DeepSeek is about to release a new model. The delay is because it has been doing deep-level underlying compute-ability adaptation with Huawei, which has held things up quite a bit. It’s said that all inference and training will run entirely on Huawei chips.

If this happens, then the worst-case scenario that Huang Renxun fears will come true.
A few days ago, in an interview with Dwarkesh Patel, he said: “If someday in the future a model at the DeepSeek level is released first on Huawei chips, it would be a terrifying outcome for our country.”
Now, that “someday” might be arriving soon.
Huang Renxun’s logic for opposing chip export controls is very simple—of course, I also think he wants to make money from China; it’s the biggest market outside the US.
1、China’s computing power has long been enough. AI training is a parallel-computing problem. Whatever a single H100 can do, you can also do by stacking a bunch of 7nm chips. China has a large amount of 7nm capacity and cheap energy. Anthropic’s Mythos was trained and produced on a “quite ordinary computing scale.” This kind of compute already exists in large quantities in China, which shows that China has enough compute to train the very top models. Huawei’s 2025 revenue is 8809 billion yuan, and it has shipped tens of millions of chips.
2、Algorithms matter more than computing power in determining the ceiling. China has more than 50% of the world’s AI researchers. Just look at the world’s top AI labs—you’ll see that the number of people of Chinese descent there is in the majority. DeepSeek isn’t built by stacking cards; it’s a breakthrough at the algorithm level. Computing power is the lower limit, while algorithms are the upper limit.
The training cost of Chinese models is less than one-tenth that of the US. China’s data center electricity bills for compute centers are only half of the US’s. So the current API pricing for Chinese models is not “selling at a loss”; the profit margin compared with the US may not be as large as we think. But if the model’s capability is the same as the US’s, while the cost is only one-fifth, then that would be the biggest nightmare. It’s like Chinese-made products with the same quality as American ones, but with costs less than one-third of America’s—will users still vote with their wallets at the end? The answer is obvious.
3、The real effect of the ban is to help Huawei nurture and build its ecosystem. Can’t buy Nvidia → So only domestic chips can be used → The software ecosystem adapts around domestic chips → Huawei chips become increasingly easier to use → Even if the ban is lifted, customers may not come back. China accounts for 40% of the global technology industry; voluntarily giving up this market isn’t protection—it’s self-harm.
Huang Renxun’s exact words: “If, in the future, AI models run best on someone else’s technology stack, then that’s America’s nightmare.”
So what?
The moat in the chip industry isn’t just technical leadership—it’s also ecosystem lock-in. When customers write code around your architecture, your moat keeps getting deeper. If customers are forced to move to someone else’s architecture, your moat will be leveled. China spent 3 years migrating, and it’s foreseeable that this will be achieved in the not-too-distant future.
If DeepSeek really completes full end-to-end validation—from training to inference—on Huawei chips, it means that for the first time, China’s AI ecosystem, from chips to frameworks to models, has run a completely independent technology stack.
The pitfalls that DeepSeek and Huawei chips have stepped into can be reused for other Chinese open-source models, and with domestically produced chips, the corresponding entire ecosystem will move along together. It might not even require 3 years to catch up to the US’s current ecosystem.
I think this part is quite reasonable. China’s internet development has happened precisely because the Great Firewall exists—only then could it develop differently from the US. If there were no firewall, China would probably already be entirely full of American internet products, and maybe the domestic internet would not have that much room to grow.
China today isn’t catching up—it’s branching off.
What Huang Renxun sees isn’t China chasing; it’s China setting up a new workshop. And the ban is exactly the push that forces them onto this road.