“The defining feature of DeepSeek is that it had similar capabilities, at the time, to the American models like Open AI and Anthropic, but using a far smaller amount of computer chips for training that model.”
In Beijing, meanwhile, there was a palpable optimism, says Selina Xu, a researcher who works on China AI policy in the office of former Google boss Eric Schmidt. “Everybody was trying to figure out, ‘How did DeepSeek do it?’. And it’s really… been a very positive catalyst for the Chinese AI ecosystem.”
It’s also highlighted a sharp difference in how the countries operate. In the US, AI firms fiercely guard their intellectual property, but in China, there’s been a greater “open source” approach. In an effort to accelerate uptake and innovation, Chinese firms often publish their codes online, so developers from other companies can look at it.
“This means that tech companies in China, when they’re building a new AI model, don’t have to start from scratch,” says Olson. “They can just take that model and build on top of it and make it better.”
As a result, the race for AI “brains” is no longer so clear cut. America thought that LLMs were a powerful tool in its arsenal; now, China can make them too.
“The American closed-proprietary models are probably better, but maybe just not by that much,” says Selina Xu. “The Chinese model, maybe it’s only 90% as good, but it is 10% as expensive.”