Author: Xiangshan Finance

The large model competition in 2024 is intense.

Sora started the year with a bang, leaving the film and television industry bewildered. Subsequently, the industry has witnessed continuous financing and the rapid rise of emerging companies. Price wars and free models have echoed through the large model arena.

The battle for "hundreds of models" is bound to be brutal. Fortunately, even in the toughest fights, there are still signs to follow.

During Baidu's Q2 earnings call, Robin Li said that "Ernie Bot" has the highest daily calls and daily tokens processed domestically, with over 600 million daily calls, a more than 10-fold increase in the past six months, and over 1 trillion daily tokens used.

In summary, Baidu's Q2 earnings report shows that "Ernie Bot" has become a formidable competitor. Today, we try to find the key to the development of the large model competition by examining how "Ernie Bot" is leading the industry.

The clues to how "Ernie Bot" leads the industry

In a nutshell, "Ernie Bot" leads the industry because it offers high quality at low cost.

High quality:

The core of "Ernie Bot's" leadership lies in its solid foundation - a full-stack layout. Moreover, its iteration speed is sufficient. Over the past year, Baidu has continuously expanded its model portfolio and iterated on large model capabilities.

In June of this year, Baidu released "Ernie Bot" 4.0 Turbo, which is fully open to enterprise users. Through core technological innovations such as data and basic models, as well as joint optimization with PaddlePaddle, "Ernie Bot" 4.0 Turbo is faster and more effective.

In particular, the joint optimization between PaddlePaddle and "Ernie Bot" has resulted in a weekly training efficiency of 98.8% for "Ernie Bot," which is 5.1 times faster and 105 times more efficient in inference compared to when "Ernie Bot" was first released, with inference costs reduced to 1% of their original level.

It is worth mentioning that Baidu's deep learning framework PaddlePaddle was also upgraded to version 3.0 in June, which can significantly improve its compatibility with Baidu's AI infrastructure and "Ernie Bot," further reducing model inference costs in the future.

Low cost:

One of the keywords in the large model competition this year is price wars, and it can even be said that the large model competition has entered a free stage.

In May of this year, Baidu Intelligent Cloud was the first to announce that its ERNIE-Speed, ERNIE-Lite, and ERNIE-Tiny series model preset services would be free to customers. During the World AI Conference (WAIC) in July, Baidu Intelligent Cloud announced significant price reductions for its flagship ERNIE 4.0 and ERNIE 3.5 models, further helping enterprises reduce large model and trial-and-error costs.

Subsidy wars and free models are basic operations for tech companies. However, not many companies can successfully navigate the free model strategy from an ecological perspective.

In the view of Xiangshan Finance:

First, free is not a strategy but a capability.

While free may seem like a competitive strategy, it is actually a capability. Everyone knows that large models are expensive, so can companies afford to offer them for free?

Baidu's strategy is to continuously optimize and improve model performance through a four-layer technology stack while continuously reducing model inference costs.

In other words, Baidu's free model offering is not a hard sell but rather a low-cost, more sustainable approach. This capability is essentially cost reduction, which not only saves money but also creates a moat. When your costs are low enough, that becomes your differentiated core competitiveness.

Second, the free model competition is not about price wars but ecological battles.

The business logic of "making money from others while incurring costs elsewhere" that tech companies excel at requires ecological capabilities to succeed. For example, in this round of large model free offers, most participants are tech giants, while start-ups in the large model space have barely kept up. On the one hand, they lack the financial resources, and on the other hand, unlike giants with their own ecosystems, they cannot make money through cloud services even if their large models are free.

The free model can help "Ernie Bot" attract more developers, thereby quickly covering various industries. From a developer's perspective, enterprises have been reducing costs over the past two years, and the free "Ernie Bot" has alleviated their operational pressures. Moreover, as a representative of new productivity, large models embody the inclusive value of technological innovation. This is also the significance of revitalizing the country through science and technology. Leading tech companies are never self-indulgent but can drive industries and help more emerging enterprises.

From an industry perspective, free and discounted offerings are the general trend. For example, in the mobile era, images and text were initially popular on the web due to expensive data plans. Nowadays, videos are booming due to cheap data plans, which is a decisive factor. Following this logic, domestic large model companies' collective price reductions and free offerings are essentially building the infrastructure of an intelligent society.

Third, the free model competition is not a sprint but a marathon.

Over the years, Baidu has been running a marathon, adhering to a long-term perspective. Under this guiding principle, "Ernie Bot" has a more far-sighted strategic vision and can capture more market share. Data shows that since Baidu announced the free use of its two main large models, ERNIE Speed and ERNIE Lite, on May 21, the daily calls for these two models have increased tenfold in two weeks, attracting a large number of developers and enterprise users to the Qianfan large model platform for business innovation.

Large model support: Steady first curve, vibrant second curve

Another factor contributing to "Ernie Bot's" leadership is that AI is not just a business for Baidu but a way of thinking. At Baidu, AI-derived technologies, such as large models, empower external entities while continuously reshaping Baidu's business fundamentals.

The Q2 earnings report shows that 18% of search results are generated by AI, compared to 11% in Q1.

The AI-enhanced search has received widespread recognition. On August 8, Baidu Search ranked first in the domestic AI Product Ranking, one of the most recognized AI application rankings in China, in its first participation.

It is worth mentioning the intelligent agent. Robin Li, Baidu's founder, chairman, and CEO, stated in a speech that intelligent agents are his most promising AI application direction, as they are the simplest AI applications to develop. He believes that millions of intelligent agents will emerge in the future, forming a vast intelligent agent ecosystem.

Traditional search returns links, while intelligent agents provide answers, which undoubtedly complements traditional search and significantly improves search efficiency. Furthermore, the generative content of intelligent agents enriches Baidu's content ecosystem.

As a tool, search is all about efficiency and content. Intelligent agents are undoubtedly an evolution of search. At the same time, search will become the primary distribution channel for intelligent agents, making them super apps once they are fully operational.

With the support of large models, the first curve of search can be described as steady, while the second curve of intelligent cloud can be termed vibrant.

In Q2, Baidu Cloud's revenue was 5.1 billion yuan, a year-on-year increase of 14%, with AI revenue accounting for 9%, driving continuous growth in Baidu Cloud's business.

The logic behind Baidu Intelligent Cloud's growth is simple: increased demand. As large model applications become more widespread, the demand for model inference will surge, which will drive a new round of growth in cloud services. Citibank also predicts that due to the transformation of AI, cloud computing may accelerate its growth.

The growth of cloud computing reflects Baidu's business line's advantage of high synergy. As the volume of "Ernie Bot" increases, it also drives the growth of cloud computing business.

To withstand the deluge of traffic generated by the large model trend, Baidu Intelligent Cloud has launched the world's first one-stop enterprise-grade large model platform - Baidu Intelligent Cloud Qianfan Large Model Platform.

The Qianfan Large Model Platform provides flexible, diverse, and efficient large model services for enterprises and developers, helping them solve all problems in the development and application of large models. The entire toolchain for large model development, known as Qianfan "ModelBuilder," covers continuous pre-training, fine-tuning, evaluation, compression, and deployment of large models. Users can fine-tune specialized models economically and easily on Qianfan "ModelBuilder."

Qianfan AppBuilder provides a complete toolchain for developing AI-native applications based on large models, helping enterprises and developers build AI-native applications agilely and efficiently, lowering the threshold for AI-native application development. With continuous optimization of the AI-native application development platform AppBuilder, over hundreds of thousands of applications have been created on the platform, covering industries such as online education, e-commerce, and government affairs.

The core of any TO B service is to help customers succeed. By genuinely assisting enterprises, commercialization will naturally follow. According to the latest IDC report, Baidu Intelligent Cloud ranks first in the Chinese large model platform market with a share of 19.9%.

In the first half of this year, Baidu continued to lead the bidding rankings. Statistics show that from January to May 2024, nearly 260 domestic large model project bid announcements were issued, with Baidu winning the most bids (10), covering the most industries (8), and having the highest total bid amount (over 64 million yuan) among mainstream large model vendors. The corresponding project procurement units include leading central and state-owned enterprises such as the All-China Federation of Trade Unions and China Southern Power Grid, demonstrating that Baidu's large models are accelerating the transformation and upgrading of traditional industries and speeding up industrial implementation.

A typical case is the introduction of large models into villages through Baidu's "Village Chief Copilot" based on "Ernie Bot." At the beginning of this year, in Henan and Chongqing, Baidu Intelligent Cloud collaborated with local grassroots government departments to build a local government knowledge base and resident Q&A assistant based on the capabilities of Qianfan AppBuilder and "Ernie Bot," enabling convenient services for residents within WeChat groups 24/7. Currently, "Village Chief Copilot" has entered over 6,000 villages.

There are many other convenient cases like this: The Guiyang Public Security Traffic Management Bureau and Baidu Intelligent Cloud jointly deployed an AI traffic signal control system at 31 key traffic nodes in Guiyang's Guanshanhu core area. By introducing a large traffic signal control model, the system precisely senses and predicts traffic flow at various intersections and directions, dynamically adjusting signal light durations at intersections, and alleviating traffic congestion. Currently, the average vehicle delay in Guiyang during weekdays has decreased by up to 15%, and Guiyang's national urban congestion ranking has dropped from fourth to 20th place.

Large aircraft need longer runways

The tech industry has always had a problem of following trends blindly, like the "Hundred Regiments War" of the past and the current "Hundred Models War."

Eventually, when the trend passes, the pigs that flew into the sky fall back to earth and fade into obscurity.

The reason why "Ernie Bot" can lead the industry lies in a crucial factor: Baidu is an "AI native enterprise" that has been involved in AI business for over a decade, not just the past couple of years.

Tech giants bet on industry cycles rather than short-term trends. Companies that have led in various fields over the years are no exception, such as CATL, BYD, and Baidu.

These giants are well aware that Rome wasn't built in a day, and large aircraft need longer runways, just like electricity was invented and then applied.

In fact, after the Frenchman Hippolyte Pixii invented the hand-cranked DC generator in 1832, people explored electrical applications for a long time. It wasn't until 1882, when General Electric established the world's first public power station, that electricity began to enter factories and spread worldwide, opening the door to the electrical age, which took several decades.

The official birth of artificial intelligence as an independent discipline can be traced back to the Dartmouth Conference in 1956, where several scientists gathered to discuss the possibility of endowing machines with intelligence. This conference marked the official birth of AI as a discipline.

Just like electricity was invented and then applied, the technological diffusion and application of AI are just beginning to reach their peak.

Years ago, Baidu chose a long and lonely road, but fortunately, its ambitions are vast. Large aircraft naturally need longer runways, and it takes time to validate them. Fortunately, that time is drawing near.

For other large model companies, the best time to plant a tree was ten years ago; the second-best time is now. Grasp the present and persevere.

Only by advancing together can we better tap into new productivity, as one flower does not make a spring, but a hundred flowers do.

Disclaimer: This article is based on the company's statutory disclosures and publicly available information and provides commentary, but the author does not guarantee the completeness or timeliness of the information. Additionally, investing in the stock market involves risks, and investors should exercise caution. This article does not constitute investment advice, and investors should make their own decisions.