Chinese AI Company DeepSeek Accused of Getting Around US Sanctions to Buy NVIDIA Chips

A Surprising Twist in the AI Arms Race

DeepSeek, a Chinese AI company, is making waves in the US market with its DeepSeek-R1 model, which is reportedly challenging other open-source large language models (LLMs) like ChatGPT-4. What’s more, DeepSeek is achieving this at a lower cost and with less computing power. In fact, the DeepSeek-R1 AI chatbot recently topped the charts as the most popular free app on Apple’s App Store, surpassing its rival ChatGPT.

Getting Around US Sanctions

A recent report suggests that DeepSeek was able to circumvent US restrictions on AI chips by purchasing advanced Nvidia semiconductors through third-party intermediaries in Singapore. This raises questions about whether China is gaining an unfair advantage in the AI arms race.

Nvidia’s Response

Nvidia, the manufacturer of the AI chips, insists that its partners must comply with all applicable laws. However, the company did not confirm whether DeepSeek violated any laws.

But How Accurate is DeepSeek?

A recent test among 11 AI platforms found that DeepSeek finished only 10th in accuracy, answering questions correctly only 17% of the time. Moreover, 30% of the questions handled by DeepSeek were answered with false claims, and more than half the time (53%) the Chinese AI chatbot gave a vague or not useful answer to news-related queries.

Investigations Underway

The White House and FBI are investigating whether DeepSeek used intermediaries in Southeast Asia to get around US restrictions. Nvidia has also stated that its partners must act within the law.

Howard Lutnick’s Take

Howard Lutnick, President Donald Trump’s pick to head the Commerce Department, has said that it’s okay for DeepSeek to compete with US AI companies, but they need to stop using America’s own tools (like Nvidia chips) to compete with us.

Singapore’s Role

Some statistics suggest that Singapore might have been involved in helping DeepSeek obtain Nvidia chips. The percentage of Nvidia’s revenue derived from the country rose from 9% to 22% over two years.

DeepSeek’s Admission

DeepSeek has admitted to using 2,048 of Nvidia’s H800 GPU chips to train its V3 model, and its R1 model was likely trained on a more powerful machine that required the use of more advanced Nvidia GPUs that can’t be sold in China.