I was reading about the ZTE story eight years ago, and honestly, the whole thing still sticks in my memory. A giant company with 80,000 employees and revenues exceeding one trillion yuan stopped operating overnight. But today, the situation is completely different.

The real problem was never the chips themselves. It was something called CUDA. This NVIDIA platform built an entire ecosystem around it—4.5 million developers, 3,000 accelerated applications, and more than 90% of the world’s AI developers tied to it. The moat is extremely deep. Every new developer who learns on it widens the gap even further.

But I noticed a real need that emerged. When agent scenarios started replacing simple conversations, token consumption jumped from 10 to 100 times. At this moment, price became a decisive factor—not a luxury. DeepSeek managed to cut inference costs to an insane degree—25 to 75 times cheaper than Claude. The result? In just three weeks, the share of Chinese models on OpenRouter rose by 127%, surpassing the United States for the first time.

But that was only part of the story. The real part was in training. Algorithms alone are not enough—you need local chips capable of real training, not just inference. The production line in Jiangsu took only 180 days from signing to production. Loongson processors and Taichu smart cards have actually started handling real training workloads. In January 2026, Zhipu AI released the first image model that is fully trained on domestically made Chinese chips. This is a qualitative shift from inference to actual training.

Another thing I noticed—America’s power shortage is starting to play a real role. Virginia and Georgia stopped approving new data center locations. By 2030, US electricity consumption will double, and it could reach 12% of total consumption. As for China? China produces 2.5 times as much electricity as the United States, and industrial electricity costs are 4 to 5 times lower. This is not a small factor.

Now what comes out of China is not a product or a factory, but the Token itself—the fundamental unit that AI models process. It is produced locally and then transmitted worldwide via submarine cables. The user distribution for DeepSeek tells a lot: China 30.7%, India 13.6%, Indonesia 6.9%, and other countries. In sanctioned countries, market share ranges between 40% and 60%.

This reminds me of the war for industrial independence from forty years ago. In the 1980s, Japan controlled 51% of the global semiconductor market, but it agreed to be the best producer inside a globally partitioned system dominated by one power. It didn’t build an independent ecosystem. When the wave receded, all it had left was production itself.

China is standing at a completely different point. Yes, we’re facing enormous pressure and three rounds of escalating restrictions. But this time, we chose the harder path: extreme algorithmic improvements, a leap in domestically made chips from inference to training, 4 million developers in the Ascend ecosystem, and then global expansion. Every step builds an independent industrial system.

The financial reports released on February 27 tell the real story. Local chip companies achieved massive growth—some are winners, and some are net losers. But these losses are not the result of mismanagement; they are the unavoidable toll of war that must be paid to build an independent ecosystem. Every loss is an investment in R&D, in software support, and in human costs to solve translation problems one by one.

The face of the war has truly changed. Eight years ago, we asked, “Can we survive?” Today, the question is, “What price do we have to pay to survive?” And the price itself is progress.