China's DeepSeek AI model challenges US dominance amid sanctions

Listen to article

A little-known Chinese AI startup, DeepSeek, has disrupted Silicon Valley’s hold on artificial intelligence with the release of R1, a reasoning model that rivals OpenAI’s ChatGPT o1.

Despite working under the constraints of US-imposed export sanctions on advanced chips, DeepSeek has developed an open-source model praised for its efficiency and accessibility.

The Hangzhou-based company claims that R1 outperforms leading AI systems in tasks requiring complex reasoning, including mathematics and coding. DeepSeek also unveiled smaller versions of R1, capable of running on local devices such as laptops, making advanced AI tools more accessible to researchers and developers with limited resources.

Innovation amid constraints

DeepSeek’s success is even more impressive given the limitations imposed by US export controls on semiconductors. These restrictions prevent China from obtaining high-performance chips like Nvidia’s H100s. DeepSeek, however, leveraged a stockpile of older Nvidia A100 chips, acquired before the sanctions, and lower-capacity H800 chips to train its models.

To overcome these hardware limitations, DeepSeek engineers optimised their training processes to use less memory and computational power. This efficiency-first approach allowed them to match the performance of systems built with far greater resources.

“The US export control has essentially backed Chinese companies into a corner where they must be far more resourceful,” said Matt Sheehan, an AI researcher at the Carnegie Endowment for International Peace. “It’s led to innovation.”

Global recognition

Microsoft CEO Satya Nadella highlighted DeepSeek’s achievements at the World Economic Forum in Davos. “It’s super impressive how effectively they’ve built a compute-efficient, open-source model. Developments like DeepSeek’s should be taken very seriously,” he said.

The model employs a “chain of thought” reasoning approach similar to ChatGPT o1, solving problems step by step. Unlike other systems, R1 emphasises accurate answers rather than detailing every logical step, reducing computation time without sacrificing quality.

Dimitris Papailiopoulos, a principal researcher at Microsoft, praised the model for its simplicity. “DeepSeek focused on core performance, which significantly cut costs while maintaining effectiveness,” he said.

Rising competition

DeepSeek’s rise reflects broader trends in China’s AI sector, which has embraced open-source principles. Startups like Minimax and 01.AI have released their models for public use, while tech giants such as Alibaba have launched over 100 open-source AI solutions.

According to the China Academy of Information and Communications Technology, China now accounts for 36% of global AI language models, second only to the United States.

This shift is driven by a new generation of Chinese researchers who value open-source collaboration, says Tufts University professor Thomas Qitong Cao. “Young researchers benefit immensely from open-source culture and are contributing back to it,” he said.

The future of AI competition

DeepSeek was founded in 2023 by Liang Wenfeng, an alumnus of Zhejiang University and the founder of the hedge fund High-Flyer. Liang’s decision to venture into AI stemmed from recognising the potential of the chip stockpile his firm had acquired.

“Chinese companies have historically used more computational power to achieve the same results. Our goal is to close those gaps through innovation,” Liang said in a 2024 interview.

As competition in the AI space intensifies, analysts expect further consolidation and collaboration. Alibaba recently partnered with Kai-Fu Lee’s 01.AI to establish an “industrial large model laboratory,” signalling a strategic move toward resource sharing.