ANBOUND advocates that China should not develop large AI systems, touted as "big AI", but instead focus on smaller ones. When discussing the rationale behind such analysis and proposition, the cost factor is one crucial point.

At the end of 2022, ChatGPT emerged and signaled the age of the rapid development of AI globally. From the United States to China, large language models similar to ChatGPT have mushroomed, becoming a valuable resource for various technology companies to showcase their strengths and attract capital. Public information shows that by the end of March 2024, nearly 120 large models have been registered and launched in mainland China. The large models that have completed filing for generative AI services include Baidu’s Ernie Bot, Alibaba’s Tongyi Qianwen, Huawei’s Cloud Pangu, Tencent’s Hunyuan, OPPO’s AndesGPT, Vivio’s BlueLM and others. If including large models that have not been registered, the total number of AI large models in China reached 238 in 2023.

From the perspective of speed and quantity of research and development, China does not seem to lag far behind the world's leading level in terms of AI large models. At many large model launch events in the countries, major tech companies emphasize the superiority of their large model performance, often claiming to be "comparable to" or even "ahead of" ChatGPT. For example, at the upgrade release conference of the StarFire Cognitive Large Model V1.5 held by iFlytek in June last year, there were several mentions of “StarFire is just a step away from ChatGPT" and "will surpass ChatGPT before October this year".

In reality, Huawei's claim of being significantly advanced has established a precedent where future competitors may feel reluctant to speak up unless they can also demonstrate superiority. When assessing the effectiveness of large models, it is essential to focus on their performance under more challenging conditions rather than just their peak capabilities. Models that undergo repeated training on smaller yet high-quality datasets often exhibit strong reasoning and decision-making abilities, a trait many large-scale models currently highlight. However, these models frequently encounter challenges when deployed across diverse user bases in real-world markets.

At a previous academic conference on China's digital economy development and governance, Xue Lan, Dean of Tsinghua University's Schwarzman College, pointed out that many Chinese large models are essentially "shells" or "assemblies" based on open-source models from abroad. There is still greater room for improvement in the logic and reasoning capabilities of such Chinese large models compared to foreign counterparts.

It is well known that the three elements of large models are algorithms, computing power, and data. With open-source large models, the algorithmic structure has relatively limited constraints on Chinese large models. Similar to how the complete open-sourcing of the Android system fundamentally changed the landscape of smartphone operating systems, developers can now shift their focus to other elements and commercial applications of large models. The real constraints on China's development of large models are computing power and data, which form a formidable cost investment gap in AI hardware that China finds difficult to overcome. In this regard, ANBOUND’s founder Kung Chan noted that due to objective conditions, China should not stubbornly follow the direction of the United States in developing general large models. Instead, China is suitable for developing what he calls "small AI", models oriented towards industry and products with strong application backgrounds.

The current AI market in China remains in a state of fierce competition. Under the pressure of rapidly declining costs and frequent product updates from foreign giants like OpenAI, Chinese large-scale AI models have initiated a massive price war without fully polishing their quality. Companies such as ByteDance, Alibaba Cloud, Baidu, iFlytek, Tencent, and others have recently announced significant price reductions for their large models, some even offering them for free. However, this price war in the large-scale AI model industry is not something that every enterprise can afford. In the era of large models, the costs of model training and inference pose growth traps that every large model startup must confront.

Behind the world's most advanced large models are powerful backers with significant financial capabilities, with investors continuously pouring funds into them. For example, Microsoft reported capital expenditures of USD 14 billion in the last quarter, a 79% increase year-over-year, largely due to AI infrastructure investments. Alphabet, Google's parent company, reported expenditures of USD 12 billion, a 91% increase year-over-year, with expectations that spending in the second half of the year will match or exceed this level. Meta has also increased its investment expectations for the year, projecting capital expenditures between USD 35 billion and USD 40 billion, a potential 42% increase at the upper limit of this range. These large AI projects involve investments often exceeding hundreds of billions, even surpassing the investments in semiconductor chip waves.

The challenges posed by the need for increased training data, larger models, additional semiconductor chips, expanded data centers, and heightened energy consumption for computational power are formidable for technology firms. These factors can exert significant strain on the majority of Chinese large-scale AI initiatives. The development and maintenance costs associated with advancing AI capabilities can be seen in several critical areas:

Firstly, there is the cost of data for AI models. As models become increasingly complex, the demand for high-quality training data continues to rise. Initially, early generative AI models were primarily trained using images, text, audio, video, and other data sourced from public web pages, some of which were copyrighted. However, due to escalating legal challenges, AI companies have started opting for paid data sources. Online community Reddit has claimed to have earned hundreds of millions of dollars by providing licensed data to entities such as Google and OpenAI. The AI training data market is projected to grow from approximately USD 2.5 billion currently to nearly USD 30 billion within the next decade. ByteDance, for instance, was previously reported to have used data generated by ChatGPT to train its large language models, leading to ChatGPT eventually banning access to its API. Additionally, data cleaning and annotation are crucial stages in ensuring data quality, requiring substantial human and automated resources to handle inconsistent and duplicate data, thereby increasing the costs associated with large-scale models.

There is also the training cost of AI models. OpenAI CEO Sam Altman stated that the cost of training GPT-4 exceeded USD 100 million. Dario Amodei, CEO of AI startup Anthropic, also mentioned that the training cost for current AI models in the market is around USD 100 million. Amodei further indicated that models currently under training, as well as those expected to be launched later this year or early next year, could cost close to USD 1 billion. By 2025 and 2026, the training costs for their company's models are projected to approach USD 5 billion to USD 10 billion.

Another crucial part of the expenditures is on developing AI chips. Even with a steady supply, these costs remain substantial as chips determine the computational power of models. Reports indicate that Nvidia's H100 graphics chip is priced around USD 30,000. Meta CEO Mark Zuckerberg previously mentioned plans to purchase 350,000 H100 chips by the end of this year to support their AI research efforts. All these investments amount to hundreds of billions.

Finally, there are the costs associated with cloud service centers, where the cost of each data center is measured in billions. For instance, Microsoft and UAE's AI company G42 announced a collaboration to invest USD 1 billion in building a data center in Kenya and EUR 4 billion in constructing AI data centers and cloud infrastructure in France. Over the past two years, Amazon has also committed USD 148 billion to constructing and operating data centers globally to meet the surge in demand for AI applications and other digital services. To ensure the smooth operation of these data centers, both Microsoft and Amazon have invested tens of billions to hundreds of billions of dollars in renewable energy sources.

Given the high costs of developing and maintaining large-scale models, AI companies with limited assets may not withstand the harsh challenges ahead. Particularly in China, without government policy support or the backing of Wall Street funds, it will be challenging for tech companies to sustain themselves through the fierce competition. According to Irene Tunkel, Chief Equity Strategist at BCA Research, despite extensive efforts by tech firms in AI, they are still in a phase of heavy investment rather than profitability. Financial reports indicate that in the first quarter, leading AI company iFlytek recorded a loss of over RMB 300 million, significantly higher than the loss of RMB 57.8953 million in the same period last year.

Therefore, from a cost-benefit perspective, ANBOUND’s viewpoint is further validated: large-scale AI models in China should not continue focusing on large AI systems, as most Chinese AI companies cannot sustain such pursuits. In the future, a few major AI companies in the country will likely monopolize large-scale models, while other enterprises in the industry should focus on developing small models tailored to specific application scenarios based on open-source large models, and this could be termed as "small AI".

Final analysis conclusion:

Currently, the AI landscape in China is witnessing intense, chaotic competition among large-scale models. Hindered by limitations in data availability and computational capabilities, Chinese "big AI" struggles to achieve significant breakthroughs and faces formidable competition from foreign counterparts. Hence, ANBOUND proposes avoiding hardware limitations and leveraging strengths to focus on "small AI". This approach is substantiated by the cost implications of large AI models. The development and maintenance costs of "big AI", encompassing data, training, chip procurement, and cloud service center fees, present challenges for tech firms. With this in mind, China's AI industry is better suited to develop small-scale models tailored for specific applications, termed "small AI", to enhance cost-effectiveness and mitigate unnecessary competition and capital inefficiencies.

______________

Chen Li is an Economic Research Fellow at ANBOUND, an independent think tank.