I was following the latest developments in the Chinese artificial intelligence industry and noticed something very important worth discussing. Eight years ago, the story of Zhongxing was a harsh lesson about dependence on global supply chains, but today the picture is completely different.

The real problem was never the chips themselves, but CUDA—the software environment from Nvidia that monopolized over 90% of the global AI market. Nearly every developer in the world trained themselves on this platform, making it almost impossible to detach from it. But what has happened in recent months indicates a real shift.

DeepSeek took a completely different approach. Instead of trying to compete directly with Nvidia, they focused on radically improving algorithms. Their model V3 has 671 billion parameters but only activates 5.5% during inference. The result? Training costs dropped from $78 million for GPT-4 to about $5.6 million. This huge cost difference directly reflected in service prices—DeepSeek is 25 to 75 times cheaper than Claude.

But the most important part relates to the ability to train locally. Tsinghua and others have started deploying fully local computing servers using Loongson processors and Chinese Taichu Yuanqi cards. This is not just inference but actual training of large models. In January 2026, the first advanced model for fully generating images on Chinese chips was trained. Now we are talking about a real qualitative leap.

Huawei, for its part, built a complete ecosystem around Ascend processors. By the end of 2025, the number of developers exceeded 4 million, with more than 200 open-source models adapted. Huawei’s new water chip prices are now approaching NVIDIA A100 levels in performance, and this is a development that cannot be ignored.

Another very important factor is electricity. China produces 2.5 times more electricity than the US, and industrial electricity prices in western China are 4 to 5 times lower than in the US. While the US faces a severe electricity shortage and halts data center projects, China is rapidly building.

What is happening now resembles the trade war between the US and Japan in the 1980s over semiconductors. But this time, China is learning from the lesson—rather than just becoming a product in a global system, it is building an entirely independent ecosystem. From algorithms to chips, to software environment, to energy.

DeepSeek data paints a clearer picture: 30.7% of users are from China, but 13.6% from India and 6.9% from Indonesia. This means the Chinese model is beginning to dominate emerging markets. Within China itself, DeepSeek has captured 89% of the market.

Financial reports from local chip companies released in February 2026 tell a true story. Yes, some lost billions, but revenues increased by hundreds of percentage points. These losses are not management failures but investments in building an independent ecosystem. Every dollar lost is spent on R&D, software support, and solving technical problems.

The real difference between us and Japan in the 1980s is that we are not content to be the best product in a system dominated by external power. We are building the system itself from scratch. This is harder, but the only way to true independence in the age of AI.