I noticed something truly interesting in recent weeks. As you follow the development of the artificial intelligence industry, you realize that the real war has never been just about chips, but about something much deeper.

Eight years ago, America imposed a simple ban on ZTE. No American components, no software, no technologies. The company nearly collapsed within weeks. But this time, the story is completely different.

The real squeeze wasn’t on the hardware, but on CUDA. This platform from NVIDIA controls everything in the AI world. Every algorithm, every model, almost every global developer is connected to it. Building an alternative system means rewriting decades of accumulated expertise. Who bears this cost?

But Chinese companies chose a different path. Instead of direct confrontation, they went for algorithmic infiltration. DeepSeek V3 is a clear example. A model with 671 billion parameters, but only activates 37 billion during operation. The cost? Just $5.576 million. Compare that to $78 million for GPT-4. The difference is enormous.

The result directly reflected in prices. DeepSeek’s API is 25 to 75 times cheaper than Claude. This price gap changed everything. By February 2026, the use of Chinese models on OpenRouter increased by 127% in just three weeks.

But reducing operational costs doesn’t solve training problems. Here, local chips entered the scene. Loongson and Taichu Yuanqi cards began handling real training tasks. In January 2026, Zhipu AI trained a full image generation model solely on local Chinese chips. This is a qualitative shift from inference to training.

Huawei Ascend now attracts millions of developers. A complete software environment is being built before our eyes. Major companies are doubling their imports of local computing servers this year.

But there’s one factor many are still ignoring: industrial electricity. Here lies the real advantage.

The United States faces a severe electricity crisis. Data centers now consume 4% of total American electricity, expected to double by 2030. States like Virginia and Georgia have suspended approvals for new data centers. Wholesale electricity costs in those areas have risen 267% in five years.

China’s situation is completely opposite. It produces 2.5 times more electricity annually than the US. Only 15% of local consumption is used domestically, compared to 36% in America. This leaves a huge amount of industrial energy available for computing. Industrial electricity prices in western China are about $0.03 per kilowatt-hour, a quarter or a fifth of US prices.

The difference in industrial electricity means a huge economic advantage. When building massive computing centers, fixed costs dominate. China has a structural edge here.

What’s coming out of China now isn’t products or factories, but tokens themselves. Tiny data units processed by AI models. Produced in local computing factories, then transmitted via submarine cables to the world.

DeepSeek now serves 30.7% of China’s local market, but also 13.6% of India, 6.9% of Indonesia, and 4.3% of the United States. 58% of new AI startups incorporate it into their tech stack. In sanctioned countries, market share ranges between 40% and 60%.

This reminds me of another war for industrial independence. In 1986, Japan signed a semiconductor agreement with America. Japan controlled 51% of the global market at that time. But after the agreement, the US exerted comprehensive pressure, supporting Samsung and SK Hynix in Korea. Japan’s share of DRAM collapsed from 80% to 10%. By 2017, only 7% of the IC market remained.

The difference is that Japan was content to be the best producer within a divided global system, without building an independent ecosystem. When the wave receded, nothing was left.

This time, China is taking a different route. From algorithmic improvements, to the leap in local chips from inference to training, to 4 million developers in the Ascend system, and finally, to tokens spreading globally. Every step builds an independent industrial system.

On February 27, 2026, three Chinese chip companies announced their results on the same day. Kimo’s revenue surged 453% and achieved its first annual profit. Moi Tun grew 243% but lost a billion. Moxi grew 121% but lost 800 million.

Half fire, half water. The flame is market appetite for alternatives. The 95% gap left by NVIDIA is gradually being filled. Regardless of current performance, the market needs an alternative option. A rare structural opportunity driven by geopolitical tensions.

Marine energy is the cost of building the ecosystem. Every real money loss in the effort to build an alternative to CUDA. R&D investments, software support, engineers dispatched to solve translation issues one after another. These losses aren’t mismanagement, but a war tax for building true independence.

The war has changed its form. Eight years ago, our question was: should we stay? Today, it’s: how much are we willing to pay to stay? The same cost is progress.