3DNews AI→ original

DeepSeek trained a new model on banned Nvidia Blackwell accelerators

Chinese company DeepSeek allegedly trained its latest AI model on Nvidia Blackwell accelerators, which remain under an export ban for China. Despite a…

AI-processed from 3DNews AI; edited by Hamidun News
DeepSeek trained a new model on banned Nvidia Blackwell accelerators
Source: 3DNews AI. Collage: Hamidun News.
◐ Listen to article

The American export control regime, built to slow the development of Chinese artificial intelligence, has encountered what may be the most striking failure in its entire history. According to sources cited by American analysts, Chinese company DeepSeek used Nvidia Blackwell accelerators — among the most powerful and strictly forbidden-from-export-to-China chips — to train its latest AI model.

To understand the scale of this story, one must return to the logic of American sanctions. Since 2022, Washington has consistently tightened restrictions on supplies of advanced computing accelerators to China. The idea was simple: without access to the most performant chips, Chinese companies would not be able to train competitive large language models and other AI systems.

Nvidia, dominant in the market for machine learning accelerators, found itself at the center of this strategy. The company was forced to create special "crippled" versions of its chips for the Chinese market, while its flagship products, including the entire Blackwell family, remained under strict prohibition. Even though in recent times the US administration has somewhat softened its overall rhetoric and partially relaxed restrictions on less powerful solutions, Blackwell remains in the category of the strictest embargo.

DeepSeek has repeatedly demonstrated the ability to achieve impressive results under conditions of limited access to computing resources. The company's models, in particular DeepSeek-V3 and DeepSeek-R1, produced a bombshell effect in the industry: they showed performance comparable to Western counterparts while requiring significantly lower training costs. At that time, the company claimed it was working on older Nvidia A100 accelerators — also formally sanctioned, but that had entered China before the restrictions were introduced or through gray channels. Now, however, we are talking about Blackwell — an architecture that only began shipping to customers at the end of 2024, that is, already under existing restrictions.

The question of exactly how Blackwell accelerators could have come into DeepSeek's possession remains open. There are several likely scenarios. The first is the use of intermediaries and shell companies in third countries, through which chips are redirected to China. This channel is well documented: American authorities have repeatedly recorded instances of re-export through Southeast Asia and the Middle East. The second scenario is access to cloud computing resources based on Blackwell through overseas data centers, which formally may not violate the letter of sanctions legislation but completely undermines its spirit. The third option is direct illegal supplies, the scale of which, according to expert assessments, could be significantly larger than commonly believed.

For Nvidia, this situation is extremely awkward. Jen-Hsun Huang's company was already losing billions of dollars due to its inability to sell its best products in one of the world's largest markets. Each such report intensifies pressure from Congress, which demands even stricter measures, and simultaneously undermines Nvidia's own arguments that excessive restrictions harm American companies without achieving stated goals. The paradox is that both sides of the debate — both proponents of tightening and its opponents — can use this story to their advantage.

The consequences for the global AI race are even more serious. If DeepSeek really did train a model on Blackwell, this means the gap between Chinese and Western AI systems may be even smaller than assumed. A company that even on older equipment managed to create competitive models, having gained access to cutting-edge chips, is capable of making a qualitative leap. This calls into question the very strategy of technological containment: if sanctions do not work as a barrier but only as a slowdown, then their actual value turns out to be significantly lower than the political price paid for them.

This story will likely become a turning point in the discussion about the future of export controls. Washington will have to face an unpleasant truth: in a globalized semiconductor world, absolute control over supply chains is an illusion. Chips, like water, find their way to where there is demand for them. And as long as that demand is measured in billions of dollars, no sanctions will be able to completely cut off the flow.

ZK
Hamidun News
AI news without noise. Daily editorial selection from 400+ sources. A product by Zhemal Khamidun, Head of AI at Alpina Digital.

Want to stop reading about AI and start using it?

AI News is a curated feed of AI/tech news. Hamidun Academy teaches you to use AI systematically in your work.

What do you think?
Loading comments…