The wind is still blowing on the large model, but the direction of the wind is gradually changing.
From the end of 2022 to 2023, the field of large models ushered in a wave of blindfolded chasing. Today, large models are still the hottest word in the global technology field, but the focus of discussion about it has changed from the initial "technological subversion" to the current "commercial landing".
In the current financing environment, how to solve the problem of commercialization has undoubtedly become a key "topic" for the next stage of the breakthrough of the large model. Looking at today's market, there are two main types of companies that are expected to run through the commercialization of large models: ecological players represented by Wondershare Technology and application-oriented players represented by Wondershare Technology.
Interestingly, these two companies are platform-based supergiants and vertical industry leaders respectively, which to a certain extent also reveals a reasonable path for different types of companies to compete for large models.
Commercialization has become a must-answer question for large model manufacturers.
Fewer and fewer people are willing to pay for a story about technology, even if it's about a big model that's hot.
The objective reality is that financial conditions continue to tighten. According to Crunchbase, 2023 is the lowest year for venture capital since 2018, with global startup financing of $285 billion, down 38% year-on-year. According to data released by Pitchbook, a U.S. data analysis company, U.S. startups will raise $170.6 billion in 2023, down about 30% and 50% from 2022 and 2021, respectively.
In the Chinese market, large models have set off a wave of financing boom in the first half of 2023, but they also tend to calm down in the second half of the year. According to data from the China Business Industry Research Institute, from May to July 2023, the number of AI-related investment events in China ranked highest, remaining above 60 per month, but has dropped to 26 in November.
Under the tuyere of large models, many companies have rushed to the top and set off a "hundred model war" for a while, but some investors who pay attention to this field said, "More than 90% of large model companies basically do not have investment value." "Some of the most sought-after star companies have accumulated financing of more than billions of yuan, but the team admits that such financing is still far from enough.
The way to break the game is, of course, "self-hematopoiesis". Small factories require survival, large factories need to see the report, no one can continue to invest in research and development regardless of returns, everyone is working hard to find a path to commercialization, "from virtual to real" is becoming the consensus of the entire industry.
So what exactly is the path? Open source or closed source? TOB or TOC?Do generic or vertical? Various large-scale model manufacturers are exploring different paths. However, it can be clearly felt that the attention of the industry is constantly tilting towards the "application layer".
Robin Li, the founder, repeatedly emphasized: "Focusing on large models itself does not make much sense, and the real opportunity lies at the application level." Wu Taibing, chairman of Wondershare Technology, said early in the morning: "The large model is the queen, but the application is the king, and the application software will become the only way for the commercial landing of the large model." β
360, which has repeatedly appeared in the spotlight recently, also has a similar judgment, its head Zhou Hongyi judged that "the application of large models will usher in huge opportunities in 2024", and its vice president Huang Jian also said that "applications are more important than large models".
Under the consensus of "application is king", giant players and vertical leaders are gradually finding their own commercial answers.
Platform-based company: do B-end, do general, and deeply cultivate the ecology.
Today, the reason why the craze for large models has calmed down is that more and more people realize that the research and development of large models is an all-round competition, involving various dimensions such as computing power, algorithms, networks, and big data. In this sense, if you really want to do the underlying development and ecological construction, there are only a handful of people who really have the admission ticket.
And even if it is a giant, competing at this level, no one dares to guarantee that there is enough ammunition, and whether it can run through commercialization and form a positive cycle is also an unavoidable hurdle.
Among the domestic super players, the layout of the large model is obviously the most radical, almost showing an all-in posture. It has to be said that it is one of the earliest and most clearly thought among the giants, and its exploration in commercialization is also at the forefront, providing a feasible sample for the commercialization of large models of platform companies.
In March 2023, the "Wenxin Qianfan" large model platform was officially launched, providing customers with enterprise-level large language model services, including a full set of Wenxin large models and corresponding development tool chains, and then connected to dozens of third-party models, aiming to aggregate more enterprises to develop the Wenxin ecosystem and make revenue through these underlying service capabilities.
Calling API is currently the charging method of many large model TOBs at home and abroad, and Alibaba and Microsoft have also made similar attempts. However, judging from the data disclosed by various companies, the ROI of API revenue is not high, and some of them are even difficult to cover costs. However, the API pattern can facilitate the binding of the platform to the customer, so as to provide the customer with other services, such as the cloud.
According to the latest data, the number of developers in the Wenxin ecosystem has reached 10.7 million. In August 2023, it announced its Q2 financial report, showing that its revenue in the second quarter increased by 15% year-on-year, and its attributable net profit increased by 44% year-on-year. The intelligent cloud, which has attracted much attention, has also continued to be profitable after achieving profitability for the first time in Q1 23. The third-quarter financial report also showed that its revenue and profit exceeded market expectations.
In fact, this is quite a rare achievement. After all, the layout of the large model of most domestic companies has not yet formed a blessing on performance, and some have even caused a decline in performance. The core reason is to think more clearly, not simply sell services, but to do ecology.
Many large-scale companies have focused on TOB services from the beginning, but it obviously takes a process to get more companies to pay. Some people in the industry once said bluntly: "The model can fire half of the people, and the company will consider using it." β
According to the large-scale model bidding demand analysis briefing, from January to November 2023, a total of 124 procurement requests were launched in the bidding market, with a procurement amount of 29.7 billion yuan. Compared with the huge research and development of various large model companies, this is obviously a drop in the bucket.
According to the Large-scale Model Bidding Demand Analysis Briefing, the procurement amount from January to November 2023 is 29.7 billion yuan.
At the stage where the large model is not yet mature to monetize directly through TOB, it is imperative to build an ecosystem and form a positive cycle.
To attract more developers and build an ecosystem, it is a test of the manufacturer's pedestal model capabilities, engineering capabilities, capital reserves, and even brand influence. But in any case, the feasibility of this path has been verified to a certain extent.
Application-oriented companies: do C-end, do vertical, focus on scenarios.
As mentioned above, "application is king" has become the consensus of the large model industry. On top of the infrastructure built by super players, there is also a surprising company among application players, Wondershare Technology (300624., which has just released a large modelSZ) is one of the typical ones.
In January 2024, the "Tianmu" sound multi-scale model created by Wondershare Technology was officially released, which attracted the attention of the industry. On the one hand, because this is a self-developed model launched by an A-share listed company and a creative software A-share listed company, from the perspective of development, everyone pays attention to its combination with creative software and applications, on the other hand, "Tianmu" has won a number of "firsts" and has been commercialized. It is the first large model in Hunan that has passed the "Regulations on the Administration of Deep Synthesis of Internet Information Services" algorithm, and it is also the first multi-model with audio as the core in China, with core functions such as one-click filming, AI art design, Wensheng, audio enhancement, sound analysis, and multilingual dialogue, and has been commercialized on a large scale overseas.
With the help of "Tianmu", you can understand the creator's intention with just one ** and a few prompts, quickly build scenes and characters, and help the creator complete the content creation of film and television dramas, cartoons, advertisements and other **.
China's first large-scale model Wondershare "Sky Canopy" was released to help creators efficiently complete the creation of various audio content.
In fact, before the official release of "Tianmu", Wondershare Technology's AIGC application layout has received widespread attention in the industry. Focusing on the creative software track, Wondershare Technology has embarked on a commercialization path of "vertical large model + AIGC application", which also provides a reasonable paradigm for vertical head companies to break through large models.
First of all, why the "vertical large model"?
Compared with the general large model, which often costs billions of dollars in R&D, the threshold for vertical large models is much lower, which can be as small as one billion parameters, and the development cost of using open source frameworks will be lower.
Moreover, there is a gap between the general large model and the specific industry application, which cannot carry the specific needs of the professional vertical field. The vertical large model is trained with unique data, which is a blank area that is difficult for the general large model to step into no matter how smart it is, which is the significance of developing the vertical large model.
In June 2023, Zhou Hongyi once said that the vertical model is the future development direction, and the real value is to make the large model change from "know-it-all" to "jack-of-all", industry-based and enterprise-oriented. In January this year, Zhou Hongyi said more bluntly at the Yabuli Forum that in 2023, the large model will look like an atomic bomb, and this year it will look like a tea egg, in 2023, everyone will be racing OpenAI to engage in a general large model, and in 2024, everyone is thinking about how to find their own application scenarios for large models, and the goal has changed. The data also shows that in 2023, the release of global general large models will decline, with less than 50; So far in 2023, the release of vertical models has soared, and there have been more than 100, including the Ziyue education model released by NetEase Youdao and the Ctrip Wendao tourism industry model released by Ctrip. However, it is worth noting that in the era of ** is king, large models with sound ** as the core are still relatively scarce, so the value of "Tianmu" is obvious. Also, why "AIGC app"?This is in line with Wondershare Technology's original business.
Wondershare Technology is mainly engaged in the field of digital creativity, and its software products such as Wondershare Filmora, Wondershare Filmora, Edraw, and Ink Knife have been integrating AI capabilities into AI capabilities for a long time, and the emergence of large models has further upgraded the AI capabilities and interaction methods of its products. Unlike many companies that build large models first and then find application scenarios, the original intention of Wondershare Technology to build large models is to empower business scenarios.
From a commercial point of view, compared with ToB's high threshold and high customization cost, ToC is easier to amortize R&D costs through user scale. In a global market where consumer applications have become extremely mature, tools may be the solution, and the direction is to make tools "more usable".
As a company with one billion users around the world, Wondershare Technology can be said to be driven by user needs.
For example, Wondershare Filmora, a creative software owned by Wondershare Technology, has insight into many YouTubers overseas, hoping to focus on creation more efficiently. With the blessing of "Skylight", Wondershare Filmora has implemented AI Copilot intelligent editing assistant, which can provide relevant creative suggestions and guidance according to the natural language description entered by users, allowing users to focus on more important and creative work. Users can also click "Apply" to complete a series of preliminary operations such as adjusting ** parameters and adding transition effects with one click, which greatly simplifies the ** creation process.
On the other hand, the massive amount of data accumulated in vertical industries is also the key to the success of vertical large models. At this point, companies with mature business and sufficient accumulation in a specific field have more advantages.
At present, the large-scale model capability of "Tianmu" has been fully applied to the products of Wondershare Technology. The first-class generation ability provided by the "Tianmu" large model is significantly higher than the industry level in terms of motion control accuracy, aesthetic quality, and comprehension, which also means that Wondershare Technology's creative software products will greatly benefit.
In the same way, Wondershare Technology's ability to apply large models has been reflected in the growth of user and other data. Its AI digital human short** marketing tool Wondershare has exploded, and the monthly active users have increased by more than 700% in one year; Its flagship software, Wondershare Filmora, saw a more than 240% increase in the number of users using AI features from August to December last year.
There is an argument that large models are not the key to the current revenue of AIGC companies, but they are a necessity, because there will be no barriers to AIGC without large models. In fact, with the blessing of AIGC, Wondershare Technology's performance in the past two years has achieved benign growth, and its just-released performance forecast shows that its net profit attributable to the parent company in 2023 is expected to be 75 million yuan to 100 million yuan, a year-on-year increase of 818%β142.39%γIn addition, the increase in the pricing of the capital market also proves the barriers and prospects of this company to a large extent.
With a more applicable vertical model and a C-end product that has been verified by the market, Wondershare Technology provides another answer. In the second half of the 100-model war, the platform company did the B-end and deeply cultivated the ecology; Application-oriented companies can do C-side, vertical, and focus on scenarios, and perhaps they can finally stay at the table.