China's 130 Large Models, ChatGPT Revenue Breaks 10 Billion, Domestic Competition Still Needs Effort

[baoshi.rao]

China's AI large models are fully 'unleashed.'

According to Titanium Media App, on August 31, eight companies/institutions—Baidu (ERNIE Bot), ByteDance (Skylark Model), Zhipu AI (Zhipu Qingyan), Chinese Academy of Sciences (Zidong Taichu Model), Baichuan Intelligence (Baichuan Model), SenseTime (SenseChat Model), MiniMax (ABAB Model), and Shanghai AI Lab (Scholar General Model)—officially announced the public release of their large model products.

These are China's first batch of AI large model products offering services similar to ChatGPT, marking a critical moment for the large model industry.

Meanwhile, Alibaba's 'Tongyi Qianwen,' 360's 'Zhiniao,' and iFlytek's 'Spark Cognitive Model' have not yet announced full public access but have indicated they have completed regulatory filings and are expected to open within the next week.

Among the first batch of domestic large models, five are consumer-facing (C-end) products, free and open to the public, potentially positioning themselves as 'China's ChatGPT.' The exceptions are the ABAB, Scholar General, and Zidong Taichu models.

However, ChatGPT, with its first-mover advantage, is accelerating commercialization. According to The Information on August 30, OpenAI, the creator of ChatGPT, is projected to generate over $1 billion in revenue in the next 12 months from AI software and computing power sales. Even Nvidia, which 'sells the shovels,' saw an 843% year-on-year increase in net profit in Q2. In contrast, China's large models collectively face losses and struggle to achieve scalable revenue.

Data obtained by Titanium Media App from CCID shows that 64 large models were released domestically from January to July this year. As of July 2023, China has introduced a total of 130 large models.

Clearly, China's 130 large models are unlikely to surpass ChatGPT, which is rapidly commercializing and projected to generate $1 billion in revenue.

Eight Large Models Fully Open
Alibaba's Model Awaits Notification Within a Week

This year, AI large model technology, represented by ChatGPT, has taken the world by storm. Seizing this opportunity, domestic companies like Baidu, Alibaba, and ByteDance have successively announced their large model products.

On July 13, seven Chinese ministries jointly released the 'Interim Measures for the Management of Generative Artificial Intelligence Services,' effective from August 15, 2023. Under the new regulations, generative AI products, including large models, must undergo two regulatory procedures: algorithm filing and security assessment, known as the 'dual new assessments.' Only after passing both can generative AI products be legally launched.

Subsequently, companies such as Baidu, SenseTime, Alibaba, and Tencent began the filing process, hoping to be among the first batch approved. On August 31, eight enterprises/institutions officially launched their AI large models to the public.

As of the time of writing, data from Qimai shows that on August 31, Baidu's Ernie Bot had approximately 313,000 downloads on the Apple App Store, while SenseChat App had 3,832 downloads.

Specifically, Baidu announced that its large model application "Ernie Bot" is now fully open to the public. Users can download the "Ernie Bot APP" from app stores or visit the official website (https://yiyan.baidu.com) to experience it. The head of Baidu's MEG group stated that Baidu's various user-facing products are being restructured using large models. Products like Baidu Search, Baijiahao, Baidu Wenku, and Baidu Input Method are developing and rolling out AI features, with new products like NEWAPP to be launched in the future.

Baidu, leveraging its search portal, has added a "Chat" button at the top of search results. Clicking it opens a chat window, allowing anyone to access Ernie Bot, similar to how Microsoft integrated ChatGPT into its Bing search.

Robin Li, Baidu's Chairman and CEO, stated that opening Ernie Bot to hundreds of millions of internet users will provide vast amounts of real-world feedback, further improving the foundational model and accelerating iterations to enhance user experience.

Baidu's official data shows that on August 31, the first day Ernie Bot was available for download, daily active users exceeded 1 million, with the platform answering over 33.42 million questions in the past 24 hours.

SenseTime announced that its AI chatbot "SenseChat 3.0" is now open to all users. Users can register and use it via the SenseChat website (https://chat.sensetime.com) to experience features like text editing, mathematical calculations, and programming. SenseChat 1.0 was first launched in April this year as a large language model chatbot with capabilities in semantic understanding, multi-turn dialogue, knowledge mastery, and logical reasoning.

SenseTime's interim report released on August 28 revealed that, driven by its large-scale computing power and the "Day by Day New" large model system, generative AI-related revenue grew by 670.4% in the first half of the year, accounting for over 20% of total revenue. SenseTime estimates that generative AI business will contribute over 30% of revenue by 2024 and exceed 40% by 2025.

Baichuan Intelligence announced that its large model product "Baichuan Large Model" was among the first to pass the "Interim Measures for the Management of Generative Artificial Intelligence Services" filing and is now open to the public. Users can visit the official website (www.baichuan-ai.com) to experience it. Founded on April 10 this year by former Sogou CEO Wang Xiaochuan, Baichuan Intelligence has released three general large language models in four months: the 7-billion-parameter Baichuan-7B (the first open-source, free commercial model in China) on June 15, the 13-billion-parameter Baichuan-13B-Base and dialogue model Baichuan-13B-Chat on July 11, and the 53-billion-parameter Baichuan-53B on August 8.

Wang Xiaochuan, founder and CEO of Baichuan Intelligence, previously revealed to Titanium Media App that the company plans to release a trillion-parameter model in Q4 2023 and a "super app" around Q1 2024.

According to MiniMax, its 'abab' large model has fully opened services starting today. Users can log in to the 'MiniMax Open Platform' to experience it firsthand. The 'abab' model is reportedly suitable for most text-processing scenarios, including logical reasoning, text continuation, copywriting generation, content summarization, code generation, and knowledge retrieval. It also offers voice model capabilities. Currently, the MiniMax Open Platform has onboarded over a hundred paying customers, serving more than ten industry scenarios such as office collaboration, interactive entertainment, customer service, search, and education.

MiniMax was founded in December 2021 by Yan Junjie, former Vice President of SenseTime and head of general intelligent technology. It is a leading company in China's AI sector. Reports indicate that in June this year, MiniMax was close to completing a funding round exceeding $250 million, with investors including Tencent and miHoYo, valuing the company at over $1.2 billion (approximately RMB 8.5 billion).

Meanwhile, Zhipu AI announced the official launch of its first generative AI assistant, 'Zhipu Qingyan,' now available in major app stores. Users can download it via Apple App Store, Android app stores, or search for 'Zhipu Qingyan' on WeChat Mini Programs. Founded in June 2019, Zhipu AI originated from the Knowledge Engineering Laboratory of Tsinghua University's Computer Science Department and has completed three rounds of funding, with investors including Meituan, Huakong Fund, Qiming Venture Partners, Legend Capital, CAS Star, and Jiangmen Venture Capital.

Additionally, three other large models have been announced: the 'Shusheng General Large Model System,' jointly released in July 2023 by Shanghai AI Lab, SenseTime, The Chinese University of Hong Kong, Fudan University, Shanghai Jiao Tong University, and Tsinghua University; 'Zidong Taichu,' a 2.0 full-modal large model released in June 2023 by the Institute of Automation, Chinese Academy of Sciences, and Wuhan AI Research Institute; and the 'Lark' model, already applied in Douyin's AI chatbot 'Doubao,' launched on August 17, offering multi-turn dialogue services in Chinese and English.

Apart from the aforementioned models, companies like iFlytek and Alibaba have stated they have completed the regulatory filing process and are awaiting official announcements from local authorities. Alibaba mentioned that its Alibaba Cloud's 'Tongyi Qianwen' has completed filing and is pending launch. The first batch of approved companies is expected to be announced by local regulators within a week. iFlytek and 360 ZhiNao also confirmed completing their filings, while Tencent and Huawei have not responded regarding their filing status.

According to Titanium Media App, the first batch of officially announced large models are all from companies based in Beijing and Shanghai. 'After completing the filing process, the timing varies due to differences in company locations,' a source told Titanium Media App.

Caixin cited sources stating that approval progress varies across provinces and cities, with differences in quotas allocated to each region.

B-Side Gradually Finding Ground

OpenAI Estimated Annual Revenue of $1 Billion

Compared to consumer-facing (C-side) large models, several enterprise-level (B-side) AI models launched via cloud or applications are currently undergoing filing and evaluation, with no company yet receiving approval for launch. Commercially, they also face fierce competition from industry giants.

A representative from a relevant company told Titanium Media App that they are currently undergoing the filing process for their large model, which has completed the "dual new evaluation." They are awaiting official notification before making an announcement, hoping to be among the second batch of approved large models, with results expected soon.

Currently, multiple B-sector large models have completed algorithmic filings. On June 20 and September 1, two batches of domestic deep synthesis service algorithm filings were announced, including 360 Zhiniao's text generation algorithm, NetEase Youdao's machine translation algorithm, Kuaishou's generation and synthesis algorithm, Tian Gong's large language model algorithm, iQiyi's generation and synthesis algorithm, Skylark's large model algorithm, and Huawei Cloud's Pangu multimodal large model algorithm.

In reality, compared to the C-sector, industry and vertical large models are becoming the development model pursued by domestic companies.

Currently, B-sector large model business models are divided into three types: selling large model API interfaces, charging companies or developers based on usage; directly selling large model development services, providing industry solutions to traditional enterprises for revenue; and using large models to transform existing businesses, enhancing product competitiveness for greater commercial returns, known as Model-As-A-Service (MaaS).

Fang Han, CEO of Kunlun Wanwei, stated in July this year that large model development will trend toward fragmentation in the B-sector and free services in the C-sector. "Due to the demand for industry data and the fact that data is already fragmented across different B-sector segments, it is difficult for any single company to obtain all industry data, leading to B-sector fragmentation. Meanwhile, Chinese users have a strong preference for free services, so companies serving C-sector users will inevitably adopt a free model with VIP options as a supplement."

However, the challenge lies in the presence of ChatGPT Enterprise and the open-source Llama2 (Meta's large language model), which have lowered the entry barriers for the large model industry, making revenue generation difficult.

Based on public financial information from multiple companies involved in large models, Titanium Media App compiled the following revenue and business details:

• SenseTime: As of May this year, SenseTime has secured over 10 large model clients, more than 40 KA (key account) clients, coverage in over 5 cutting-edge fields, and implementation in over 20 industry scenarios.
• 360 Company: In the first half of this year, the company generated nearly 20 million yuan in revenue from AI services provided to SMEs based on 360 Zhiniao, primarily including membership fees for Soda Office and SaaS services for Enterprise Security Cloud.
• Baidu: In Q2 2023, Baidu Intelligent Cloud revenue reached 4.5 billion yuan, a 5% year-on-year increase, with non-GAAP operating profitability sustained.
• iFlytek: In the first half of this year, iFlytek reported revenue of 7.84 billion yuan, a 2.26% year-on-year decline, and net profit attributable to shareholders of 73.572 million yuan, a 73.54% year-on-year decrease.

Additionally, companies like Tencent and Huawei have previously mentioned launching dozens or even hundreds of large model solutions across multiple industries. MiniMax also announced that its enterprise-facing open platform has onboarded over a hundred paying clients.

However, so far, the market has not seen significant revenue growth from domestic enterprise-level large models. Even 360 Company's nearly 20 million RMB earnings come from software memberships and SaaS services for enterprise security cloud, with most firms not disclosing separate revenue figures attributable to large model technology.

In stark contrast to the "unremarkable" performance of domestic companies, international large model-related firms like ChatGPT and Nvidia are achieving rapid growth.

OpenAI: According to The Information's August 30 report, OpenAI, creator of ChatGPT, is projected to generate over $1 billion in revenue within the next 12 months through AI software sales and computational capacity.

Nvidia: In Q2, Nvidia's data center GPU chip-related business revenue reached $10.3 billion, a 171% year-over-year increase, with total net profit surging 843% to $6.188 billion.

CoreWeave: This general-purpose cloud computing platform built on Nvidia GPUs expects 2024 revenue to hit $1.5 billion.

Microsoft: Microsoft's Azure OpenAI GPT-4 cloud service for enterprises now serves 11,000 clients, doubling from 4,500 in mid-May. For fiscal year 2023, Microsoft's intelligent cloud revenue exceeded $110 billion, with Azure accounting for over 50% for the first time.

Overall, the large model sector shows dramatic divergence between domestic and international markets, with OpenAI demonstrating clear leadership and reshaping industry dynamics. As new rounds of large model deployments emerge, China's leading companies may advance in scenarios like digital marketing and text generation.

PitchBook data shows Q2 2023 generative AI investments in the U.S. grew 65% to $3.3 billion.

Gartner predicts over 100 million people will collaborate with generative AI by 2026, with nearly 15% of new applications being AI-generated by 2027. McKinsey forecasts generative AI could add $2.6-$4.4 trillion in global economic value.