Nvidia and DeepSeek: How Engineer 'Complicity' Bypasses US Sanctions

Remember that shock when Chinese DeepSeek rolled out its V3 and R1 models, which in terms of intelligence were no worse than GPT-4o, but cost tens of times less to train? Back then everyone wondered: how did they manage it under the most severe shortage of top Nvidia chips? It seems the answer has been found, and American regulators won't like it. Head of the U.S. House Committee on China John Moulton directly stated that Nvidia not only sold "authorized" hardware, but actively helped DeepSeek in designing and optimizing their architecture. This is exactly the case when business finds a loophole in the fence that politicians have been building for years.

The situation looks ironic. While Washington methodically banned supplies of H100 and B200 to China, Nvidia engineers, it appears, worked "in the field," helping Chinese colleagues squeeze the maximum out of what they had on hand. The term "co-design" in this context means deep software optimization for CUDA architecture. This is critically important: you can have a mountain of graphics cards, but without the right software they will remain just expensive heaters. DeepSeek showed the world that the Mixture-of-Experts (MoE) architecture and intelligent load distribution can compensate for the lack of computing power. And now it turns out that behind this triumph of Chinese engineering could have been technical expertise from Silicon Valley.

Why did Nvidia take such a risk? The answer is simple: money and influence. China remains a huge market, and if a company cannot sell its best chips there, it tries to sell its ecosystem. By getting Chinese developers hooked on CUDA through consulting and collaborative work, Nvidia guarantees itself dominance for decades to come, even if the hardware itself is weaker than American counterparts. This is a classic "soft power" strategy that this time collided with the harsh reality of geopolitics. For Jensen Huang it's a game on a knife's edge — he's trying to maintain global monopoly without angering the White House.

The consequences of this scandal could be tectonic. If before sanctions concerned only physical goods — teraflops and memory bandwidth — now regulators can tackle "export of brains." We are entering an era when an American engineer could be banned from calling a Chinese colleague to discuss code. This sounds like a dystopia, but in the logic of trade war it's the next logical step. DeepSeek has already proven that algorithm efficiency is more important than the number of transistors, which means the battle for minds is becoming more important than the battle for TSMC factories.

The most amusing thing about this story is that DeepSeek's success ultimately hurt Nvidia itself. When the market realized that creating powerful AI doesn't necessarily require buying tens of thousands of the latest GPUs, Nvidia's stock wavered. It turns out that by helping DeepSeek, the company with its own hands created a case that questions the need for endless purchases of its most expensive hardware. This is the highest degree of irony in the world of high technology: help a competitor prove that your main product can be used much more economically than you used to tell investors.

Currently Nvidia is under the microscope. Any confirmation that technical support went beyond standard instructions will lead to new Congressional hearings. For the industry this is an important signal: the era of free exchange of ideas in the field of AI is finally closing. Now every commit in GitHub and every line in documentation can be viewed as a matter of national security. We are witnessing how a "digital iron curtain" is being erected around artificial intelligence, and Nvidia seems to have found itself on both sides of it at once.

The main point: Washington realized that banning chips is useless if you don't ban engineers from talking to each other. Are we waiting for the emergence of "consulting sanctions"?