DeepSeek is about to release a new model, but progress has been delayed due to deep low-level compute optimization efforts with Huawei. It is reportedly running both training and inference entirely on Huawei chips. If this comes to fruition, the very scenario Hensen Huang fears most will have become reality. A few days ago, in an interview with Dwarkesh Patel, he said: “If, one day, a model of DeepSeek’s caliber is first released on Huawei chips, it would be a terrifying outcome for our country.” That “one day” may be just around the corner. Huang’s opposition to chip export controls is straightforward (though I believe he still wants to profit from China—its largest market outside the U.S.): 1. China already has sufficient compute power. AI training is a parallel computing problem; what one H100 can do, a cluster of 7nm chips can also accomplish. China has abundant 7nm manufacturing capacity and low-cost energy. Anthropic’s Mythos was trained on “fairly ordinary” compute scales—infrastructure that already exists in large quantities in China. This proves China already has sufficient compute to train top-tier models. Huawei’s 2025 revenue is projected at 880.9 billion RMB, with millions of chips shipped. 2. Algorithms determine the upper limit, not compute. China accounts for over 50% of the world’s AI researchers. Look at the world’s leading AI labs—they are dominated by Chinese diaspora talent. DeepSeek was not built by throwing more GPUs at the problem; it succeeded through algorithmic breakthroughs. Compute sets the floor; algorithms set the ceiling. Chinese model training costs are less than one-tenth of those in the U.S., and electricity costs at Chinese data centers are only half those in the U.S. Therefore, current Chinese model API pricing is not being sold at a loss—profit margins may not be as far behind the U.S. as we assume. But if Chinese models match U.S. performance while costing only one-fifth as much, that would be the real nightmare. Just as Chinese-manufactured goods match U.S. quality at less than one-third the cost, the market will inevitably vote with its wallet. 3. The real effect of the ban is helping Huawei build an ecosystem. Unable to buy NVIDIA chips → forced to adopt domestic alternatives → software ecosystems adapt to domestic chips → Huawei chips become increasingly capable → even if the ban is lifted, customers may never return. China represents 40% of the global tech industry; voluntarily abandoning this market isn’t protection—it’s self-harm. Huang’s exact words: “The real nightmare for the U.S. would be if future AI models perform best on someone else’s technology stack.” So what? The moat in the chip industry isn’t just technological leadership—it’s ecosystem lock-in. When customers write code around your architecture, your moat deepens. When they’re forced to switch to another architecture, your moat gets filled in. China has spent three years migrating—and it’s clear that in the near future, it will catch up. If DeepSeek successfully completes end-to-end training and inference on Huawei chips, it means China’s AI ecosystem—from chips to frameworks to models—has for the first time established a fully independent technology stack, entirely separate from NVIDIA. The challenges DeepSeek and Huawei chips have overcome can now be reused by other Chinese open-source models and domestic chips alike—every layer of the ecosystem will follow suit. It may not take three years to catch up to the current U.S. ecosystem. I find this argument compelling. China’s internet industry flourished precisely because of its firewall—it created space to compete with the U.S. Without it, China’s internet would likely be dominated entirely by American products, leaving little room for domestic innovation. China is no longer playing catch-up—it’s branching off. What Huang sees isn’t China chasing—it’s China building something new. And the export ban? It’s precisely what pushed them down this path.