I noticed something very important in the story of Chinese artificial intelligence, and I believe most people completely overlook it.

Eight years ago, ZTE was suffering a real disaster – a full American ban that halted its operations overnight. 80,000 employees, revenues exceeding one trillion yuan, and in a single moment, everything collapsed. Now in 2026, we are witnessing a completely different story. DeepSeek announces that it built a full multimodal model without NVIDIA. The difference? This time, China built a truly independent ecosystem.

The real problem wasn’t chips to begin with. Anyone talking about chip bans is misdiagnosing. What really stifles Chinese AI companies is CUDA – NVIDIA’s platform that controls everything. Imagine that 90% of the world's AI developers are tied to this platform. Every line of code, every project, every team learns on it from day one. That’s a very deep trench – you can’t just switch out the “stomach” and move on.

But China chose a different path. Instead of trying to compete directly with NVIDIA, it started rethinking the algorithms themselves. Hybrid expert models – a simple but powerful idea. Instead of running the entire model, only the necessary parts are activated. DeepSeek V3, for example: 671 billion parameters, but only 5.5% are active during inference. The result? Training with just 2048 H800 units for 58 days at $5.6 million. Compare that to GPT-4, which costs about $78 million. The price difference directly reflected – DeepSeek is 25 to 75 times cheaper than Claude.

This price gap changed everything. By February 2026, Chinese model usage on OpenRouter increased 127% in just three weeks. A year ago, the share was 2%, now close to 60%. But what’s happening behind the scenes is even more important – applications shifted from simple conversations to smart agents consuming 100 times more tokens. When token consumption becomes huge, price becomes the decisive factor.

The real leap came from local chips. In Jiangsu, they built an entire production line in just 180 days. Loongson 3C6000 processors and Taichu Yuanqi cards – 100% Chinese chips. The most important? They started handling real training tasks. Zhipu AI trained its first image generation model on Chinese-made local chips. China Telecom trained its massive model on a full local Chinese computing pool. This isn’t just inference – this is real training. The difference between the two is enormous. Inference is relatively easy, training requires ten times the computing power, higher bandwidth, and an advanced software ecosystem.

Huawei Ascend is at the heart of this system. 4 million developers, over 3,000 partners, 43 main models trained on it. In March 2026, Huawei launched SuperPoD – a new computing architecture where Ascend 910B’s processing power reached NVIDIA A100 levels. Not perfect yet, but the gap has shifted from “unusable” to “usable.” Don’t wait for perfection – start deploying now and use market needs to develop chips and software.

There’s another advantage no one talks about much: electricity. The US is facing a real energy crisis. Virginia and Georgia have suspended approvals for new data centers. Power consumption of American data centers could reach 12% of total consumption by 2030. The power grid is already strained. Wholesale electricity costs have risen 267% in regions with data centers.

China, on the other hand, produces 2.5 times more electricity than the US. Residential consumption accounts for only 15% of the total ( compared to 36% in America ), meaning massive industrial energy is available. Industrial electricity prices in western China are about $0.03 per kWh – a quarter to a fifth of American electricity prices. This huge energy advantage completely shifts the equation. Water-intensive and energy-intensive chips become less costly in an environment with abundant electricity.

What’s coming out of China now isn’t products or factories – it’s Tokens. The smallest unit processed by AI models becomes a new digital commodity. Produced in Chinese computing factories and then transmitted via submarine cables worldwide. The distribution of DeepSeek users tells you a lot: China 30.7%, India 13.6%, Indonesia 6.9%, the US 4.3%. 26,000 global companies have accounts. In emerging markets, adoption is massive.

This reminds me of Japan’s semiconductor story in the 1980s. Japan controlled 51% of the global market in 1988. But after the US-Japan semiconductor agreement, everything changed. Intense pressure, support for competitors, and ultimately Japan’s share in DRAM dropped from 80% to 10%. By 2017, Japan’s share in ICs was only 7%. The tragedy is that Japan accepted being the best producer in a global system without building a truly independent ecosystem.

This time, China is taking a completely different route. From extreme algorithm improvements, to a leap in local chips from inference to training, then to 4 million Ascend developers, and finally to global Token proliferation. Every step builds a truly independent industrial system.

The financial reports published on February 27, 2026 tell the real story. Kimo’s revenue soared 453% and turned a profit for the first time. Moi Tun grew 243% but lost a billion. Muxi grew 121% but lost 800 million. Half of it is fire, half is water. The flames are the huge market appetite. The 95% gap left by Jensen Huang is gradually being filled. The oceanic costs are the cost of building the ecosystem – every real money loss in the pursuit of building a CUDA alternative. R&D investments, software support, engineers solving translation problems one after another.

These losses aren’t mismanagement – they are a war tax that must be paid. The war has changed its form. Eight years ago, we asked: can we survive? Today, the question is: what price must be paid to survive? The price itself is progress.