I noticed something interesting as I follow the developments of Chinese artificial intelligence these days. Exactly eight years ago, a dramatic turning point occurred when ZTE's heart stopped in April 2018. A single American ban brought everything down—no Qualcomm chips, no operating system. The company that was the fourth-largest global telecom equipment provider shut down overnight.

But here comes the most important part: now, after eight years, we are witnessing a very different ending. China did not succumb to the same AI scenario. Instead, it chose a much harder path.

The real problem was not always chips—it was CUDA. This platform from NVIDIA has become the backbone of the entire global AI industry. More than 4.5 million developers are connected to it, and over 90% of AI developers work within this ecosystem. It’s a sustainable wheel— the more developers use it, the more the environment thrives, attracting additional developers.

China’s solution came from an unexpected place: algorithms. Instead of trying to compete directly with NVIDIA, Chinese companies focused on radically improving algorithms. Hybrid expert models became the trend—breaking down large models into smaller experts and activating only what is necessary. DeepSeek V3 has 671 billion parameters but activates only 37 billion—just 5.5%. The cost? $5.576 million compared to $78 million for GPT-4. That’s a huge difference.

The price was directly reflected in APIs. DeepSeek offers prices 25 to 75 times cheaper than Claude. In February, the use of Chinese models on OpenRouter increased by 127% in just three weeks, surpassing the United States for the first time.

But price and algorithms alone do not solve the training problem. This is where local chips come in. Loongson 3C6000 and Taichu Yuanqi cards have started working seriously. In January 2026, Zhipu launched the first fully Chinese-chip-trained image model. Afterwards, the "Stars" model was trained on a local computing pool with tens of thousands of units. This is a qualitative shift—from "inference capability" to "training capability."

Huawei’s Ascend has become the main engine here. 4 million developers, 3,000 partners, and 43 main models trained on its basis. Ascend 910B has reached the level of NVIDIA A100—from unusable to usable. Perfection cannot be waited for; development must start now and use real business needs to drive progress.

Another often-overlooked factor: electricity. China produces 2.5 times more electricity than the US, and residential consumption here is only 15% compared to 36% there. Industrial electricity prices in western China are about $0.03 compared to $0.12–$0.15 in the US—a quarter to a fifth of the price. Meanwhile, the US faces a real electricity crisis—(Virginia and Georgia have suspended new data center approvals), while China exports tokens globally.

DeepSeek is now available in 37 languages. 30.7% of its users are local, but 13.6% are from India, 6.9% from Indonesia. 26,000 global companies have accounts. In countries under sanctions, market share ranges between 40-60%. This resembles an industrial independence war—similar to what happened with Japan and semiconductors 40 years ago, but with a different ending this time.

Earnings reports from February 27 were very honest: some companies achieved profits for the first time, while others lost billions. But these losses are not failures—they are a tax for building a truly independent ecosystem. Every loss is an investment in R&D, in supporting software, in engineers solving translation problems one after another.

The question has shifted from "Can we survive?" to "How much do we need to pay to survive?" And the answer itself is real progress.