“Our real challenge has never been funding, but the ban on the export of state-of-the-art chips,” DeepSeek founder Liang Wenfeng said in a recent interview, referring to US restrictions on semiconductor exports to China.
It was because of the United States measures against the Chinese chip industry that DeepSeek and other Chinese Artificial Intelligence teams began to wonder if they could improve the algorithm and training mode of Large Language Models (LLM), says Zhuang Fei, content director at Wave Media.
The Internet and technology expert believes that the ban on exporting the most advanced chips to China “forced Chinese programmers and engineers to innovate.” “They realized they could get better results through innovation [rather than more advanced chips].”
Chinese startup DeepSeek made a big impact on the Artificial Intelligence (AI) sector last month when it launched its latest DeepSeek R1 Artificial Intelligence model, which is open source, meaning anyone can access, modify and distribute it.
The announcement caused the biggest drop in the value of a company on the stock exchange in just one day in the United States. The company NVIDIA lost almost 600 billion dollars, 17% of its value.
The reason is that DeepSeek's tool managed to perform like the most advanced models without using the most powerful chips on the market and with significantly fewer resources than companies such as Sam Altman's OpenAI and Mark Zuckerberg's Meta, both from the United States. Around 80% of the AI chip market is dominated by NVIDIA.
Innovation first instead of profit
For Zhuang Fei, the emergence of DeepSeek may have transformed the AI industry forever. It changed from being an investment and capital-intensive sector to an innovation-intensive one.
The fact that DeepSeek has achieved the same efficiency and functionality as the most advanced AI models on the market, such as OpenAI’s ChatGPT, through innovation in training methods and code has completely overturned the idea created by the US technology sector that you need to have enough chips and computing power to develop an AI model that is good enough, argues Zhuang.
The company reportedly used around 2,000 H800 chips from NVIDIA, which are not the most advanced. The marketing of the latest NVIDIA chips was banned in the October 2022 round of US sanctions on China.
Part of the impact also lies in the free nature of the Chinese tool. ChatGPT has a free version and versions with monthly subscriptions ranging from 20 to 30 dollars, and one launched in December that costs 200 dollars a month.
Google's Gemini tool also comes in versions that cost around 20 dollars. Meta's tool is priced by “token”, which are units or blocks in LLMs. LLMs are models that can generate and interpret human language trained on a huge amount of data. Tokens can be words, scores, etc.
“People like Sam Altman [OpenAI’s director] have created a way of thinking about the development of Artificial Intelligence based on profit,” says Zhuang Fei.
This will make companies like Altman's and other big techs very cautious about charging when launching their AI products in the future. At least they won't dare charge very high prices for new AI services, the Chinese expert argues.
The Global North’s reaction to DeepSeek
Following DeepSeek’s launch of its latest tool, some countries, mostly in the Global North (United States, Italy, Australia, South Korea), began to ban or restrict access to it. The argument is that DeepSeek poses a risk to their national security.
China's Permanent Representative to the UN, Fu Cong, rejected the insinuations last week at a press conference at United Nations Headquarters in New York, during the beginning of China's presidency of the Security Council.
“Technological containment and technological restriction do not work. I believe this is a lesson the whole world should learn, particularly the US. I've said it before, and I'll say it again: Never underestimate the creativity of Chinese engineers and scientists,” said Fu Cong.
Zhuang Fei considers the reaction of the countries of the Global North “sad and ridiculous”. “The ridiculous thing is that DeepSeek is open source. They released a document with detailed instructions on the optimizations and what the company did with the training data,” explains Zhuang. “A Brazilian or Indian programmer can move DeepSeek to their local server and then modify it, tweak it, debug it, and completely transform it into a tool tailored to their needs.”
DeepSeek has also been accused of utilizing OpenAI AI models to develop its tools. Zhuang Fei replied that almost none of the Chinese company's employees studied abroad or worked in Silicon Valley. “Almost all of its team of around 150 people graduated from Chinese universities and ed DeepSeek. In other words, it has little to do with the United States,” he concluded.
