Summary of Key Points
This article uses Darwin's theory of "adaptive radiation" observed in the finch species on the Galapagos Islands to explain the unique landscape of China's AI industry. Unlike the American AI sector, where a few giants (such as OpenAI) dominate (creating a "champion structure"), China's AI industry features a fragmented landscape due to scarce computing power, widespread use of open-source technology, and diverse monetization strategies. The article suggests that cities should not blindly compete for the title of "AI hub" but instead identify their own "ecological niches"—especially those closely aligned with local industries. In the future, the core metric for measuring the success of the AI industry will be the "morpheme economy," similar to the concept of internet traffic.
I. AI in China and the US: Different Approaches
American AI follows a pattern of "convergent evolution"—organisms like sharks, plesiosaurs, and dolphins, despite their genetic differences, all evolved to have streamlined bodies because there is only one optimal design for high-speed movement in water. Companies like OpenAI, Anthropic, and DeepMind all focus on creating chat interfaces and APIs, using a subscription-based business model with pay-per-use pricing, and their technological approach is to build larger models and more computing power. The winner takes all, leading to a monopoly by a few giants.
Chinese AI, on the other hand, follows a pattern of "adaptive radiation"—similar to how finches have evolved different beak shapes for various tasks (e.g., cracking seeds, extracting nectar, or using tools to poke insects). Each company focuses on its own niche. For example, ByteDance invests in models with 10 trillion parameters, YueZhiMian develops open-source large models, DeepSeek focuses on low-cost inference, and Alibaba emphasizes cost-effective multimodal models. It's like comparing a screwdriver with a wrench; they serve different purposes.
II. Why Has Chinese AI Diversified?
The differentiation in China is not the result of conscious choice but rather three constraints:
1. Scarcity of Computing Power: With limited computing resources, companies have to find alternative approaches. While the US has ample computing power, China's shortage (7-8 times less) forces them to explore non-computing-based strategies. For instance, DeepSeek focuses on reducing costs significantly (inference prices at 2% of others' levels), and YueZhiMian uses small amounts of context to process long texts.
2. Widespread Use of Open-Source: Although models can be copied, core capabilities are difficult to replicate. Chinese models like Qwen and Kimi are open-source, allowing others to quickly learn from them. Therefore, companies focus on non-copyable aspects such as cost-effective architectures (DeepSeek) or developer-friendly ecosystems (Alibaba's use of Qwen for fine-tuning).
3. Diverse Monetization Strategies: Revenue sources vary greatly across countries. The US relies on consumer subscriptions (ChatGPT Plus) and corporate APIs, while Chinese consumers are less willing to pay (as shown by the backlash against subscription services like DouBao). As a result, companies in China use different monetization methods: Alibaba uses cloud services for models, ByteDance leverages e-commerce and advertising, and Zhipu targets government and enterprise clients.
III. Cities Should Stop Competing for AI Hubs and Focus on Ecological Niches
Many city plans aim to become AI hubs, but with 50 such hubs, it would simply be a market saturation. Ecological niches are limited and stratified. Most cities lack the necessary resources to compete at the "model layer":
- Model Layer: Top-tier talent is concentrated in specific areas. For example, Haidian in Beijing hosts half of China's leading AI institutions. Talent mobility depends on whether there are like-minded partners nearby, not on subsidies.
- Computing Power Layer: There is an oversupply of computing power, which is becoming increasingly homogeneous due to similar requirements (electricity, water, green energy).
- Application Layer: This is the most viable opportunity for cities, as it can be closely tied to local industries. For instance, manufacturing cities can develop industry-specific applications, medical cities can use data for research, and cultural and tourism cities can create content generation tools.
IV. The Future of AI: Measured by the Morpheme Economy
The National Data Administration has defined "tokens" (the smallest units in AI processing) as the "settlement units of the intelligent era," similar to internet traffic. This means that the success of AI industries will be measured by how many morphemes are consumed and in what contexts. Cities should shift their focus from attracting large model companies to creating high-quality datasets and open ecosystems. For example, manufacturing cities can share factory data, and cultural and tourism cities can provide content generation materials.
V. The Right Question: How Much Morpheme Can You Consume?
The diversity of Chinese AI is not a weakness but a result of adapting to local conditions. Cities should ask themselves: "How much morpheme can my city consume annually, and in what contexts?" This is more relevant than competing for AI hubs. Just as Darwin did not evaluate the strongest finches but studied their beak shapes, cities should consider how their unique features (i.e., their ecological niches) will help them thrive in the current environment.