The development of large models seems to have reached a period of observation. When new technologies represented by generative AI emerged, the market attitude was initially overly optimistic, thinking that large models could do anything. Later, it gradually turned pragmatic and neutral, recognizing the limitations of large models, and now there is even a slightly pessimistic sentiment.
At the end of June, many experts expressed that people's expectations for AI are too high and the investment is too large, but its existing benefits and potential benefits are too small. At present, AI has a huge bubble risk.
Does AI have a huge bubble at present? At least Baidu doesn't think so.
On the evening of August 22, Baidu released its Q2 2024 financial report, showing a total quarterly revenue of 33.9 billion yuan, with Baidu's core revenue at 26.7 billion yuan and Baidu's core operating profit at 5.6 billion yuan, a year-on-year increase of 23%, exceeding market expectations.
Robin Li, founder, chairman, and CEO of Baidu, said in the earnings call that the daily number of calls for the Wenxin large model exceeded 600 million times, and the daily processing of Tokens text exceeded 1 trillion, both of which are the highest in the country.
Advertisement
At the same time, AI-driven Baidu's intelligent cloud continues to maintain strong growth, with a year-on-year increase in revenue of 14% in this quarter, among which AI-related revenue accounted for an increase of 9%.
Baidu is also continuously reconstructing internal businesses including search with AI. Currently, 18% of the search results on Baidu search are generated by AI; the subscription revenue of Baidu Wenku has increased by more than 15% year-on-year, and AI functions are popular with users; the distribution volume of intelligent bodies in the Baidu ecosystem is rapidly increasing, with an average daily distribution of more than 8 million in July, twice that of May.
The existence of bubbles is due to the irrationality of the market and the inability of technology to meet expectations. Looking at the three-level rocket of AI business that Baidu has built, the continuous growth of the number of calls for the Wenxin large model, the optimistic revenue expectations for the intelligent cloud that provides infrastructure for large models, and the huge monetization potential of a large number of businesses that have been reconstructed by AI.
Baidu does not believe in the bubble of large models because the key to realizing the value of AI is in its hands.
The large model industry is integrating.
"I think in the next two to three years, the competition in the field of artificial intelligence will be very fierce. As for who will be the final winner, my view is that whoever makes money will survive," said Robin Li.
Views on the future directly affect the actions of artificial intelligence manufacturers. Last year, when China's artificial intelligence industry was still in the so-called "hundred schools of thought" period, Robin Li predicted that the artificial intelligence industry would eventually move towards integration, and only a few basic large model companies would survive.
At present, the daily number of calls and the number of Tokens processed daily by the Wenxin large model is the highest in the country. Compared with the 50 million daily calls announced in Q4 of 2023, the growth has exceeded 10 times in half a year. A little observation can reveal another fact: other domestic large models, with a lower base than Wenxin, will also difficult to catch up in growth speed, and the gap will only widen, forming a de facto Matthew effect.
Compared with the overseas market, OpenAI took the lead step by step and is still crushing other large models in scale. Baidu is the world's first listed company to launch a model similar to ChatGPT, and since then, it has continuously upgraded the model through Baidu's self-developed four-layer artificial intelligence architecture. The 4.0 version of the Wenxin large model has become China's first large model to match ChatGPT GPT-4.
In June this year, Baidu launched the 4.0 Turbo version of the Wenxin large model. Compared with the 4.0 version of the Wenxin large model, the Turbo version is faster and lower in cost. It has already supported model fine-tuning to facilitate enterprises and developers to train larger models that are more in line with needs and improve the use effect in business.
The popularization of new technology is inevitably accompanied by a significant reduction in the threshold of use, not only in technical thresholds but also in cost thresholds. Whether it is a large enterprise or a small and medium-sized enterprise, the strategy of free and reduced prices for large models can widely promote the breadth of large model adoption and drive the number of users and API calls for the Wenxin large model to increase significantly.
At present, Baidu's intelligent cloud's ERNIE-Speed, ERNIE-Lite, and ERNIE-Tiny series of model preset services are free for customers to use, and the flagship models ERNIE 4.0 and ERNIE 3.5 have been significantly reduced in price, reducing the trial and error costs for enterprises with large models. ERNIE 4.0 Turbo is fully open to enterprise customers, with input and output prices as low as 0.03 yuan/thousand Tokens and 0.06 yuan/thousand Tokens, respectively.
Robin Li mentioned that Baidu has been committed to the popularization of the Wenxin large model, hoping that more and more users can afford it, and allowing more users to use the Wenxin large model to solve real-world needs. Therefore, it continues to reduce the cost of model inference, expand the model portfolio, and develop tool kits for model construction and application construction, so that users, partners, and developers can use the various powerful functions of the Wenxin large model more efficiently and effectively.
Better models and lower prices can be exchanged for high-frequency use of large models, and outputting better results at lower prices not only tests the funds of large model manufacturers but also reflects the comprehensive strength of the manufacturers. With the in-depth popularization of large models, the industry is accelerating integration.
Intelligent cloud has reached the harvest time
The growth of intelligent cloud is both expected and unexpected.
What is expected is that Baidu's intelligent cloud will benefit from large models. Baidu's intelligent cloud revenue reached 5.1 billion yuan in this quarter, a year-on-year increase of 14%, and continued to achieve profitability (Non-GAAP) and profit margin improvement. At the same time, the proportion of AI-generated revenue has further increased to 9%, higher than the previous quarter's 6.9%.
What is unexpected is that the outside world still underestimates Baidu's intelligent cloud. Shen Dou, President of Baidu's Intelligent Cloud Business Group, said that he is very confident that Baidu's intelligent cloud business revenue will maintain a strong growth momentum in the next few quarters and achieve long-term growth, and profits will also maintain sustainable and healthy growth.
"We believe that the performance of Baidu's intelligent cloud business is healthier than ever, and it can continue to bring us operating profits (calculated according to non-generally accepted accounting principles). We will also continue to focus on profit improvement," he said.
The landing of large models is inseparable from infrastructure. The strong demand from all walks of life for generative artificial intelligence and large language models (LLMs) has largely been converted into demand for intelligent clouds. Combined with the strength of the Wenxin large model, Baidu's intelligent cloud can naturally better meet customer needs.
Robin Li said that Baidu provides the industry's most advanced and cost-effective AI infrastructure and an excellent MaaS platform, becoming the common choice of more and more enterprises. According to the latest report from IDC, Baidu's intelligent cloud ranked first in the market share of China's large model platform in 2023, reaching 19.9%.
It was learned from the earnings call that in this quarter, the expenditure of GPU users on Baidu's public cloud has increased significantly. Currently, Baidu's intelligent cloud is cross-selling CPU cloud services to GPU users, and the CPU cloud revenue from these users has increased significantly. Over time, Baidu's intelligent cloud market share will continue to grow, and the standard profit of generative artificial intelligence cloud business will exceed traditional cloud business.
Regarding price wars, Baidu also said that the market for generative artificial intelligence and large language models is still in a very early stage of development. Convenient operability and competitive prices are crucial for expanding the market. Therefore, while continuously improving the capabilities of the Wenxin large model, the intelligent cloud will also continue to reduce inference costs to ensure Baidu's leading position in the industry.
Some users may wish to conduct model training, and the intelligent cloud provides powerful and cost-effective artificial intelligence infrastructure; some users hope to use large language models and call APIs, and the intelligent cloud provides Baidu's Wenxin large model series, combined with a series of comprehensive tool kits, allowing users to easily achieve model fine-tuning, model customization, and develop appropriate applications for specific needs.
In terms of customer landing, in the field of public services, Baidu cooperated with an enterprise to help more than 6,000 villages improve grassroots services. Since the large-scale landing in April, the daily usage of the service has increased by more than 30 times, reaching more than 2 million times; in the medical industry, enterprises have trained industry-specific models through the Wenxin large model and model customization tools ModelBuilder, which can assist doctors in automatically generating medical records. Two months after deployment, the average number of patients treated by doctors has increased by 50%.
To further reduce the threshold for using large models, Baidu continues to optimize the development tools on the Qianfan large model platform. In the quarter, the model customization tool ModelBuilder has been significantly upgraded, introducing a variety of mixed training datasets, which can support users in fine-tuning high-performance industry-specific models. At the same time, based on the continuous optimization of the AI-native application development platform AppBuilder, more than hundreds of thousands of applications have been created on the platform, covering online education, e-commerce, government affairs, and other industries.
The internal business reconstructed by AI has the potential to be released
"Baidu has always adhered to the development concept of 'application-driven', which has further expanded our competitive advantage and made us stand out among a series of competitors. We always believe that if we cannot develop practical applications on the basic large models, then developing more basic large models is useless," said Robin Li.
Therefore, Baidu has also deeply transformed its applications, such as Baidu Search and Baidu Wenku, iterating these products into applications with artificial intelligence capabilities.
It is reported that currently, 18% of the search results on Baidu search are generated by AI. AI-generated search results can not only provide users with more accurate and direct answers but also increase information that was previously unobtainable; its newly added interactive features support users in refining their needs through multi-round dialogue, enhancing the user experience. In addition, Baidu is also accelerating the distribution of intelligent bodies in search results, providing users with intelligent assistants.
From an industry perspective, since the outbreak of large models, various search engines have been exploring the capabilities of large models. In addition to Baidu, global technology companies such as Google, Microsoft, and Meta have all entered the AI search field. At present, most AI search engines mainly increase the capabilities of large models, and users only feel that they have added "AI-generated content" to the search.
Robin Li pointed out that intelligent bodies can help users solve complex problems and make decisions. For example, the college entrance examination intelligence body launched by Baidu in June can meet the personalized needs of different candidates and help them choose universities and majors. After the college entrance examination, the daily active user peak of this intelligent body approached 2 million, and its practicality was recognized.
Intelligent bodies have changed the way search is used and greatly improved the user's search experience. In the past, users used search engines to obtain mostly web pages. Now, when users enter their needs in the search box, Baidu search's "AI assistant" can accurately interpret user intentions and push the most matching intelligent bodies to the forefront of users, allowing users to call them at any time. The reconstruction of AI search by intelligent bodies will become a greater potential created by Baidu.
The distribution volume of intelligent bodies in the Baidu ecosystem is rapidly increasing, with an average daily distribution of more than 8 million in July, twice that of May. The most commonly used intelligent bodies include content creation, personality tests, schedule planning, and other types. Users, developers, service providers, and merchants are all important participants in the intelligent body ecosystem. Currently, there are already 16,000 merchants developing intelligent bodies on Baidu, covering education, law, and B2B industries.
Robin Li mentioned that although artificial intelligence has great potential for product iteration, this process is not something that can be achieved overnight. In the short term, the investment in artificial intelligence will have a certain negative impact on revenue. However, we believe that this transformation will ultimately lead to a completely different new type of artificial intelligence ecosystem, a comprehensive, efficient, and user-friendly ecosystem.