A friend in China asked to buy AI compute from us, and I was baffled
A friend in China recently reached out to purchase AI compute from us, and I was baffled.
Don’t they have an endless supply of powerful, inexpensive AI models?
Then it occurred to me.
They have the models. They don’t have enough chips.
U.S. export controls have limited China’s access to the advanced chips needed to build and run AI. China can produce impressive models, but getting enough compute to run them at scale is another matter.
Take GLM. It’s a Chinese model that anyone can download, but it’s also available through OpenRouter, a service that connects users with companies around the world that have the compute to run it.
So if the chips can’t come to China, the work can go to the chips.
A developer in China sends a request abroad. Foreign computers process it. The answer comes back over the internet.
I don’t think this was some mastermind strategy from the beginning. It’s more likely an emergent behavior.
China released open models to gain adoption and compete globally. Chip restrictions created a shortage of compute at home. Put those together, and foreign infrastructure becomes a practical way for Chinese developers to run Chinese models.
It reminds me of how China has been moving manufacturing to Mexico to work around tariffs.
As a wise man once said: “Life finds a way.” 🌶️