"Domestic AI Model Race Begins: Baidu Takes the Lead in Sprint"

"Volume" is a single word that is not spoken once; they only express it through action.

This year's "618" wind has unexpectedly blown into the domestic large model circle—manufacturers have begun to fiercely compete on "price," pushing the competition into a new stage: switching from the "qualification round" to the "knockout round."

The reason why large model manufacturers choose to "retreat" on price is to seek "advancement" in business models and user growth. Without reducing the price, it is impossible to accelerate implementation. However, what will be the final outcome of this game of "exchanging price for volume"? How can large model manufacturers balance between "fast, good, and economical"?

As the top AI player in China, Baidu quickly provided an exciting result and answer.

01 Baidu "sprints," leading the "inclusive" era comprehensively

The latest data for 2024 shows that the cumulative user scale of Wenxin Yiyan has reached 300 million. According to the financial report conference call, with the increase in AI demand from enterprises, the daily usage of Wenxin's large model exceeds 600 million times, and the daily Token usage exceeds 1 trillion, both of which are the highest in the country. Compared to the 50 million daily usage announced in the fourth quarter of last year, the growth has exceeded 10 times in half a year. It can be seen that in terms of the momentum of domestic large model application implementation, there should be no one who can surpass Baidu.

Advertisement

While sprinting on the road of "practicality," Baidu is also continuously gaining positive feedback from the commercialization brought by the new AI engine, which is reflected in Baidu's just-released second-quarter financial report.

According to the financial report, in this quarter, Baidu's core revenue was 26.7 billion yuan; the core operating profit was 5.6 billion yuan, a year-on-year increase of 23%. Among them, the intelligent cloud continued to accelerate growth in the second quarter and will maintain the growth momentum.

The surge in the usage of Baidu's Wenxin large model is undoubtedly the result of market choice and also a full display of its model capabilities.

Generally speaking, the performance of large models and usage will form a positive "flywheel." Only larger usage can polish better models, and only better models will have larger usage. In addition, the larger the usage of large models, theoretically speaking, the more room there will be to optimize the inference cost. Therefore, Baidu, which leads in scale, is expected to further optimize costs and efficiency in the future, thereby gaining a stronger competitive advantage.In light of the above, the surge in usage is undoubtedly an effective realization and strong evidence of Baidu's "free and price reduction" strategy, including the full free opening of three models (ERNIE Speed, ERNIE Lite, ERNIE Tiny), as well as a significant reduction in the prices of two flagship models (ERNIE 4.0, ERNIE 3.5).

While price reductions will inevitably intensify short-term market competition, they also reflect the model manufacturers' optimistic expectations for future cost reductions and profit increases. Behind this is the urgent desire to capture market share and obtain the "entry ticket" to the application market. Increasing user activity, generating a data "flywheel effect," and strengthening user stickiness remain key issues that large model manufacturers need to focus on.

It can be anticipated that manufacturers like Baidu taking the lead in price reductions will accelerate the implementation of large models and move towards a more inclusive phase. As the implementation of large models gradually enters the "deep water area," the focus of enterprises is shifting from "being able to use" to "using well." To truly be used well in actual business operations, it will be a comprehensive test of the model's quality, performance, speed, and cost.

Over the past year, Baidu has also been continuously expanding its model portfolio and iterating the capabilities of large models. For example, following the launch of the Wenxin Large Model 4.0 tool version in April, Baidu released Wenxin 4.0 Turbo in May, once again achieving improvements in speed and effectiveness. To date, the Wenxin Large Model has iterated six versions, and Baidu has done so in just four years.

02 Intelligent Cloud Business: AI Continues to Assist, Large Model Commercialization Presses the "Accelerate Button"

The implementation boom of large models has also added fuel to the fire of cloud computing, and Baidu has also reaped the benefits.

In the subsequent telephone meeting, Baidu's management revealed that the intelligent cloud revenue reached 5.1 billion yuan in the second quarter, a year-on-year increase of 14%, and continued to achieve profitability under non-U.S. GAAP; at the same time, the proportion of revenue contributed by AI further increased to 9%, higher than the previous quarter's 6.9%.

In addition, according to the latest report released by IDC, "China Large Model Platform Market Share, 2023: The First Year of Large Models - The Initial Game," Baidu Intelligent Cloud ranks first in the large model platform in 2023 with a share of 19.9%.

Baidu Intelligent Cloud is an important business that supports the commercial implementation of Baidu's large models. The continuous increase in AI contributions actually reflects the "acceleration" of Baidu's large model commercialization process.

According to statistics, in the first half of this year, the implementation of the domestic large model industry continued to accelerate. The number of large model-related winning projects exceeded 200, and Baidu Intelligent Cloud successively won large model projects for top customers such as the All-China Federation of Trade Unions, the National Library, China Southern Power Grid, China Unicom, and China Tower Corporation. Among mainstream large model manufacturers, it achieved the most winning project numbers, the most industry coverage, and the highest total winning amount, accumulating rich practical experience in the implementation of large models in various industries.It appears that Baidu is not only leading in "speed," but also managing to take into account "breadth" and "depth."

Following the explosion of large models, the demand for GPUs has also surged. However, the scarcity of resources and high prices are significant drawbacks. Therefore, the MaaS (Model as a Service) model, which integrates storage, computing, and applications into a one-stop service, has begun to be accepted by the market and has become a new basic service that large model manufacturers and cloud providers are competing to deploy.

In this regard, Baidu's Intelligent Cloud has also launched the Qianfan Large Model, positioned as a "one-stop" enterprise-level large model service platform. Essentially, it is a platform that covers the entire lifecycle of large models, not only including its own ERNIE but also integrating many mainstream large models.

Over the past year or so, the explosion of domestic AIGC applications has been truly indispensable. Public data shows that Baidu's Intelligent Cloud Qianfan Large Model platform has served a cumulative total of 150,000 customers, helping users fine-tune a cumulative total of 21,000 large models.

Nowadays, basic models have gradually become the "territory" of tech giants and a few unicorns. AI-native applications, seen as the future of large models, provide a new opportunity for entrepreneurs and developers.

As large models accelerate towards the application phase, Baidu has also sped up the process of building an AI-native application ecosystem. In April of this year, Baidu released three powerful tools at once to help the industry accelerate the layout of AI-native applications and quickly welcomed a surge in volume. Among them, the construction of applications by Qianfan AppBuilder has reached hundreds of thousands. In addition, in July, Qianfan AppBuilder was upgraded with RAG capabilities to help enterprises efficiently build knowledge question and answer systems.

03 Search continues to distribute "AI" dividends, the intelligent entity ecosystem gradually takes shape, and Baidu Wenku "transforms"

In addition to cloud services, Baidu has also received a lot of "supplies" from AI in its "core territory" of search.

According to financial reports, in June, the monthly active users of Baidu App reached 703 million, a year-on-year increase of 4%; and according to the information disclosed in the financial report conference call, 18% of search results are generated by AI, which continues to increase quarter-over-quarter.

At the same time, one year after Baidu Search was reconstructed based on AI large models, it has begun to show its potential.On August 8th, according to data from one of the most recognized AI application lists in China, the "AI Product List," Baidu Search's "AI Intelligent Answer" topped the domestic overall list for the first time with an absolute advantage. It is worth mentioning that in this AI Product List's domestic overall list, Baidu occupied 3 out of the top 5 spots. In addition to the first place, Baidu Wenku and Wenxin Yiyan secured the 3rd and 5th places, respectively. Moreover, the brand-new product "Orange Chapter" launched by Baidu Wenku also landed on the "Global New Product Growth List," becoming the fastest-growing AI product in China.

The improvement of Baidu Search still owes much to the assistance of intelligent entities. According to the data from the conference call, the average daily distribution of intelligent entities in July exceeded 8 million, which is twice that of May. This means that Baidu's intelligent entities, which are experiencing explosive growth, have gradually formed an ecosystem.

Baidu hopes to use AI to consolidate the basic market and leadership position of search. At present, it seems that intelligent entities, which are regarded as the "last mile" to the landing of large models, are very likely to become a "trump card" in Baidu's hands to turn the situation around.

For Baidu, "intelligent entities + search" can be called a "match made in heaven":

On the one hand, intelligent entities will redefine search and greatly enhance the user's search experience; on the other hand, Baidu Search has a daily request volume of billions of times, with a large user base, super traffic, and intelligent recommendation as three major advantages, providing the best natural scene for the landing and distribution of intelligent entities. Search will also become the largest entrance for the distribution of intelligent entities. In addition, Baidu also relies on the search ecosystem to provide distribution and monetization support for intelligent entities created for enterprises and developers, helping customers increase sales conversion and improve delivery efficiency.

Reconstructing AI search through intelligent entities is a Baidu specialty. Although global giants such as Google, Microsoft, and Meta are heavily involved in AI search, exploring the combination with the application of large model capabilities, most AI search engines are still mainly focused on enhancing the capabilities of large models. What users feel is just an increase in the search of "AI-generated content." This also fits with the new search connotation proposed by Robin Li at "Baidu World 2023" last year, that is, the new search has three characteristics of "extreme satisfaction," "recommendation inspiration," and "multi-round interaction."

In fact, Baidu's heavy bet on intelligent entities has long been traceable, especially since June this year, Robin Li has expressed his high expectations and forward-looking views on intelligent entities in public many times: "In the development direction of AI applications, the most optimistic about intelligent entities." "Intelligent entities are like websites in the AI era, forming a huge ecosystem with millions of quantities."

Baidu, which has achieved KPI, has also won the favor of major banks.

For example, J.P. Morgan said that Baidu has increasingly placed Gen-AI content in its main advertising inventory, and Baidu's search result pages will change over time: from simple queries to complex queries, from single-round queries to multi-round conversations. From a business perspective, such changes will expand Baidu's role in the value chain from a purely traffic-oriented one to pre-sales consulting. It is expected to increase the flow of transaction conversations and will have a positive impact on profits in the medium and long term.Baidu's vision of "using AI to do everything that can be done" seems to be rapidly becoming a reality, and with it, the industry is unlocking new opportunities for growth. Beyond search, Baidu Wenku appears to be another product that has been successfully "lifted off" by large models.

According to information revealed in the earnings call, this quarter, Baidu Wenku's subscription revenue has grown by more than 15% year-over-year. The cumulative number of AI users has surpassed 180 million, and the number of AI feature usages has broken through 2.2 billion. Additionally, the revamped Baidu Wenku leads in over a hundred multimodal AI features such as intelligent PPT, intelligent documents, and intelligent drawing books. Among them, the intelligent PPT has captured 80% of the market share, holding a dominant position. It is not an exaggeration to say that out of every 5 intelligent PPTs, 4 are created by Baidu Wenku.

04 Conclusion

As the competition of large models moves into the second half, the consensus is that "application" is king, which more tests the "internal strength" of each company. As the leader, Baidu is clearly "running" faster, not only leading in speed but also in the depth and breadth of industry implementation, it is at the forefront in China. It not only has a better position in the ecosystem but also has laid a solid foundation for the subsequent performance.

A glimpse at Baidu's second-quarter report shows a clear increase in the AI application market. As manufacturers are increasingly "getting involved," the adoption rate of AI applications in China will also accelerate. Is the "freedom" of large models far away? Perhaps the real sweet spot will also arrive.