Since early 2023, generative AI, particularly large models, has reshaped the technology landscape globally. Companies involved in computing power, algorithms, data, cybersecurity, cloud computing, and AI have seen significant growth amid this development. However, China's major AI models have yet to create a notable impact on production and lifestyle dynamics. Entrepreneur Kaifu Lee recently pointed out that while "the ChatGPT moment" occurred 17 months ago in the U.S., Chinese users are still anticipating their own AI breakthrough, as current chatbots and tools fail to meet expectations. He emphasized the need for China to develop its own version of ChatGPT to spark public interest and drive broader applications and investments in AI technology.
Industry leaders are now voicing the opinion that "large models without practical applications are meaningless." Baidu's CEO, Robin Li, criticized the proliferation of foundational models in China as a waste of societal resources and called for a focus on practical integration with various industries and the exploration of potential super applications. The market demands AI products and services that deliver quick value, making mere parameter comparisons irrelevant for developers and users of large models. Tan Dai, president of Volcano Engine, posed the critical question: "How can more people across different industries utilize these models?" This sets the parameter for defining a "good model" today.
Experts predict a crucial turning point for generative AI applications in China within this year, with pricing becoming a pivotal factor. OpenAI's recent announcement of its latest model, GPT-4o, highlights this trend, showcasing improved performance alongside a 50% price reduction. OpenAI has slashed prices four times since early 2023, primarily billing clients per 1,000 tokens. For reference, the input cost for GPT-4 dropped from $0.03 to $0.005—an 83% reduction—while output costs fell from $0.06 to $0.015, a 75% decrease. OpenAI anticipates continuing to lower costs by 50-75% annually.
Other companies in China's market are also pursuing significant cost reductions to enhance application adoption. Zhiyu Model recently announced an 80% decrease in entry-level pricing for its GLM-3 Turbo model, reducing costs from 5 to 1 per million tokens. The second-generation MoE model, DeepSeek-V2, released in May, matches the capabilities of GPT-4 and LLaMA 3-70B with costs of 1 and 2 per million tokens, respectively—approximately 1% of GPT-4 Turbo's price. ByteDance's DouBao model, launched in mid-May, has drastically cut AI usage costs, with the main DouBao Pro 32k model priced at 0.008 per 1,000 tokens, reflecting a 99.3% reduction compared to the industry average.
In this environment of widespread price cuts, the value proposition has shifted. For 1, users can acquire 2,400 tokens from GPT, while Chinese large models provide over 8,000 tokens. Users opting for the open-source LLaMA model can access around 30,000 tokens, and an impressive 1 RMB spent on the DouBao model yields 1.25 million tokens—equivalent to processing three volumes of "Romance of the Three Kingdoms."
The ongoing “hundred model battle” raises the question of how to accelerate model implementation. Although large AI models possess enormous potential value, the industry is still in its preliminary exploration phase. As a cost-driven productivity revolution unfolds, the true worth of large models lies in minimizing the marginal cost of creativity through enhanced image and language understanding.
Tan Dai asserts that price reductions significantly drive value creation in the B2B market, where applications remain limited. With OpenAI continuously lowering prices, the shared objective of expanding the market becomes evident. Reducing trial and error costs is essential for industry growth. However, current market sizes for AI large models in China fall short compared to the substantial training costs incurred by market players. The lack of a positive cycle in B2B services, where revenue gaps surpass two orders of magnitude, poses challenges. In response, large model enterprises are making strides to enhance their offerings' accessibility through pricing strategies, recognizing that affordable solutions represent a viable path forward.
Ranking
