One thing I’ve recently noticed that’s worth stopping to consider. Last February, DeepSeek announced that their new model would collaborate with entirely local chip manufacturers, without relying on Nvidia. "We don’t use Nvidia" — a simple phrase but one with huge implications.

Initially, the market was skeptical. Is it commercially feasible to abandon Nvidia, which controls over 90% of the training chip market? But what’s happening here is deeper than just a business decision. It’s a matter of true independence in computing power.

The alarming truth is that what’s choking Chinese companies isn’t the chips themselves, but something called CUDA. This Nvidia platform practically controls everything in the AI world. Every developer, every framework, every project — everyone is intertwined with it. Building an alternative environment means rewriting decades of development, tools, and expertise. Who will bear that cost?

But China has chosen a different path. Instead of direct confrontation, it’s penetrating through algorithms. By the end of 2024 and early 2025, Chinese AI companies collectively shifted toward hybrid expert models. The simple idea: instead of activating a large full model, divide it into smaller experts and only activate the most relevant ones. DeepSeek V3 is a perfect example — 671 billion parameters but only 37 billion are active at a time, just 5.5%. Training cost? Only $5.576 million. GPT-4 cost about $78 million. The difference is enormous.

This difference directly reflected in the price. The API price for DeepSeek is 25 to 75 times cheaper than Claude. The result? In February, the share of Chinese models on OpenRouter increased by 127% in just three weeks, surpassing the US for the first time. A year ago, it was less than 2%. Now it’s approaching 60%.

But here comes the real part. Reducing inference costs doesn’t solve the core problem — training. And that requires enormous computational power.

In Changzhou, a small city known for stainless steel, a local production line was built measuring 148 meters. From signing to production took only 180 days. The essence? Fully local chips: Loongson 3C6000 processor and Taichu Yuanqi acceleration card for industrial use. When fully operational, five units are produced every minute. The really important part is that these chips are already handling massive training tasks.

In January 2026, Zhipu launched, in collaboration with Huawei, the GLM-Image model — the first advanced image generation model trained entirely on Chinese local chips. Immediately after, a huge "Star" model was trained on a local Chinese computing pool.

This is a real paradigm shift. Inference requires relatively low demands. Training? It demands processing vast amounts of data and complex gradient calculations. The requirements increase tenfold in terms of computing power, bandwidth, and software ecosystem.

Huawei Ascend is the real powerhouse here. By the end of 2025, the number of Ascend developers exceeded 4 million. Over 3,000 partners. 43 major industry models trained on its platform. More than 200 open-source models adapted.

At the MWC conference in March 2026, Huawei launched SuperPoD for the first time in foreign markets. The processing power of Ascend 910B reached the level of NVIDIA A100. The gap shifted from unusable to usable.

But there’s another aspect many haven’t talked about: energy. The end of computational power is energy itself. And here, the gap is completely opposite.

China produces 10.4 trillion kilowatt-hours annually. The US produces 4.2 trillion. China produces 2.5 times what America does. More importantly? Residential consumption in China accounts for only 15% of the total, while in the US it’s 36%. This means huge industrial energy that can be directed toward building computing infrastructure.

In terms of costs, electricity prices in American AI hubs range between $0.12 and $0.15 per kilowatt-hour. In western China? About $0.03. a quarter to one-fifth of the American price. And while the US faces real electrical issues — Virginia and Georgia have suspended new data center approvals — Chinese AI quietly expands abroad.

But this time, what’s coming out isn’t a product or factory. It’s a Token — the smallest unit processed by AI models. Produced in Chinese computing factories, then transmitted via submarine cables worldwide. It’s a completely new digital commodity.

DeepSeek user distribution data tell the story: China 30.7%, India 13.6%, Indonesia 6.9%, the US 4.3%. Supports 37 languages. Very popular in emerging markets. 26,000 global companies have accounts. In China, it captured 89% of the market.

This is very much like the industrial independence war from forty years ago. In 1986, Japan signed a semiconductor agreement with the US. Japan’s industry was at its peak — controlling 51% of the global market in 1988. But after the signing? Its share of the DRAM market fell from 80% to 10%. By 2017, only 7% of the IC market remained. Giants withdrew through division, acquisition, or ongoing losses.

The difference is that Japan accepted being the best producer in a global system dominated by one power, but it didn’t build an independent ecosystem. When the wave receded, it realized it had nothing but production itself.

China stands at a similar crossroads but entirely different. Facing enormous external pressures — three rounds of chip restrictions with ongoing escalation — this time it chose a much harder path: maximum algorithmic improvements, local chip leap from inference to training, accumulating 4 million developers in the Ascend system, then spreading Tokens globally. Every step builds an independent industrial system that Japan never had.

On February 27, 2026, three local chip companies published performance reports on the same day. Kimo’s revenue soared 453% and turned a profit for the first time. Moi Tun grew 243% but lost $1 billion. Muxi grew 121% and lost nearly $800 million.

Half fire, half water. Fire is the market’s insatiable appetite. The 95% gap left by Nvidia is gradually being filled. The market needs a real alternative. This is a rare structural opportunity amid geopolitical tensions.

Water is the enormous cost of building the ecosystem. Every real money loss paid in the quest to build a local CUDA. R&D, software support, engineers dispatched to solve translation issues one by one. These losses aren’t mismanagement — they’re a war tax that must be paid.

These three financial reports honestly depict the true picture of this war for computational power. It’s not an inspiring victory, but a fierce battle fought on the front lines, blood being shed.

But the shape of the war has already changed. Eight years ago, we discussed: can we survive? Today, we ask: what price must be paid to survive? The very price is progress.