Chinese Autonomy Strategy: When Demand Increases, Alternatives Decrease

When it comes to the global artificial intelligence sector, we face a strange dilemma: the higher the demand for computing power, the fewer reliable sources are available. This is exactly what China is experiencing today—and how it responds is redefining the balance of power in the industry.

The real bottleneck isn’t chips but the software environment

Some may think that chip bans are the main threat. But the reality is much deeper. The true chokehold on Chinese AI companies isn’t the chips themselves, but a software environment called CUDA.

Since 2006, NVIDIA has built an empire around its CUDA platform. This platform, which turns graphics processing power into a supercomputing tool, has become the foundation for nearly all modern AI models. After two decades of development, over 4.5 million developers are connected to CUDA, spread across more than 40,000 global companies.

Here’s the problem: an AI developer can’t simply abandon CUDA and switch to another technology. Every line of code written in CUDA, every software library developed, every bit of expertise accumulated over years—it’s all deeply tied to this environment. Transitioning would require rewriting the collective knowledge of thousands of the world’s smartest minds. Who will bear this cost?

From algorithms to independence: China’s alternative path

Instead of confronting the ban directly, Chinese companies have chosen a completely different route. From late 2024 to 2025, they undertook a comprehensive strategic shift toward mixed expert models—techniques that divide large models into multiple smaller experts, activating only what’s needed for each task.

DeepSeek launched its V3 model with 671 billion parameters, but only runs 37 billion during inference. The result: training costs of just $5.576 million—compared to $78 million for OpenAI’s GPT-4. This isn’t just a technical detail; it’s a leap in efficiency.

This improvement directly impacted prices. DeepSeek’s API costs between $0.028 and $0.28 per million tokens, versus $5 for GPT-4. This—25 to 75 times cheaper—wasn’t just a pricing advantage; it became a strategic weapon.

Within three weeks in February 2026, Chinese model usage on OpenRouter, the largest AI interface distribution platform worldwide, increased by 127%. The share of Chinese models, which was under 2% a year earlier, rose to nearly 60% after a year—a growth of 421%.

Local infrastructure matures: from inference to training

Now, the real transformation is happening. Chinese domestic chips have moved beyond “inference capability” to a more critical stage: “training capability.”

In Changzhou, a 148-meter-long local production line began operation in 2025—going from concept to production in just 180 days. This line produces Loongson 3C6000 processors and T100 AI cards from Taichu Yuanqi—100% Chinese-made chips designed locally.

The result: a full server produced every five minutes. With an investment of 1.1 billion yuan, it’s expected to produce 100,000 units annually.

Most importantly, these chips are already capable of handling real training tasks for large models. In January 2026, Zhipu AI and Huawei jointly launched the GLM-Image model—the first advanced image generation model trained entirely on Chinese local chips. A month later, China Telecom trained its massive “Stars” model entirely on Chinese computing hardware.

This isn’t just a technical breakthrough—it’s a qualitative shift. Training requires processing huge amounts of data, complex calculations, and parameter updates—demands ten times greater than inference.

The backbone of this shift is Huawei’s Ascend chips. By the end of 2025, over 4 million developers were working in the Ascend environment, with 3,000 industry partners. Forty-three major industry models had been trained on Ascend, and more than 200 open-source models had been adapted.

In March 2026, Huawei launched its new SuperPoD computing technology outside China for the first time. The processing power of the Ascend 910B chip reached the level of NVIDIA’s A100. While a gap remains, the fundamental difference has shifted: from “useless” to “usable efficiently.”

Electricity and the new world: when energy becomes a strategic weapon

While attention is on chips and algorithms, something quieter but more impactful is happening: the electrical gap is widening at a staggering pace.

In early 2026, the U.S. faced a severe energy crisis. Virginia suspended approval for new data center projects, followed by Georgia until 2027. The Eastern U.S. grid faces a capacity shortfall of 6 gigawatts. By 2033, the country will face a total energy gap of 175 gigawatts—enough to power 130 million homes.

American data centers consumed 183 terawatt-hours in 2024—about 4% of the national total—and this is expected to double by 2030. The AI sector alone could consume 20-25% of U.S. electricity by then.

Wholesale electricity prices in regions hosting data centers have risen 267% over the past five years.

China’s situation is radically different. China produces 10.4 trillion kWh annually—2.5 times more than the U.S. (4.2 trillion). Most importantly, residential consumption accounts for only 15% of the total, compared to 36% in the U.S. This means vast industrial energy resources are available for computing investments.

Industrial electricity prices in western China are around $0.03 per kWh—one-quarter to one-fifth of the rates in AI-focused regions in the U.S. ($0.12–$0.15).

This isn’t a marginal difference—it’s structural. Moving heavy computing operations from energy-scarce, drought-prone areas to regions with abundant power fundamentally changes economic equations.

Tokens replace products: how China redefines exports

While the U.S. faces an energy crisis, Chinese AI quietly enters global markets. But this time, what’s emerging isn’t a factory or a product—it’s “tokens,” the tiny units processed by AI models.

These tokens are produced in Chinese data centers and transmitted via global networks worldwide. They are a completely new digital commodity—no shipping or customs needed, just internet connectivity.

User distribution data from DeepSeek tells the story: 30.7% in mainland China, 13.6% in India, 6.9% in Indonesia, 4.3% in the U.S., 3.2% in France. Supporting 37 languages, it has gained strong footholds in emerging markets like Brazil.

Twenty-six thousand global companies have opened accounts, and 3,200 institutions have deployed enterprise versions. By 2025, 58% of new AI startups have adopted DeepSeek as part of their tech stack.

In China, it dominates 89% of the market. In sanctioned countries, market share ranges from 40% to 60%.

A historical lesson from Japan: building the system, not just the product

Forty years ago, Japan faced a similar challenge. In 1986, under intense American pressure, Japan’s semiconductor industry signed an agreement with the U.S.—stripping it of technological independence.

By 1988, Japan controlled 51% of the global semiconductor market, while the U.S. held 36.8%. Six of the top ten semiconductor companies worldwide were Japanese: NEC, Toshiba, Hitachi, Fujitsu, and others. Intel lost $173 million that year and nearly went bankrupt.

But after the agreement, everything changed. The U.S. used comprehensive investigations and simultaneously supported Samsung and Hynix to flood the Japanese market with low prices. Japan’s share of DRAM fell from 80% to 10%.

By 2017, Japan’s share of the IC market shrank to 7%. Companies that once dominated disappeared or were acquired, suffering ongoing losses.

The real reason for Japan’s hemorrhage wasn’t technical deficiency but a strategic choice: accepting to be the “best product” in a globally controlled system, rather than building an independent ecosystem.

When the wave receded, Japan realized it had nothing but its existing production lines.

The Chinese path: same challenge, a completely different choice

Today, China faces the same pressures—and more. Three rounds of chip restrictions (2022, 2023, 2024), with ongoing escalation. The walls of the CUDA environment remain high.

But the response is decisive. Instead of seeking to be the “best product” in a system dominated by NVIDIA, China is building an independent ecosystem.

It started with fundamental algorithm improvements. Then, local infrastructure advanced from inference to training. It accumulated 4 million developers in the Ascend environment. Finally, it exported tokens globally to emerging and advanced markets.

Each step is building real independence—something Japan never achieved.

On February 27, 2026, three Chinese AI chip companies published their financial reports on the same day. Kemo increased revenue by 453% and posted its first annual profit. Moitun grew 243% but lost $1 billion. Moxi grew 121% but lost $800 million.

Half fire, half water. Fire is the crazy market appetite. Water is the cost of building the ecosystem.

Every loss is real money paid in the race toward independence—investments in R&D, software support, engineers solving translation problems one after another. This isn’t poor management; it’s the toll of independence.

These three financial reports honestly reflect the front lines of this war for computing power more than any industry report. It’s not an inspiring victory but a fierce battle fought on the front lines, blood flowing.

But the nature of the war has changed. Eight years ago, we asked: “Can we survive?” Today, the real question is: “What price will we pay?”

And that price is progress itself.