The day after Christmas, a small Chinese startup, DeepSeek, announced a new AI system that rivals the capabilities of cutting-edge chatbots from companies like OpenAI and Google.
That alone would have been a groundbreaking event. But the team behind a system called DeepSeek-V3 has described an even bigger step. In a research paper describing how they built the technology, DeepSeek engineers say they use only a fraction of the highly specialized computer chips that major AI companies use to train their systems. said.
These chips are at the center of a tense technology race between the United States and China. The U.S. government is seeking to maintain its lead in the global AI race and is trying to limit the number of powerful chips it can sell to China and other rivals, such as those made by Silicon Valley company Nvidia.
However, the performance of the DeepSeek model raises questions about the unintended consequences of the US government's trade restrictions. The restrictions are forcing Chinese researchers to get creative with a wide range of tools available for free on the internet.
Benchmark tests used by US AI companies show that DeepSeek chatbots answer questions, solve logic problems, and create unique computer programs that are as capable as anything already on the market.
And it builds on the popular idea that only the biggest companies in the tech industry (all based in the US) can afford to create cutting-edge AI systems on the cheap and challenging. Ta. Chinese engineers said they only needed about $6 million in computing power to build the new system. That's about a tenth of the amount tech giant Meta spent building its latest AI technology.
“The number of companies spending $6 million is far greater than the number of companies spending $100 million or $1 billion,” said Chris V. Nicholson, an investor at venture capital firm Page One Ventures. he says. AI technology.
Since OpenAI sparked an AI boom in 2022 with the release of ChatGPT, many experts and investors had concluded that no company could compete with market leaders without spending hundreds of millions of dollars on specialized chips. .
The world's leading AI companies use supercomputers with more than 16,000 chips to train their chatbots. Meanwhile, DeepSeek engineers said they only need about 2,000 of Nvidia's specialized computer chips.
China's restrictions on chips meant that DeepSeek's engineers “needed to train the chips more efficiently to stay competitive,” said George Washington University professor of emerging technology and international relations. Associate Professor Jeffrey Ding says.
Earlier this month, the Biden administration announced new rules aimed at blocking China from obtaining advanced AI chips through other countries. The rules build on multiple previous rounds of restrictions that prevented Chinese companies from buying or manufacturing cutting-edge computer chips. President Trump has not yet indicated whether he will enact or rescind the rules.
The U.S. government has sought to keep advanced chips out of the hands of Chinese companies over concerns that they could be used for military purposes. In response, some Chinese companies are stockpiling thousands of chips, while others are sourcing them from a thriving underground market of smugglers.
DeepSeek is operated by a quantitative stock trading company called High Flyer. By 2021, the company plowed its profits into acquiring thousands of Nvidia chips, which it used to train early models. The company did not respond to requests for comment, but is known in China for scouting recent graduates from top universities with the promise of high pay and the ability to pursue the most intriguing research questions.
Zihan Wang, a computer engineer who worked on early DeepSeek models, said the company was trying to understand the technology to be able to generate poems and ace questions for China's notoriously difficult university entrance exams. The company also hires people without computer science backgrounds.
DeepSeek does not manufacture any consumer products and its engineers focus entirely on research. This means the company's technology is not bound by the strictest aspects of China's AI regulations, which require consumer technology to comply with government information controls.
Leading U.S. companies continue to advance the cutting edge of AI In December, OpenAI announced a new “inference” system called o3 that outperforms existing technology, but is not yet widely used outside the company. No. But DeepSeek continues to show that it's not far behind. This month, the company released its own impressive inference model.
(The New York Times sued OpenAI and its partner Microsoft for copyright infringement of news content related to AI systems. OpenAI and Microsoft have denied these claims.)
An important part of this rapidly changing global market is the old idea of open source software. Like many other companies, DeepSeek is open sourcing its latest AI systems. This means we share the underlying code with other companies and researchers. This allows others to build and distribute their own products using the same technology.
While employees at large Chinese technology companies are limited to collaborating with colleagues, “working on open source means you're working with talented people from all over the world,” says the CEO at Baseten in San Francisco. said Inen Chan, lead software engineer working on the open source SGLang. project. He helps other people and companies build products using DeepSeek's systems.
The open source ecosystem for AI gained momentum in 2023, with Meta freely sharing an AI system called LLama. Many thought this community would only thrive if companies like Meta – a tech giant with massive data centers packed with specialized chips – continued to open source their technology. . But DeepSeek and other companies have shown that the power of open source technology can be extended. ”
Many executives and experts have argued that large American companies should not open source their technology. Because technology can be used to spread disinformation and cause other serious harm. Some US lawmakers are exploring the possibility of preventing or curbing this practice.
But some argue that China would gain a significant advantage if regulators stifled open source technology advances in the United States. They argue that if the best open source technologies come from China, U.S. developers will build systems on top of those technologies. In the long term, China may become the center of AI research and development.
“The center of gravity of the open source community is shifting toward China,” said Ion Stoica, a computer science professor at the University of California, Berkeley. “This could pose a significant risk to the United States” because it would allow China to accelerate the development of new technologies.
Hours after taking office, President Trump rescinded a Biden administration executive order that threatened to curb open source technology.
Dr. Stoica and his students recently built an AI system called Sky-T1 that rivals the performance of OpenAI's latest system called OpenAI o1 in certain benchmark tests. The computing power required was only $450.
They did this by building on two open source technologies released by Chinese tech giant Alibaba.
Their $450 system isn't as powerful as OpenAI's technology or DeepSeek's new system. And the technology they used is unlikely to produce a system that exceeds the performance of the leading technology. But this project showed that it is possible to build a competitive system even with very low resource operations.
Reuven Cohen, a technology consultant in Toronto, has been using DeepSeek-V3 since late December. He says it's comparable to state-of-the-art systems from OpenAI, Google and San Francisco startup Anthropic, and costs much less to use.
“DeepSeek is a way for me to save money,” he said. “This is the kind of technology that people like me want to use.”
