Two rounds of financing exceed last year's total financing in China's large model sector.
Just over a month into 2024, the capital competition in China's large model sector has entered a new phase.
LatePost has learned that the large model startup Moonshot AI has recently completed a Series B financing of over one billion dollars, with investors including Alibaba, Meituan, Xiaohongshu (Little Red Book), and Sequoia China, among others. After this round, Moonshot AI's valuation has exceeded 2.3 billion dollars.
This is currently the largest single round of financing in China's large model sector, but this record may not hold for long.
It is understood that another leading large model company is also conducting a new round of financing of one billion dollars, and the transaction has entered its later stages.
The combined two billion dollars in financing have already exceeded the total public financing amount in China's large model sector last year (about 10 billion yuan).
Large model startups are competing for the remaining oxygen: overall investment from dollar funds is sluggish, and yuan funds place more emphasis on certainty. The main entities willing and capable of making significant investments in large models are large internet or technology companies.
As limited investors, China's internet giants are also engaged in fierce competition and have differentiated into different actions and strategies: one is to invest heavily externally to serve strategic goals; the other is to hold back for now and focus on developing their own large models.
The new shareholders of Moonshot AI in this round, Alibaba and Meituan, belong to the first group.Alibaba has previously invested in large model startups such as Smart Spectrum AI, Baichuan Intelligence, and Zero One Infinity; Meituan has invested in Smart Spectrum AI and has invested in and acquired Beyond Light Year.
Companies that cast a wide net also include Tencent and Xiaomi: Tencent invested in Baichuan Intelligence, Smart Spectrum AI, MiniMax, and Beyond Light Year last year; Shunwei, founded by Lei Jun, invested in Baichuan Intelligence, Smart Spectrum AI, and MiniMax, with Baichuan and Smart Spectrum AI also being invested in by Xiaomi.
Baidu and ByteDance have not yet invested in large model companies.
Alibaba is the most aggressive among Chinese internet giants in investing in large models.
An investor said that in Baichuan and Smart Spectrum AI, which are invested in by both Alibaba and Tencent, Alibaba has a higher stake. This time, Alibaba has also participated in the new financing of the Dark Side of the Moon, which exceeds 1 billion US dollars.
The difference in actions comes from different strategic intentions and degrees of determination.
Alibaba has formed a relatively clear large model strategy in 2023: the ultimate goal is not to make money by self-developed models, but to provide cloud computing capabilities for a group of large models, and to seize the opportunity to become an AI computing power infrastructure.
A survey by Silicon Valley venture capital firm a16z found that 80%-90% of early financing of large model and generative AI startups is used to purchase computing power from cloud computing platforms; a16z estimates that in the future, when the industry is more mature, model and application companies will also have to give 10%-20% of their annual revenue to cloud computing companies. It is equivalent to a "AI tax".
One way to collect the AI tax is to cooperate deeply with potential large model companies, that is, buyers of AI computing power: for example, Microsoft promised to invest 10 billion US dollars in OpenAI at the beginning of last year, most of which is settled in cloud computing resources rather than cash; Google and Amazon also promised to invest a total of 6 billion US dollars in Anthropic last year, Anthropic will spend more than 3 billion US dollars to purchase Google cloud computing services within four years, and will also purchase cloud computing services from Amazon.
Heavy investment in large models helps Alibaba Cloud to continue to consolidate its leading market position in China and seize the incremental opportunities brought by AI computing.Compared to Alibaba, Tencent's investment strength is weaker, possibly for two reasons: first, cloud computing is less important to Tencent than to Alibaba, and CSIG (Tencent Cloud and Smart Industries Group) has been downsizing staff over the past two years; second, Tencent wishes to focus more on refining its own Hunyuan large model.
Baidu and ByteDance, which have not yet invested in large model startups, each have their own logic.
According to Baidu's Q3 2023 financial report, the cash and cash equivalents at the end of the period were less than 35 billion yuan, while ByteDance's cash reserves during the same period exceeded 30 billion US dollars.
With limited resources, Baidu's priority is to develop its own large models and corresponding products. A person close to Baidu stated that Baidu had already clarified last year that it would not invest in large model companies, focusing instead on application companies.
ByteDance, which has a large amount of investable funds, has wavered several times.
It is understood that in the first half of 2023, ByteDance once considered investing in large model companies MiniMax and Step Star, but ultimately did not proceed.
In the middle of 2023, ByteDance decided no longer to invest in large model companies, but to concentrate on developing its own large model, which is key for ByteDance to seize the opportunity of the AI era's super App.
Just recently, ByteDance has once again approached top large model companies, reassessing the necessity of investment.
The background is that ByteDance's own development progress of the large model has not met expectations; Alibaba's aggressive moves and the recent undercurrents in large model investment and financing competition have also brought certain pressures.
OpenAI's recently released video generation model Sora is a new impact.A previous market perspective was that ByteDance, which owns TikTok, and Google, which owns YouTube, have a significant advantage in creating video multimodal models due to their vast amounts of video data. However, Sora's impressive performance has demonstrated that a sufficiently powerful large language foundation model can encompass other modalities such as images and videos, suggesting that model technology capabilities may be more important than data resources.
Large internet technology companies that originally had ready-made content or social products and thus access to massive data resources now have to reassess whether they can maintain a lead. "ByteDance might be a bit panicked now; they didn't have any investment cards in their hand before," said an industry practitioner.
The Dark Side of the Moon, established in March 2023 and invested in by Alibaba, Meituan, and others, was founded by Yang Zhi Lin, an assistant professor at the Institute for Interdisciplinary Information Sciences at Tsinghua University.
Yang Zhi Lin, 32 years old, graduated from the Department of Computer Science and Technology at Tsinghua University and later pursued his Ph.D. at the Language Technologies Institute (LTI) at Carnegie Mellon University. During this time, he published Transformer-XL and XLNet as the first author; the former was the first attention language model to fully surpass RNNs, and the latter outperformed Google's BERT model in 20 tasks. Yang Zhi Lin's academic citations have exceeded 20,000 times since 2019.
Several investors close to Yang Zhi Lin have commented on his rare intelligence and astonishing growth rate. The Dark Side of the Moon is characterized by its focus on Chinese, long text, and consumer product-driven approaches. In October 2023, the first model released by The Dark Side of the Moon, Moonshot, and the intelligent assistant Kimi could support input of 200,000 Chinese characters, which is 2.5 times that of Anthropic's Claude 2-100k (about 80,000 characters) and 8 times that of OpenAI's GPT-4-32K (about 25,000 characters).
An AI industry practitioner stated that, at the end of last year, The Dark Side of the Moon was also one of the Chinese top large model companies with a relatively high investment cost-performance ratio: after its 2 billion yuan (approximately 300 million USD) Series A funding in October last year, its valuation was around 800 million USD, lower than the nearly 2 billion USD valuations of MiniMax and ZhiPu AI at that time, as well as the unicorn status of Baichuan Intelligence.
A single round of 1 billion USD in financing is not something that typically occurs in the early stages of an industry. Even during the last AI boom, which was a period of prosperous financing, SenseTime took 4 years to accumulate 1 billion USD, and Megvii took about 7 years, while The Dark Side of the Moon had just been established for 11 months.
The unusual level of investment is backed by extraordinary opportunities and extremely high expectations from investors. The B-round investment in The Dark Side of the Moon by Alibaba, Meituan, and others may ignite the FOMO (Fear of Missing Out) dust that has been pervasive, like a spark.
Comments