The "truth" of AI large-scale model melee: investors see more and invest less, 20 companies only received 6 billion yuan in financing

Source: Sohu Technology

Author: Liang Changjun

Editor: Yang Jin

Figure: At the 2023 World Artificial Intelligence Conference, many companies displayed AI large-scale model products and applications

"Look more, vote less", "the atmosphere of waiting and watching is relatively strong". Talking about the attitude of investment institutions towards AI large-scale model entrepreneurship this year, this is the feeling of some domestic investors and entrepreneurs.

But in the industry, AI large models are still popular. At the just-concluded World Artificial Intelligence Conference, more than 30 large-scale models collectively showed their muscles. Domestic Internet companies and many AI companies have stepped into the hot track of large-scale models, and the battle of 100 models has already begun.

In the venture capital market, big names such as Wang Huiwen, Wang Xiaochuan, and Li Kaifu entered the market one after another, calling out to be China's OpenAI, to be the best large-scale model in China, etc., becoming an important force in this wave of AI large-scale models .

According to incomplete statistics from Sohu Technology, in the first half of this year, at least 20 large-scale model companies received more than 6 billion yuan in financing. From a global perspective, the number of related financing exceeds 50, and China and the United States take the lead, with more than 20 each, involving a total of 100 billion yuan.

Although the number of domestic transactions is relatively large, the amount only accounts for 6% of the world. Domestic investors are not generous, and the AI large model investment market is a little deserted.

At the same time, all parties have not reached a consensus on the entrepreneurial value of large models. Zhu Xiaohu, the managing partner of GSR Ventures, and Fu Sheng, the chairman of Cheetah Mobile, have been arguing with each other for this earlier.

Cheng Hao, the founder of Yuanwang Capital and Xunlei, judged that there will not be more than 10 general-purpose models in the world, and there is little opportunity for entrepreneurship. This has basically become a consensus in the investment circle. Baidu CEO Robin Li also said before that there is no need to reinvent the wheel.

But many entrepreneurs find it hard to agree. Wang Xiaochuan believes that start-up companies must have opportunities, and without the burden and greater commercial pressure, they will run faster than large companies.

Li Wei, Vice President of Engineering & Chief Scientist of Going out to ask, told Sohu Technology that it is unlikely that major manufacturers will monopolize general-purpose large models. "Many startups have launched large models, proving that this matter is not a big problem anymore."

Despite the differences, there is consensus that application will be key. From waiting to see whether to do it, to the entry of players from all walks of life, the large model has reached the stage of answering how to implement it. Whether it is a big factory or a start-up company, this is a must-answer sheet.

The boss has his own halo when he enters the game, who is investing in the big model?

ChatGPT was born, setting off the peak of the large model of the third wave of AI, and a group of bigwigs started their own businesses. According to Sohu Technology’s incomplete statistics, in the first half of this year, at least 20 large-scale companies obtained financing, and most of them were in the early angel round or A round.

These entrepreneurs basically have their own halo, including Kai-fu Lee, Wang Huiwen, Wang Xiaochuan, Li Zhifei, Zhou Bowen, etc., who have entrepreneurial or large-scale backgrounds, as well as academic rookies such as Tsinghua University, National People’s Congress, and West Lake University, among which Tsinghua University is particularly typical. Tsinghua professors stand behind Shengshu Technology, Shenyan Technology, Dark Side of the Moon, Qingmao Intelligence, Face Wall Intelligence, and Lingxin Intelligence.

Figure: AI large-scale model companies that publicly obtained financing in the first half of this year Source: public information, Tianyancha

From the perspective of investors, there are not only industrial capitals such as Tencent, Baidu Ventures, Ant Group, and TAL, but also venture capitals such as Sequoia, IDG, ZhenFund, Sinovation Ventures, Qiming Ventures, Matrix Partners, and Qiji Ventures. cast. Sequoia is the most active, with at least 5 shots, including Light Years Beyond, Shenyan Technology, The Dark Side of the Moon, and Project AI 2.0, while Tencent has invested in Light Years Beyond, MiniMax, and Shenyan Technology.

In terms of financing scale, the total financing disclosed by the projects in this statistics exceeds 6 billion yuan. According to public data, in the first half of this year, there were 51 corporate financings involving AI large models in the world, with an investment and financing amount exceeding 100 billion yuan, that is, the number of domestic transactions accounted for nearly 40%, but the financing amount was only about 6%.

Multiple large-scale transactions in the U.S. market accounted for the majority, including Microsoft’s $10 billion investment in OpenAI, Bill Gates, Microsoft, Nvidia, etc. led the investment of $1.3 billion in Inflection AI, and the U.S. data company Databricks acquired MosaicML for $1.3 billion . These three transactions alone accounted for about 90% of the global total.

This also shows to a certain extent that although domestic large-scale models are fiercely involved and concept stocks in the capital market have doubled and skyrocketed, investors in the primary market are still relatively cautious, especially in large-value transactions.

Qu Kai, the founder of Chapter 42, which is engaged in venture capital FA business, revealed that almost all US dollar funds are currently looking at AI, and some RMB funds are also interested. "Many institutions are very positive, but in the end there are not many who will make a move. There will probably be fifty or sixty."

Li Wei, Vice President of Engineering & Chief Scientist, can also feel that the investment community is actually very concerned about and fancy big models, but they are also very cautious. "After all, this is a new technology direction with a large investment, but the business model is currently unclear."

Guo Tao, an angel investor and a senior expert in artificial intelligence, analyzed Sohu Technology and said that more and more investment institutions have realized that AI large models face large investment amounts, long payback periods, low success rates, fierce competition in the industry, and increasingly stringent regulations. And other issues, so the shot tends to be cautious, and the current wait-and-see atmosphere is relatively strong.

At the same time, Guo Tao believes that there are not many large-scale AI companies worth investing in at present. Most start-up companies have almost no significant advantages in technology, data, and ecology. of favor.

In addition, the valuations of many promising star companies are too expensive, typically light-years away. In just three months, the valuation has jumped from 200 million US dollars to 1 billion US dollars, and many institutions are discouraged. The ending of this company is also quite embarrassing. Wang Xing’s Meituan “saved” his brothers and investors with more than 2 billion yuan, and Wang Huiwen left early due to illness.

Continuing the Internet investment preference, and starting to fight after the fever subsides

In terms of the specific investment direction of AI large models, investment institutions have almost continued the preferences of the Internet investment era, and the application layer is the most popular. Cheng Hao said that Yuanwang Capital mainly invests in middleware and application layer companies.

According to the data disclosed by Qu Kai, among the AI projects that have received money this year, 10%-20% are for models, 20%-30% are for infra/intermediate layers, and 60%-70% are for application layers. . If you don't count the projects that get money, the projects that make applications may reach 95%.

From the underlying infra (such as chips, frameworks and other infrastructure), to the model layer, middleware, and various applications, the AI large-scale model industry has also formed an inverted pyramid structure similar to the chip industry. In the case that the bottom layer relies on foreign open source technology and is difficult to break through, most domestic startups bet on the application layer, and the high-cost model layer is only a game for a few companies.

Li Wei believes that start-up companies mainly do basic large-scale models and downstream applications of large-scale models, and it is neither realistic nor necessary to make basic large-scale models. "Therefore, many start-ups tend to apply in vertical scenarios. By means of API calls or OEM privatization deployment, they will leverage on the large model services of large model suppliers to focus on innovative research and development of data and applications. This will be a relatively clear direction."

However, this wave of big model craze seems to have begun to cool off. The often cited example is the peak of ChatGPT traffic. According to the data of the third-party website SimilarWeb, the global traffic of ChatGPT’s website and mobile client decreased by 9.7% month-on-month in June this year, the first decline since the end of last year, and the time spent by visitors on the website also decreased by 8.5%. In addition, the traffic of websites such as Microsoft Bing and Character.AI also declined to varying degrees in June.

"The market has been getting colder in the past one or two months, because the qualitative changes in large models have been decreasing recently, and there are fewer new entrepreneurs and new stories to tell than at the beginning of the year. But every track and hotspot has its own advantages and disadvantages. Fu, this is normal." Qu Kai said.

He believes that in the next stage of the big AI model, we must work hard to implement it. The next wave of hot spots will probably be two or three months later. For a large number of projects that have received money in the first half of the year, it will take a few months to complete the product launch. See if there are more and better killer apps, and see who will be the leader of the application layer.

In fact, major manufacturers are currently striving for the landing and commercial application of large models. When Huawei released Pangu 3.0 recently, it expressed that it hopes to use it to help all walks of life, rather than focusing on the voice model level. "We're so busy doing things that we don't have time for poetry."

Guo Tao believes that although the AI model has cooled down, it will last for about a year as a whole, and the focus will gradually shift from the underlying technology level to the vertical application level.

Li Di, CEO of Xiaoice, judged that the homogenization of large models is serious now, and there is no need for so many large models on the market. The fever should subside in 2024, and it is found that who is on the shore and who is swimming naked.

No chance for general model? The homogenization of the vertical track is serious

For entrepreneurs, how to think about the direction before entering the market is extremely critical. Zhu Xiaohu said that ChatGPT is very unfriendly to start-up companies, and will give up financing fantasies in the next two to three years. Fu Sheng complained about this, saying "Our investors are ignorant and those who are fearless."

Later, Zhu Xiaohu explained that he did not deny entrepreneurial opportunities in the field of large-scale models, but reminded entrepreneurs not to be superstitious about general-purpose large-scale models. "For most entrepreneurs, scenarios are the priority and data is king."

This view has basically become the consensus of the current domestic investment circle. Guo Tao believes that general-purpose large-scale models will form a certain monopoly situation, and entrepreneurs and "small factories" will be at a disadvantage in terms of capital, technology, data and ecology when they deploy general-purpose large-scale models.

"The moat of the general-purpose large-scale model is very high, and its network effect is also strong. User feedback will make it smarter and smarter, and the company that makes it first will have a first-mover advantage." Cheng Hao also believes that there is no chance for start-up companies to make a general-purpose large-scale model. Big, only big factories can afford it.

At the same time, he believes that the general base model will not require so many companies in the future. "Maybe there will be no more than 10 closed-source and open-source general-purpose models that can be used in the world in the future, which is too much."

But many entrepreneurs disagree. "Many start-up companies have launched large-scale models, proving that this matter is not a big problem. If it is to reach the level of GPT-4, it is also very difficult for top manufacturers, and it seems impossible to monopolize it now." Li Wei said .

He believes that although start-up companies cannot compete with large-scale manufacturers in terms of hardware resources and engineering strength, their advantages are that they are more down-to-earth, have an overall landing path or product demand, and do not make large models for the sake of large models.

"The basic large model is not monolithic, and there is room for expansion. Startups can start with a model of one billion or ten billion, and then decide whether to continue to develop a model of 100 billion according to their own needs and follow-up resources." Li Wei said .

Many start-up companies follow this route when making large-scale models. For example, Baichuan Intelligence first launched a 7 billion parameter model, and is still training a 50 billion parameter model. Basically, no startup company chooses to make a large model with 100 billion parameters.

In Li Wei's view, the size of the model is not the only indicator, and the application scenario is also an important reference dimension. "For most applications, a super-large-scale model is like a cannon hitting a mosquito. Not only does it take a long time to infer, it costs a lot, and it is difficult to apply and deploy, and it does not make much sense in terms of actual effect."

Therefore, Li Wei believes that blindly competing for perfection should not be the mainstream of large-scale model R&D and innovation. It is a more meaningful and healthy competition to promote large-scale models to be lightweight and verticalized. important.

Cheng Hao believes that in the future, many small and medium-sized enterprises will have their own vertical models, and startups will have more opportunities to develop applications or tool chains on the vertical track. "Most investment institutions are still looking at the track of vertical industries, like Wang Xiaochuan and Wang Huiwen, to a large extent they are betting on people, not on making general-purpose large-scale models."

However, choosing the technical route of the vertical track also faces many challenges. Li Wei said frankly that the large-scale implementation of large-scale models in the vertical field will be more difficult than imagined.

He believes that one of the current challenges is that the changes are too fast and dazzling. It is difficult for start-up companies to connect and select large-scale suppliers. Most large-scale models have not yet produced mature services that can be applied, and the upstream and downstream cannot be seamless. The problem of docking and matching.

At the same time, there are still problems such as technical homogeneity and commercial inter-rolling. "Encouraging technology competition and differentiated development of models, coordinating business cooperation, and reducing business inter-involvement are the directions that technology companies and communities should work together." Li Wei appealed.