Baidu AI, making money through Killer Apps

Author: Zhang Yifan

Editor: Shen Siqi

Source: Hard AI

Baidu, a technology giant that rose to prominence in the PC era, is facing challenges and competition in the mobile internet era. It seems to have not fully seized the opportunities of the times and appears somewhat behind compared to other tech companies in this stage of rapid development.

As time enters the era of Artificial Intelligence (AI), Baidu hopes to use this opportunity to make up for its past losses in the mobile internet era and re-establish its leadership position in the technology field.

In the latest financial report released by Baidu, the company revealed the latest progress in its AI business:

• AI cloud service revenue increased by 12% year-on-year;
• 11% of Baidu search content is generated by AI;
• Large model API calls surged: in the past five months, it increased from 50 million to 200 million;
• Significant optimization of AI model performance: training efficiency increased by 5.1 times, and inference costs reduced to one percent;

Baidu's Chairman and CEO, Robin Li, stated during a conference call that the company is shifting from being internet-centric to AI-first, advancing the reconstruction of To C and To B businesses using the Wenxin large model. Currently, the Wenxin large model processes approximately 250 billion tokens of text daily and has an average daily call volume of 200 million times.

Robin Li emphasized that search is most likely to become the "killer app" of the AI era, and the AI reconstruction work of Baidu search is still in its early stages.

Furthermore, Baidu's CFO, Lu Rong, stated that in the coming quarters, Baidu will continue to support high-quality growth in its AI business. The company believes that as China's era of generative AI unfolds, it will bring more opportunities to Baidu.

"All in AI" - Baidu's New Opportunity?

To achieve Robin Li's goal of "shifting from being internet-centric to AI-first," Baidu is heavily investing in AI technology research and applications, especially in areas such as autonomous driving, deep learning, and natural language processing.

Baidu is trying to regain the ground lost in the mobile internet era.

• In the PC era, with its unique search engine technology, Baidu became the leader in domestic search;
• In the mobile internet era, as user traffic shifted from PC to mobile, the company, despite its multi-faceted layout, still lags behind Tencent/Alibaba;
• In the era of recommendation algorithms, it is significantly behind the decentralized model of ByteDance;
• Stepping into the AI era, based on its deep accumulation in the AI field, Baidu hopes to break through in this wave;

The effectiveness of AI is gradually highlighted in the financial report. In the first quarter of this year, the revenue of the intelligent cloud business reached 4.7 billion yuan, a year-on-year increase of 12%, with revenue from generative artificial intelligence and basic models accounting for 6.9% of the intelligent cloud business revenue.

2. AI Cloud Services, from Chips to Platforms

In 2016, Baidu CEO Robin Li proposed the cloud service strategy of "AI + Cloud + Big Data".

Since then, Baidu Intelligent Cloud has built a full-stack AI infrastructure of "Chips-Framework-Model-MaaS Platform", which supports end-to-end AI services from data storage to model training, deployment, and operation.

It is worth mentioning that Baidu's PaddlePaddle platform is already compatible with over 50 different chips, many of which are locally designed, and the developer community has grown to 13 million.

This AI infrastructure deployment model also allows Baidu to better match domestic AI cloud demands in the face of external constraints. Robin Li believes that this model enables Baidu to use less advanced chips for highly effective model training and inference.

During a conference call, Robin Li stated that Baidu integrates GPUs from different suppliers into a unified computing cluster for training large language models. He said, "Our platform has demonstrated very high efficiency on the GPU cluster, which consists of hundreds or thousands of GPUs. This is a significant breakthrough achieved against the backdrop of restrictions on imported GPUs."

1) Self-developed Chips

The company has independently developed AI chips (Kunlun) and speech interaction chips (Honghu), used to optimize AI technologies such as speech, natural language processing, and image, and support the company's deep learning framework.

Latest developments in chips:

• AI Chip: Iterated to Kunlun 2, with a 2-3 times performance improvement compared to Kunlun 1, providing 128TFLOPS computing power and 512GB/s memory bandwidth;

• Speech Interaction Chip: Iterated to Honghu 900, with a 200% improvement in CPU performance and 160% improvement in GPU compared to the previous generation Honghu 818 chip, deployed on Huawei Smart Screen V5 Pro;

2) Deep Learning Framework - PaddlePaddle

PaddlePaddle is Baidu's self-developed deep learning framework based on BERT, integrating core training and inference frameworks for deep learning, basic model libraries, end-to-end development kits, and a rich set of tool components.

This framework serves various industries including energy, finance, industry, healthcare, agriculture, etc. For example, Lianxin Medical has developed an "AI system for pneumonia screening and disease pre-assessment based on CT images" on the PaddlePaddle platform, which has been put into use at Xiangnan College Affiliated Hospital in Chenzhou, HunanPaddlePaddle has adopted an open-source approach to facilitate Baidu in building its AI ecosystem more effectively.

According to the company's disclosure:

• As of mid-April 2024, the PaddlePaddle developer community has reached 13 million members;
• By the end of 2023, PaddlePaddle served 235,000 companies, and the development community had created 860,000 models.

In addition, data shows that PaddlePaddle ranks second only to leading overseas platforms Pytorch and TensorFlow in terms of GitHub users, contributors, and technical iteration speed.

3) AI Large Models - Ernie Series Models

Based on the powerful base model (PaddlePaddle), Baidu has developed the Ernie series models (fun fact: the English name for 文心一言 is "Ernie," derived from a character in the American children's show "Sesame Street," and Ernie's good friend in the show is Bert, who is a Google AI model released in 2018).

Currently at version 4.0, the Ernie model may have a parameter scale exceeding 10 trillion, about 4-5 times that of Ernie 3.5.

According to the company's disclosure, as of April this year, the Ernie model processes approximately 200 million API calls per day, far exceeding the approximately 50 million calls in December last year, indicating that Baidu's Ernie model is being adopted by more and more people and signaling strong revenue potential for future model inference.

Furthermore, to make the model more cost-effective, the company continuously improves the model's efficiency through its unique four-layer AI architecture and powerful end-to-end optimization capabilities. Compared to the version on March 23, the training efficiency of the Ernie model has increased by 5.1 times, and the inference cost has been reduced to one percent.

4) MaaS (AI Development Cloud Platform)

In order to simplify the development of AI models for developers, Baidu provides three sets of tools on its cloud platform:

• AppBuilder and ModelBuilder: for enterprises and individual developers to develop applications and build models;
• AgentBuilder: to easily create an AI Agent.

In terms of profitability, starting from 2022, Baidu has focused on improving the profit margin of its AI Intelligent Cloud business, eliminating low-quality businesses, and striving to customize standardized AI solutions for customers in different industries.

During the 2024 Q1 earnings call, the company stated that the growth in intelligent cloud revenue is mainly driven by generative AI and model training. Currently, most of the revenue comes from model training, but revenue growth from model inference is rapidly increasing.

Looking ahead, Baidu believes that model inference is one of the most important long-term opportunities and a key driver of future revenue growth for artificial intelligence cloud services

Section Three: Empowering "AI+" Business

AI technology is widely used within Baidu in areas such as search engines, information flow recommendations, Baidu translation, and more. During a conference call, Robin Li expressed strong confidence in the future AI empowerment of the core search business, stating that search is the product most likely to become a killer application in the AI era: "AI search allows users to do things they couldn't do before, and Baidu currently has no plans to charge for this."

Since the second quarter of last year, Baidu has been using Wenxin Yiyuan to rebuild Baidu search. Now, more and more search results are composed of Wenxin Yiyuan in different formats such as text, images, third-party links, and more.

In terms of model construction, Baidu has utilized 3 lightweight models and 2 models tailored for specific tasks, combined with expert hybrid models to allocate tasks to large models, achieving a better balance between performance and cost.

From a data perspective, AI is indeed enhancing the efficiency of various Baidu businesses:

1. AI+ Baidu Search: As of 24Q1, 11% of Baidu search content is generated by AI, and AI search has driven the growth of advertising revenue.

2. AI+ Baidu Wenku: By introducing generative AI functions, 18% of new paying users have been attracted. These functions include content summarization, creation, expansion, and one-click conversion of ideas into Powerpoint presentations. In 24Q1, the number of paying users in Baidu Wenku achieved double-digit year-on-year growth.

3. AI+ Baidu Maps: The introduction of the "AI Guide" function intelligently calls upon numerous map functions and services based on understanding user needs, providing solutions quickly and accurately.

4. AI+ Baidu Netdisk: The launch of the intelligent assistant "Cloud One" based on the Wenxin large model helps users quickly search for files and videos, summarize knowledge, translate documents, and even engage in content creation with just one sentence. As of 23Q3, the number of "Cloud One" users has reached 20 million.

Section Four: Cooperation with Terminal Manufacturers in AI Phones/AI PCs

In addition to improving efficiency internally, Baidu has further expanded its external cooperation this quarter.

According to the company's disclosure, in the previous quarter, the company had already partnered with smartphone manufacturers such as Samsung China and Honor. This quarter, the cooperation has expanded to more leading smartphone manufacturers, including Oppo, VIVO, and Xiaomi.

The business scope has also expanded from smartphones to PCs and electric vehicles, further enlarging Baidu's AI ecosystem. Lenovo is leveraging the Ernie API to support its AI assistant in the default browser, while Nio has started using the Ernie API to enhance the in-car experience.

Furthermore, the company has attracted many new customers, including Ctrip, Gaotu, Zhaopin, Zuoyebang, and the Singapore Tourism Board, helping them reconstruct all consumer-facing products through AI technology to provide a better user experience

V. Outbound Investments

In addition to independent research and development, Baidu also continuously expands its AI ecosystem through outbound investments, covering the three major aspects of chips, models, and model applications. However, it is obvious that, unlike Alibaba, Baidu's outbound investments tend to focus more on the upper and lower ends.

This reason is not difficult to understand, as Baidu will rely on application-side support such as search for future income to complete the business model loop.

One of the most representative proofs comes from Robin Li's judgment on whether large models should be open source or closed source: "There is little significance in open sourcing large models, they must be closed source to establish a viable business model, which can make money, and only by making money can we gather computing power and talent."

"In terms of cost, closed source actually has an advantage. As long as the capabilities are the same, the inference cost of closed source models will definitely be lower, and the response speed will definitely be faster." This is also a different business model thinking from Alibaba Cloud.

Therefore, following Baidu's strategy, the future path is already quite clear, but the competition in the AI era will obviously be more intense than in the PC era