I just noticed this year how much the landscape of Chinese AI has changed. When the chip embargo started, everyone thought the dream was over. But do you know what’s really carrying all of us? It’s not just about hardware— the real problem is CUDA.

If you haven’t heard yet, CUDA is the ecosystem NVIDIA built over more than ten years. All major AI frameworks, from Google TensorFlow to Meta PyTorch, rely on it. Now, there are 4.5 million developers using it, and 90% of AI developers worldwide are already locked into this system. It’s a flywheel that’s almost impossible to break—more people use it, more tools get built, and it becomes harder to get out.

But this time, China didn’t give up. The strategy is smarter—go after the algorithms. DeepSeek V3 has 671 billion parameters, but it only uses 5.5% of them per inference. The training cost is only $5.576 million, while GPT-4 costs nearly $78 million. Do you know what happened? DeepSeek’s API price is $0.028 per million tokens, while GPT-4o is $5. That’s 25 to 75 times cheaper. This isn’t just a discount—it's a structural shift in the industry.

And now for the amazing part: local chips have started training large models. In January 2026, Zhipu AI released GLM-Image together with Huawei, the first image generation model fully trained on local chips. The Loongson 3C6000 processor and the TaiChu Yuanqi T100 accelerator card aren’t just for inference anymore—they’re capable of training. In Jiangsu Xinghua, there’s a production line 148 meters long that produces servers every 5 minutes. This is what production looks like in the new era—not just physical goods, but computational capacity that can be exported worldwide.

The Huawei Ascend ecosystem has grown to 4 million developers, 3,000 partners, and 43 major models that have been pre-trained here. That number keeps increasing. By 2026, China’s intelligent computing capacity will reach 1590 EFLOPS. This isn’t wishful thinking anymore—it’s reality.

Now, the energy situation. America is in trouble. Virginia, Georgia, Illinois, Michigan—everyone has paused new data center projects due to power shortages. By 2033, the US will face a 175GW capacity deficit. But China? Its annual electricity generation is 10.4 trillion units—2.5 times that of the US. And residential usage in China is only 15% of the total, while in the US it’s 36%. This means there’s more industrial capacity available for computing infrastructure. Electricity prices in western China are $0.03 per kilowatt-hour, while in the US they range from $0.12 to $0.15. That’s a 4 to 5 times difference.

So tokens—the smallest unit of AI information—have started being produced in China and exported worldwide. DeepSeek’s user distribution is 30.7% China, 13.6% India, 6.9% Indonesia, 4.3% US, 3.2% France. There are 26,000 companies worldwide, and 3,200 institutions in the enterprise version. In 2025, 58% of new AI startups will integrate DeepSeek into their tech stack. In China itself, it has a 89% market share.

What happened here is like what happened in Japan 40 years ago. In 1988, Japan controlled 51% of the global semiconductor market, but after the US-Japan Semiconductor Agreement, their share fell to just 10% in DRAM. Their mistake was relying only on being the best manufacturer—without an independent ecosystem. Now China has chosen a different path—from extreme algorithm optimization, to developing local chips from inference to training, to gathering millions of developers in the Ascend ecosystem, and exporting tokens globally.

The March 27, 2026 reports show an interesting picture. Cambrian revenue is up 453%, and it has its first full-year profit. Moore Threads revenue is up 243% but with a 1 billion loss. Muxi revenue is up 121% but with an 8 billion loss. Half fire, half water. But the point is clear—the market needs an alternative where NVIDIA isn’t dominant. This is a structural opportunity driven by geopolitics.

Building an ecosystem comes at a cost. Every loss is real money spent on learning, software subsidies, and deploying engineers to clients. But these losses aren’t due to bad execution—this is a war tax for independence. Eight years ago, the question was “Can we survive?” Now, the question is “What cost do we have to pay to survive?” The cost itself is progress.