companies, everything has just begun.

Zhipu AI, with its RMB-denominated structure and insistence on localization, has relatively firmly chosen the B2B route. One investor who participated in multiple Zhipu AI rounds said that while this route isn't as sexy as B2C, it is "most pragmatic, can provide steady flow." It has first-mover advantage, able to provide services early when many consumer products haven't launched. Additionally, "private deployment and customization" — the dirty, tedious work — is typically not where giants want to invest heavily, giving startups breathing room.

MiniMax and Moonshot AI are clearly exploring more on the consumer side. MiniMax has moved faster on application commercialization: its overseas AI character-playing product Talkie previously grew faster than Character.ai; later it launched domestic AI chat + gacha app Xingye and AI dialogue robot product Hailuo Wenwen. Moonshot AI emphasizes product strength combined with LLMs, releasing its Kimi Chat intelligent assistant supporting 200,000 Chinese character input last October.

Due to different emphases, some jokingly characterize these three LLM companies' commercial exploration as: Zhipu AI is most pragmatic, more like Allen Zhu's market-believer camp; Moonshot AI is more romantic, more like Zhilin Yang's technology-believer camp; and MiniMax sits right in between.

One LLM investor believes that at a 20x price-to-sales ratio, for these LLM companies to reach $3-4 billion valuations means at least $200 million in revenue. Yet among current companies, even those doing more stable B2B business have only about 150-200 million RMB in revenue, including portions not from LLMs.

Perhaps therefore, LLM startups have become a rare风口 where exit is discussed from birth. Li Gangqiang, author of Invest Well, Exit Well, also believes after Allen Zhu's interview that early-stage investors should quickly exit foundational LLM investments: "Though LLM companies are still early on the industry curve technologically, the capital cycle curve may likely be reaching market peak."

Of course, this isn't unique to China. Even in Silicon Valley, this AI wave shows tech giants crowding out VC institutions. Li Guangmi, founder of Shixiang Technology, pointed out in a previous interview with Xiaojun Zhang | Business Interview that Silicon Valley VC's failure on LLMs, plus SpaceX and Tesla, made him reflect that for "Manhattan Project"-like scientific exploration and discovery with characteristics of "heavy investment, unclear business model, high failure risk," traditional VC may be an unsuitable financial product.

Some disagree. One early MiniMax investor told us that current social sentiment's various doubts are actually confidence issues caused by uncertainty.

AI remains a "half-glass of water" state: some see only that half a glass isn't enough; others see that there is already half a glass. In his view, GPT-3.5 meant LLMs had passed the technological singularity, transforming from a scientific problem to an engineering problem — so investing in LLMs is no longer scientific charity, but a commercial investment before commercial singularity arrives.

"Just as continuously declining lithium battery and charging infrastructure costs transformed new energy vehicles from being dismissed in 2019 with 1.2 million sales to 9.5 million in 2023, LLM performance improvements and continuously declining inference costs will also bring new possibilities," he said.

As for expected LLM investment returns, one early investor gave "2.5x," another gave "either 0 or 10."

When will be the exit point for investing in LLM companies?

One LLM investor's answer remains the same as last March: at $10 billion valuation.

But the difference is, when we discussed SenseTime with him in the first half of last year, he felt this was the floor for this wave of LLM companies; now, he feels that "independent IPO at certain scale" would already be no small feat.

In one investor's memory who participated in both AI waves, the "AI Four Dragons" era was still "a time of AI investor patience." Then, AI companies caught windfalls in security, smart cities, and other sectors, frequently encountering municipal or provincial orders worth tens of millions or even over 100 million, so most companies always had continuous investment inflows. Today's LLM companies "can no longer burn money so gloriously."

Barely a year apart, this direction that was still being romantically imagined in 2023 has been pulled back to more pressing reality.

Image source | Visual China