A Chinese AI startup, DeepSeek, has released an open-sourced problem-solving model, R1, that’s wowed Silicon Valley. The model uses far less computing power and far fewer chips – therefore far less money, roughly 3-5% of the development costs for ChatGPT – to achieve the same or better results as its US counterparts.

The model is open-source, which has allowed engineers outside of China to audit its parent company’s claims. Venture capitalist Marc Andreessen, who is advising the Trump White House, called R1 “AI’s Sputnik moment”, a bona fide breakthrough. The global AI community widely considered the US the leader in AI, but R1 has called that dominance into question.

DeepSeek’s training, done with tens of thousands of Nvidia’s chips, may also undermine the effectiveness of the US’s AI-focused trade embargoes on China. The sale of the company’s products in China are strictly regulated, but DeepSeek was able to secure about 50,000 graphics processing units (GPUs) anyway, per VentureBeat. That’s a far cry from the roughly 500,000 that OpenAI is reported to use, though. Another complicating factor: trade restrictions take effect over long periods of time, so they may not have impinged on DeepSeek’s work.