I was following the story of ZTE eight years ago when the disaster happened – a complete American ban that stopped it immediately. A company with 80,000 employees and revenues exceeding one trillion yuan disappeared from the map in one day. The scene was heartbreaking: without Qualcomm chips, without Android licenses. The story ended with a $1.4 billion settlement fee and annual losses reaching 7 billion yuan.

But what happened in February 2026 is completely different. DeepSeek announced a V4 multimodal model that will rely entirely on local chips, without NVIDIA. The initial market reaction was skepticism, but behind this announcement lies a deeper issue: how did China build true independence in computational power?

The truth is, what stifles Chinese AI companies isn’t just chips. It’s CUDA – NVIDIA’s platform that has monopolized the ecosystem since 2006. Over 4.5 million developers worldwide are connected to it, and every line of code they write deepens the trench. When the US imposed bans on A100, then H100, then H20, Chinese companies realized that the real problem isn’t the chip, but the entire software environment.

But there is a way out. From late 2024 to 2025, Chinese companies shifted to a completely different strategy: hybrid expert models. The simple idea – instead of activating the entire model, it is divided into smaller experts, activating only the most relevant ones for the task. DeepSeek V3 is a clear example: 671 billion transactions but only 37 billion are active, just 5.5%. Training cost? $5.576 million compared to $78 million for GPT-4. The difference is huge.

This evolution in algorithms directly impacted prices. DeepSeek’s interface costs $0.028 per million tokens versus $5 for GPT-4o. Licensing fees are 25 to 75 times cheaper. In February 2026, the use of Chinese models on OpenRouter increased by 127% in just three weeks, surpassing the US for the first time.

But reducing inference costs doesn’t solve the training problem – that black hole of computational power. Where do the “shovels” for trained models come from?

The answer lies in a small city in southeastern China. In 2025, a full computing production line was built in just 180 days. Loongson 3C6000 processors and T100 cards from Taichu Yuanqi – both fully local. Production: one server every five minutes. More importantly? These chips have started training large, real models. Zhipu AI trained the GLM-Image entirely on Chinese chips. Telecom companies trained their massive models on a local computing pool with tens of thousands of processing units.

This isn’t inference – it’s training. A complete qualitative shift.

Behind this stands Huawei Ascend. By the end of 2025, the number of developers exceeded 4 million, with over 3,000 partners. 43 main models were trained on Ascend. In March 2026, Huawei launched SuperPoD in international markets. The processing power of Ascend 910B reached the level of NVIDIA’s A100. The gap shifted from unusable to usable.

While the US faces a real electricity crisis – Virginia, Georgia, and Illinois suspended new data center projects. Power consumption could reach 12% of US electricity by 2030, and the grid is already strained.

China, on the other hand, produces 2.5 times more electricity than the US, and industrial electricity prices in the West are only $0.03 – a quarter of America’s price. This means enormous energy capacity that can be directed toward computing.

Now, Chinese AI is quietly going global – but not the product or factory. What’s exported is the (Token). The small information unit processed by models has become a new digital commodity. Produced in computing factories, transported via submarine cables worldwide.

The distribution of DeepSeek users says a lot: China 30.7%, India 13.6%, Indonesia 6.9%, the US 4.3%. Supports 37 languages. In 2025, 58% of AI startups integrated DeepSeek into their tech stack. In China, 89% of the market was monopolized.

This reminds me of another war forty years ago. Japan in 1986 signed a semiconductor agreement under US pressure. It was at its peak – controlling 51% of the global market by 1988. But after the agreement, everything changed. Now its share has dropped to just 7%. The giants withdrew one after another.

The Japanese tragedy is that they accepted being the best product in a global division system but didn’t build an independent system. When the wave receded, they had only production left.

Today, China stands at a similar but different crossroads. We face enormous pressures – three rising rounds of chip bans. But this time, we chose a harder path: maximum algorithmic improvements, a leap in local chips from inference to training, the accumulation of millions of developers in the Ascend system, then a global spread of Tokens in emerging markets.

Each step builds an industrial system that Japan never possessed.

On February 27, 2026, three local chip companies published their results on the same day. Half of them are fire, half are water. Fire: revenues grew by 453%, 243%, and 121%. Water: losses reached 910B dollars. Every loss is real money in the effort to build an independent ecosystem – R&D, software support, field engineers solving problems one after another.

These are not management failures. They are a war tax that must be paid.

Eight years ago, we asked: can we survive? Today, the question is: how much must we pay to survive? The same price is progress.