In May 2023, Beijing issued the "Several Measures for Promoting the Innovation and Development of General Artificial Intelligence in Beijing", announcing 21 measures such as "carrying out research on large-scale model innovation algorithms and key technologies", taking the lead in launching the first shot of large-scale models among cities across the country. More than a year later, Beijing has formed a large model "roadmap" that comprehensively covers the entire industrial chain of large models, AI computing power, data, applications, solutions, and the integration of industry, academia and research. In Beijing, we see not only China, but also the world. Behind every "first", "strongest" and "heavy" Beijing's progress, China's large-scale model industry has moved from catching up to leading. Why Beijing? How did Beijing occupy the C position in the "100 model war"?

Author | Li Shuiqing

Editorial Review | Time

Edit | Peninsula

Source | Zhi Dongxi (ID: zhidxcom).

In 2024, it is quite exciting to count China's various breakthroughs in the field of large models:

Domestic models such as Baidu Wenxin Yiyan and Zhipu GLM-4 defeated GPT-4 in the global authoritative test; Three domestic large-scale model unicorns such as the dark side of the moon broke through the "mark" of 20 billion yuan in valuation; "Domestic Sora" dark horses Aishi Technology and Shengshu Technology have refreshed the AI video generation time to 6 seconds, 14 seconds, and 30 seconds; Vanka cluster breaks out of domestic GPU dark horses such as Moore Threads, and supports the training of trillion-parameter AI models......

These are all related to Beijing, China's "first AI city".

In July this year, the relevant person in charge mentioned at the artificial intelligence special forum of the 2024 Global Digital Economy Conference that as of that time, there were 166 large-scale model products that had passed the record in the country, of which 71 were in Beijing, accounting for nearly half of the country. From public information, it can be seen that the number of unicorns in Beijing accounts for eighty percent of the country, and at the same time, these large models have accelerated their implementation in many fields such as the Internet, education, medical care, finance, and government.

△2024 Global Digital Economy Conference Artificial Intelligence Forum

At the end of July, Zhidong counted the financing of more than 10 million yuan in the first half of 2024 and in July, and the results showed that Beijing enterprises received 43 financings, while Shanghai and Shenzhen received 11 and 9 respectively, showing an obvious gap.

There are all kinds of signs that Beijing is becoming China's "first capital of big models".

From the start to the lead, it took just over a year for the breakthrough of Beijing's large model industry. In May 2023, Beijing issued the "Several Measures for Promoting the Innovation and Development of General Artificial Intelligence in Beijing", announcing 21 measures such as "carrying out research on large-scale model innovation algorithms and key technologies", taking the lead in launching the first shot of large-scale models among cities across the country.

More than a year later, Beijing has formed a large model "roadmap" that comprehensively covers the entire industrial chain of large models, AI computing power, data, applications, solutions, and the integration of industry, academia and research. In Beijing, we see not only China, but also the world.

Behind every "first", "strongest" and "heavy" Beijing's progress, China's large-scale model industry has moved from catching up to leading. Why Beijing? How did Beijing occupy the C position in the "100 model war"? Let's find out.

1. Nearly half of the top models in the country are "made in Beijing"

Entering 2024, the strength of Beijing's "abundance" of large models has been continuously confirmed.

At the 2nd Beijing Artificial Intelligence Industry Innovation and Development Conference in March, the relevant person in charge of Beijing mentioned that the number of AI-related enterprises in Beijing is about 2,200, accounting for about 4% of the country;

At the 2024 Global Digital Economy Conference in July, Beijing announced that 71 large models have been filed and launched, accounting for nearly half of the country's total of 166.

Previously, in October 2023, the "White Paper on the Innovation and Application of Large Models in the Beijing Artificial Intelligence Industry (2023)" showed that about half of China's large model manufacturers with more than 1 billion parameters are in Beijing, and the number is as high as 122.

Star large-scale model products and the company's "Made in Beijing" have happened one after another in China's science and technology circle.

Major technology companies have set off a "100-model war", and Beijing has become the most cutting-edge innovation base. Alibaba, Tencent, Baidu, Byte, Huawei, 360, JD.com, Xiaomi, NetEase and other major technology manufacturers have announced the latest progress in the research and development and application of large models.

For example, in October 2023, Baidu launched version 4.0 of the Wenxin large model benchmarked against GPT-4, and more than 10 AI native applications and large model hardware were launched, and the user scale has exceeded 100 million by the end of the year.

The domestic large-scale model "Entrepreneurial Five Tigers" has emerged and become a unicorn with a valuation of nearly 20 billion yuan. Among them, Zhipu AI, Baichuan Intelligence, Dark Side of the Moon, and Zero One Everything are registered in Beijing, and MiniMax is located in Shanghai.

Founded in 2019, Zhipu AI will launch a new generation of pedestal model GLM-4 in January 2024, and the overall evaluation results are close to GPT-4. In July 2024, the "Chinese version of Sora" will be launched, which supports the generation of 6-second videos. The company has received more than $400 million in financing this year, and its latest valuation exceeded 20 billion yuan.

Founded in 2023 by former Sogou CEO Wang Xiaochuan, Baichuan Intelligence has launched 12 large models, including Baichuan-7B/13B, Baichuan2-7B/13B, Baichuan2-192K, Baichuan 4, and "Pengcheng-Baichuan· Mind 33B". The company just completed a 5 billion yuan A round of financing in July this year, and its valuation has also exceeded the 20 billion yuan "mark".

Founded in 2023, the dark side of the moon has built Kimi into an AI explosive product with more than 20 million monthly visits with its unique "To C" strategy, and the company has received more than $1 billion in financing in February 2024, setting a new record for the largest single financing in the domestic AI field, with the latest valuation of up to $3 billion, which is also over 20 billion yuan.

The AI2.0 company incubated by Dr. Kai-Fu Lee personally opened sourced the Yi-9B model in March 2024 and released the Yi large model API open platform, and it has recently been rumored to have completed a new round of financing of hundreds of millions of dollars.

Some large-scale model "new seedlings" have not yet entered the unicorn camp, but they have only been established for about 1 year and have become the "kings" of the subdivision track.

In the popular field of AI video generation, Aishi Technology, led by Wang Changhu, the former head of Byte Vision Technology, launched the overseas version of PixVerse, a Sora-like video generation product, as early as January 2024, and Shengshu Technology, led by Professor Zhu Jun of Tsinghua University, launched China's first original self-developed video model Vidu with long duration, high consistency and high dynamics in April 2024. All ahead of OpenAI's "futures" product Sora.

In the field of AI Infra, which has attracted much attention, SiliconCloud, a silicon based company founded by former OneFlow founder and CEO Yuan Jinhui, just launched a one-stop large model API cloud service platform SiliconCloud in June, and it is reported that its self-developed SiliconLLM large model inference engine is more than 10 times faster than similar open source products.

Industry-university-research joint innovation has become a trend. For example, in the field of general large models, Beijing Academy of Artificial Intelligence and China Telecom jointly released the world's first low-carbon, high-performance open-source multilingual large model Tele-FLM in July 2024, and announced that 100 billion, 600 billion, and trillion parameter large language models will be released next.

The application of large models in the industry has achieved initial results. In June 2023, Beijing released the first batch of 10 typical application cases of industry large models, and 10 application cases of technology suppliers including Baidu, Zhipu Huazhang, Institute of Automation, Chinese Academy of Sciences, iFLYTEK, Yunzhisheng, Scientific Intelligence Research Institute, 4Paradigm, Momo Zhixing, Zhiyuan, and Facewall Intelligence were selected, covering energy, healthcare, finance, consumption, intelligent driving, government, real estate and other industries.

In July 2024, the Beijing Industry Large Model Innovation and Application Competition was awarded, and finally 6 companies won the first prize, namely Rely Technology (government affairs), Zhongke Wenge (finance), Ruibo (industrial upgrading), Shukun Technology (medical), Zhongke Brain (culture and education), and Zhongke Ruijian (smart city).

In just one year, the number of Beijing's main large models has grown from 40 in July 2023 to 71 in July 2024.

Beijing has run out of China's speed in the development of the large-scale model industry, and also out of China's intensity.

2. The four E-level intelligent computing centers launched 36,000P intelligent computing power to activate data elements

To open the "model" power of Beijing, it is inseparable from the "key" of computing power.

As of December 2023, Beijing has formed a computing power supply scale of 12,000P (PFlops, petaflops per second) - including 8672P of general computing power, 3402P of intelligent computing power, and 340P of supercomputing power. The intelligent computing power deployed by the capital cloud service provider in Tianjin, Hebei, Jinmeng and Mongolia is about 4082P, which provides a solid foundation for the development of the large model industry.

What is this concept? 1P refers to the speed of 1,000 trillion calculations per second, and 100P is equivalent to the computing power of 50,000 high-performance computers, for example, to locate a certain characteristic star in the sky map of 200,000 celestial objects, only 100 seconds. What kind of "modular" power will 12,000P bring to Beijing can be imagined.

Entering 2024, Beijing's intelligent computing infrastructure construction will remain strong, providing continuous energy supply for the development of the large model industry.

In January 2024, the Beijing Artificial Intelligence Public Computing Platform (Shangzhuang) located in Haidian District was released, and the first phase of 500P computing power has been officially launched, providing inclusive computing services for the development of the large model industry.

In March 2024, as the largest public computing platform in Beijing, Beijing Yizhuang Artificial Intelligence Public Computing Platform was officially launched, which can provide computing power of 3000P, which is equivalent to the computing power of more than one million high-performance computers.

The district will issue "100 million computing power coupons + 100 million model coupons" service enterprises every year; Facing the future, the economic development zone where the platform is located has planned a computing power scale of more than 10,000P.

In May 2024, the infrastructure construction of the Beijing Digital Economy Computing Center project in Chaoyang District has started construction, which is expected to be completed by the end of 2024, and will achieve a total of 2,000PFlops of intelligent computing power supply after it is put into operation.

△ Renderings of Beijing Digital Economy Computing Center

According to the newly released "Beijing Computing Infrastructure Construction Implementation Plan (2024-2027)", Beijing plans to build four E(EFlops, floating point operations per second) level intelligent computing power highland, and by 2027, the new intelligent computing power will reach 36,000P.

Beijing not only provides impetus for the development of the large model industry from the level of computing power, but also provides sufficient "fuel" from the aspect of data elements.

According to data from the Beijing Municipal Bureau of Economy and Information Technology, after the Beijing International Big Data Exchange achieved the landing of the license, as of November 2023, the exchange has issued 27 data asset registration certificates, with a data transaction scale of more than 2 billion yuan.

Behind the scenes, Beijing took the lead in creating the country's first data-based system pilot zone, and released a list of plans and policies for the creation of data-based system pilot zones; Innovative incentives such as the first registration, first transaction, first entry and first opening of data assets have been introduced to continue to release policy dividends.

In March 2024, Beijing will open the first AI data training base in China, and at the same time, the "regulatory sandbox" system will ensure the training of large models. The AI data training base and Beijing computing infrastructure have joined forces to greatly help enterprises in the field of large models to train their "strongest brains".

It can be seen that by grasping the three major elements of model, computing power and data, Beijing has grasped the "bull's nose" of the development of the large model technology industry.

3. 30 years of AI evolution: Beijing has become China's "No. 1 capital of large models"

How did Beijing become the capital of AI? It all started more than 30 years ago.

The rise of Beijing's AI industry has a long history. Artificial intelligence research at the Chinese Academy of Sciences began in 1987, Microsoft Research Asia in 1998, and the famous Tsinghua Yao Ban in 2005...... They have created batches of talents for Beijing's AI development, and with the first wave of AI set off by SenseTime, Megvii, and KLCII in the early 2010s, they have also become the mainstay of the current large-scale model industry development in the 2020s.

With the maturity of the AI technology industry represented by computer vision, 2020 is also the first big outbreak of AI in Beijing.

According to the "2020 Beijing Artificial Intelligence Development Report" released by Beijing Academy of Artificial Intelligence, the scale of Beijing's AI-related output value in 2019 has reached 170 billion yuan; As of the end of June 2020, the total financing of AI startups exceeded 300 billion yuan, and 32 AI unicorn companies such as SenseTime and Megvii were born; The number of AI-related papers published and patent applications ranked first in the country, and a number of major original achievements with international influence were produced, all of which ranked first in the country.

This is inseparable from Beijing's ability to step on the node of AI programmatic policies.

In September 2017, the Management Committee of Zhongguancun Science and Technology Park took the lead in releasing Beijing's first AI policy, the Action Plan for Cultivating the Artificial Intelligence Industry in Zhongguancun National Independent Innovation Demonstration Zone (2017-2020), and in December, Beijing issued the Guiding Opinions on Accelerating Scientific and Technological Innovation and Cultivating the Artificial Intelligence Industry in Beijing, providing a policy basis and support for the development of Beijing's AI industry. In 2019, Beijing was approved to establish the first national new-generation artificial intelligence innovation and development pilot zone, which also provided conditions for its AI explosion.

With the explosion of large model applications represented by ChatGPT around the world, 2023 is the year of evolution of Beijing's AI industry.

1. Fired the first shot of the national AI large model policy, and 100 billion yuan to support the investment and financing environment

In terms of programmatic policies for AI large models, Beijing got up early, did not hesitate, and did not wait and see.

On May 30, 2023, Beijing took the lead in firing the first shot of the local large-scale model policy. The Beijing Municipal Government issued two major policies, the "Implementation Plan for Accelerating the Construction of an Artificial Intelligence Innovation Source with Global Influence" and "Several Measures for Beijing to Promote the Innovation and Development of General Artificial Intelligence", and proposed 21 measures to strengthen the research and development of large model training data collection and governance tools, setting the tone for the development of Beijing's large model industry.

In September 2023, Haidian District, Beijing will be the first to make proofs. The district issued the "Implementation Plan for the Innovation-led Development of General Artificial Intelligence in Zhongguancun Science City" (hereinafter referred to as the "Plan"), which proposes to build a core industry scale of 230 billion yuan, and give talent subsidies from various aspects such as talent settlement, capital subsidies, and public rental housing.

In April 2024, Beijing will make another policy move. The National Development and Reform Commission and other departments further jointly issued the "Several Measures of Beijing Municipality on Accelerating the Leading Development of the General Artificial Intelligence Industry" (hereinafter referred to as the "Measures"), which focused on improving the supply of intelligent computing power and accelerating the innovative application of large models.

These "gift packages" include, but are not limited to: a maximum of 100 million yuan for AI technology research projects included in major national strategic tasks, a maximum support of 50 million yuan for projects to build artificial intelligence benchmark applications, an investment of more than 100 billion yuan in the next five years to optimize the investment and financing environment, and reduce the operating costs of large model enterprises through capital subsidies, the issuance of computing power coupons, model coupons, etc.

2. Build 45EFLOPS intelligent computing supply, and domestic computing power supports trillion-level parameter large model training

In terms of large-scale model infrastructure construction, Beijing has maintained its previous strong investment.

In April 2024, the Beijing Municipal Bureau of Economy and Information Technology and the Beijing Municipal Communications Administration issued the Implementation Plan for the Construction of Computing Power Infrastructure in Beijing (2024-2027) (hereinafter referred to as the "Plan"). The plan proposes that by 2025, the supply scale of intelligent computing in Beijing will reach 45EFLOPS; By 2027, the full-stack independent and controllable software and hardware products of intelligent computing infrastructure will be realized, the overall performance will reach the leading level in China, and the ability to build a 100% independent and controllable intelligent computing center will be realized.

Beijing also announced the creation of more than 200 industry benchmark cases driven by intelligent computing power, focusing on the construction of E-level intelligent computing power highlands such as Haidian, Chaoyang, Yizhuang, and Jingxi (Shijingshan and Mentougou), and giving priority to accelerating the construction of two 10EFLOPS large-scale intelligent computing clusters.

However, Beijing's computing power territory does not stop there, building a Beijing-Tianjin-Hebei-Mongolia computing power supply corridor, planning to build a super-large-scale intelligent computing cluster that supports the training needs of trillion-level parameter large models, and accelerating the commercialization of artificial intelligence large models.

It is worth mentioning that accelerating the independent and controllable core software and hardware products will be a key task.

In the plan, Beijing proposes that for enterprises that purchase independent and controllable GPU chips to carry out intelligent computing services, the safeguard measures proposed in the implementation plan are to provide support according to a certain proportion of the investment amount to accelerate the realization of independent and controllable supply of intelligent computing resources.

In May, Beijing also clearly proposed to promote the breakthrough of domestic artificial intelligence chips in the "Implementation Plan for Accelerating the Construction of an Artificial Intelligence Innovation Source with Global Influence (2023-2025)". Actively guide large-scale model R&D enterprises to apply domestic artificial intelligence chips, and accelerate the localization rate of artificial intelligence computing power supply.

Beijing's real investment in domestic computing power is also a "big deal".

For example, in October 2023, the Beijing Municipal Bureau of Economy and Information Technology issued the "2023 Beijing High-tech Industry Development Fund Implementation Guide (Second Batch)" last night, proposing to focus on supporting the first round of tape-out rewards for integrated circuit design products, and the annual reward amount for a single enterprise should not exceed 30 million yuan.

3. Build a nest to attract phoenixes, mobilize the national scientific and technological forces to "form a bureau", and invigorate the ecology of government, industry, academia, research and research

The ecological integration of government, industry, academia and research is the source of vitality for the development of Beijing's large-scale model industry.

Building a nest to attract phoenixes, Beijing has allocated a number of plots of land in the capital to develop the AI industry.

In September 2023, the "Zhongguancun Artificial Intelligence Large Model Industry Cluster" was unveiled, creating four industrial parks: Wudaokou Artificial Intelligence Industrial Park, Peking University West Gate Artificial Intelligence Industrial Park, Zhongguancun West District Artificial Intelligence Industrial Park, and Tsinghua Science and Technology Park, covering an area of 670,000 square meters, which provides a broad physical space for the grand blueprint of Beijing's large model industry.

In July 2024, the concept of an artificial intelligence innovation block was released. Beijing plans to build an artificial intelligence innovation block in an area of 53 square kilometers in the south of Haidian, connecting 52 national key laboratories, 10 new R&D institutions, 106 national scientific research institutions, 37 top universities, 89 AI2000 global top scholars, 12,300 artificial intelligence scholars, 1,300 artificial intelligence enterprises, more than 90 large model-related units, and 37 registered large model units.

Through a series of partnership programs and alliances, Beijing has "livened" the AI startup space.

In May 2023, Beijing released the General Artificial Intelligence Industry Innovation Partnership Program, announcing the list of the first batch of 39 members that month, and the second batch of 63 companies in July, including computing partners such as Alibaba Cloud and Beijing Super Cloud Computing Center, data partners such as Beijing Big Data Center, Beijing Academy of Science and Technology, and Beijing Academy of Artificial Intelligence, model partners such as Baidu, Zhipu AI, Ali Damo Academy, Kunlun Wanwei, Beijing Municipal Service Administration, Bank of Beijing, Application partners such as Kingsoft Office, investment partners such as IDG Capital, CDH Innovation, and MiraclePlus.

In July 2024, the Zhongguancun Artificial Intelligence Industry Alliance was established to launch the general artificial intelligence industry innovation ecological plan. Alliance members include Beijing Institute of General Artificial Intelligence, China Academy of Information and Communications Technology, Peking University, Tsinghua University, Chinese People's University, Institute of Automation, Chinese Academy of Sciences, KLCII, Haitian AAC, Zhipu AI, Tencent, CRRC, Baidu, Baichuan Intelligence, Zhongguancun Science City Company and other hundreds of universities and institutions.

It can be seen that in order to reflect the market-oriented open cooperation docking mechanism, Beijing has mobilized most of China's investment, industry and academic circles to jointly help the development of the large model industry.

In the end, the results of these actions are not just "showing muscles", but proofing for the whole country, and using the results to feed back the national large-scale model industry and the development of the digital economy.

4. Conclusion: To build a benchmark for the digital economy, Beijing occupies the C position of the "100 model war".

From digitalization to intelligence, Beijing is once again at the forefront of the times.

In 1994, when China had its first access to the Internet, Beijing took the lead in establishing the Municipal Information Office in 1997, and in 1999 formulated a long-term three-step goal of "Digital Beijing": from 2000 to 2010, the basic framework of "Digital Beijing" was formed. From 2011 to 2020, improve the digital level of the capital. From 2021 to 2050, Beijing will be built into a modern international digital city.

Leading the trend of Internet broadband, high-definition digital television, etc., stepping on the outlets of e-commerce, digital logistics, digital government, etc., we have created the "Window of the Capital" website that has gone down in history, the popularization of municipal transportation card, the 2008 digital Olympics, and the 3G network service network...... According to official data, the added value of Beijing's digital economy reached 871.9 billion yuan in 2015 and has grown to 1.7 trillion yuan in 2022, becoming one of the main engines of economic growth.

As an important sub-topic of digitalization and a representative of new quality productivity, AI has not only taken root in the fertile soil of Beijing's digitalization, but also sprouted under the watering of active AI special policies. With a multi-pronged approach of policy, industry, learning and research, capital and application, Beijing has become the leading talent center of the artificial intelligence industry in China and even in the world.

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com