中国AI调用量首超美国,国产算力电力大涨,英伟达市值一夜 (English)

Generated: 2026-06-24 04:57:20

---

Guess what? It’s not what you think!

Speaking of which, I still remember exactly how I felt scrolling through my feed that day.

It was an ordinary weekday evening in March 2026. I was about to put my phone down and call it a night when a post jolted me upright—China’s AI model token call volume had surpassed America’s for the first time.

My social feed exploded. Some people cheered, “China AI has fully overtaken the US!” Others were already calculating “power equipment ten-bagger stocks,” and still others asked, “Is it too late to jump in now?”

I stared at the screen for a long moment, then smiled.

All three of those things did happen—but if you look at them together, the story is just getting started.

---

First, those 140 billion tokens: the number is real, but you have to know how to read it

According to OpenRouter, the world’s largest AI model API aggregation platform, in the last week of February: Chinese models processed 4.12 trillion tokens, while US models processed 2.94 trillion. By March, the average daily token call volume for Chinese models on domestic platforms had also exceeded 140 billion (roughly 0.14 trillion)—a 1,000-fold increase in two years, from just 140 million per day.

Sounds explosive. But let’s break it down.

1. Who’s using them?

On OpenRouter, US users account for 47% of traffic; Chinese users only 6%. Developers from overseas are voting with their feet and choosing cheaper Chinese models. This isn’t domestic hype—it’s foreign developers making that choice themselves.

2. Which models are carrying the load?

MiniMax M2.5, Kimi K2.5, Zhipu GLM-5, DeepSeek V3.2—these four hold four of the top five spots globally for Chinese model call volume. To be honest, I’ve used models from both camps, and on complex reasoning, code generation, and long-document logical consistency, GPT-4o still feels ahead.

3. Most important point: token call volume does not equal technical capability.

Last year, I ran tests in my own project, Huanxing AI. For the same task, GPT-4o would nail it in one go, while some domestic models needed two prompt tweaks and an extra round. Higher call volume doesn’t always mean higher efficiency. Token count tells you “how much was used,” not “how well it was done.”

**But the 140 billion daily figure does reveal something else—China is truly putting AI to work. It’s no longer a lab