The first two articles talked about the ebb of open source, anti distillation iron curtain, and the pain of neck 2.0. Some friends may think that with this, will Chinese AI cool down? I think it's the opposite. Let me show you a set of data. In the first week of April 2026, the weekly usage of Chinese large models reached 12.96 trillion tokens, a month on month increase of 31.48%. What about the United States? 3.03 trillion yuan, a month on month increase of 0.76%. China is 4.27 times that of the United States. Even more exaggeratedly, the top six models in terms of global call volume are all from China. Alibaba Qwen 3.6 Plus ranks first with 4.6 trillion tokens. The model lags behind by 6 months and leads in call volume by 4 times. You appreciate how magical this scene is. Americans are desperately pushing for the upper limit of intelligence in the laboratory, while Chinese people are desperately cramming AI into every usable scenario outside. The two groups of people are not doing the same thing at all. Why can adjusting the dosage crush? The answer is simple, cheap. The token pricing of domestic top models is one sixth to one sixteenth of that of similar models in the United States. MiniMax's million token input cost is $0.3, while Claude Opus is priced at $5. 16 times the price difference. DeepSeek V3.2 costs only 0.28 yuan per million tokens, and looking at OpenAI's price list, the difference starts at more than ten times. This is not just a false prosperity subsidized by burning money. Behind it are real efficiency advantages, such as lower electricity costs, engineering optimization of computing power clusters, inference efficiency of MoE architecture, and a complete local supply chain. After the price war, Chinese big models have instead accumulated cost barriers. Ability difference of 3 to 6 months, price difference of 10 to 16 times. You are an application developer, who do you choose? What if the domestic model is equivalent to the current Claude Opus 4.6 in three months For the vast majority of application scenarios worldwide, the model does not need to be the smartest, as long as it is sufficient. But the price must be cheap enough, otherwise the business model won't work at all. This is the key, only when the inference cost is sufficiently low, can the large-scale outbreak of multimodal applications and AI agents have commercial soil. The high pricing of top models in the United States determines that they can only serve high-end enterprise customers and cannot penetrate into massive small and medium-sized scenarios. And China's ultimate cost-effectiveness has completely broken the commercialization threshold for AI applications. Returning to China, the explosion on the application side is even more outrageous. The most magical story of this year is probably OpenClaw. An open-source AI Agent framework developed by an Austrian developer, featuring a red lobster logo and 315K stars on GitHub, has surpassed Linux to become the world's top open-source project. As a result, in China, it has become a nationwide movement. Baidu Science Park has launched a "Lobster Market", Tencent Building in Shenzhen has set up a free installation point, and Taobao has launched a "door-to-door installation lobster" service, ranging from 30 yuan to 5000 yuan. OpenClaw is a geek tool in the United States. It is a consumer frenzy in China. Of course, the trend also fades away quickly. The buyer show dilemma of security vulnerabilities, high token consumption, and landing effects has led the first batch of shrimp farmers to start paying people to uninstall. But this is precisely the necessary process for the industry to mature, from fanaticism to calmness, from concept to implementation. Compared to the excitement on the consumer side, the actions on the industrial side are actually more worthy of attention. In the 2026 Two Sessions, "Creating a New Form of Intelligent Economy" was first included in the government work report. This is not shouting slogans, it's really pushing. The AI workshop of Zhongce Rubber has produced one tire on average every 3 seconds, increasing production efficiency by 300% and reducing the defect rate to 0.5%. BOE's AI factory covers six major scenarios including production planning, material supply, and quality management, with full chain intelligent manufacturing. According to IDC data, the penetration rate of intelligent agents in Chinese industrial enterprises has jumped from 9.6% in 2024 to 47.5% in 2025. The goal of the Ministry of Industry and Information Technology is to promote 500 typical application scenarios of "AI+manufacturing" by 2027. These are not concepts on the PPT, they are things running on the production line. I sometimes feel that the competition between China and the United States in AI may not ultimately be a question of 'whose model is smarter', but rather a question of 'whose model is more useful'. No matter how strong the model is, if only Silicon Valley tech companies can afford it, its value is limited. The model is slightly inferior, but small and medium-sized enterprises, factories, and individual developers around the world are using it, and its influence is actually greater. This reminds me of a historical story. In the 1880s, electricity began to become popular in the United States, and there was a dispute over two routes: Edison's direct current and Tesla's alternating current. Direct current is superior in many technical indicators, but alternating current has an overwhelming advantage of being cheap and capable of long-distance transmission. The ultimate winner is AC power. Not because it's better, but because it allows more people to access electricity. The path that Chinese AI is currently taking may be the 'alternating current' path. Models may not necessarily be the strongest, but tokens are the cheapest, have the most applications, have the richest scenarios, and have the deepest penetration. The United States builds bunkers on mountaintops, while China builds roads to every village at the foot of the mountains. Of course, I'm not saying that models are not important. Basic research is always important, and the gap always needs to be caught up. Reverse distillation+closed source may indeed widen the generation gap in basic innovation, and short-term pains are inevitable. But China's accumulation in the application layer and cost side is giving rise to another AI industry paradigm. Our core competitiveness lies not in making the model the smartest, but in making the model the most useful and cost-effective. The two paths are not either or. The United States won the model, and China won the market. And the most tragic ones are those countries that have no two roads. Great era, my friends.