DeepSeek released their V3.1 model in August 2025, and it got people’s attention for one reason above all else: it runs entirely on Chinese-made chips. No NVIDIA hardware anywhere in the stack. Given that US export restrictions were supposed to slow China’s AI progress, this is a significant result. The key technical innovation is a UE8M0 FP8 precision format that squeezes more performance out of less memory. If you’re following the generative AI landscape, this one matters.
What’s under the hood
The V3.1 is DeepSeek’s third major upgrade this year. The FP8 format lets them train and run large models with significantly less memory than standard approaches. It’s a workaround born from necessity — when you can’t buy the best chips on the market, you find ways to do more with what you have.
Their engineers optimized every layer of the model to work well with domestic silicon. The payoff: training times and costs that are competitive with labs using top-end NVIDIA hardware. That wasn’t supposed to be possible this soon.
The trade restriction backstory
The January 2025 “Foundry Rule” cut off TSMC access for Chinese chip designers. That was supposed to be a major setback. And for some companies, it was. DeepSeek treated it as a constraint to work around rather than a wall to stop at.
Chinese semiconductor manufacturing has been moving fast. SMIC and others are closing the process node gap, and V3.1 proves you don’t need the absolute latest fabrication technology to build a competitive model. It’s a different path, but the destination looks similar.
Actual performance improvements
The FP8 precision format cuts memory usage substantially compared to standard FP16. Training runs faster on domestic hardware than most analysts expected. The model handles complex language tasks — reasoning, code generation, long-context understanding — at levels that benchmark close to GPT-5.
Sam Altman publicly called DeepSeek a competitive threat. When the CEO of OpenAI singles out your lab by name, you’ve clearly done something right.
What this means going forward
V3.1 is a proof point. Trade restrictions can slow hardware access, but they don’t stop software and optimization innovation. If anything, the pressure seems to be accelerating domestic R&D in China rather than stalling it.
The broader implication is practical: hardware diversity in AI is increasing. The assumption that cutting-edge AI requires NVIDIA’s latest silicon is weaker than it was a year ago. For the rest of the industry, that changes the competitive math. For more on how recent AI developments are shifting the field, see our coverage.
4.5