DeepSeek shows the US failure to contain Chinese tech

For the past couple years, the narrative out of Silicon Valley has been clear: generative AI is transforming everything, bigger models would inevitably deliver better outcomes, and virtually anything had to be sacrificed on the altar of artificial general intelligence. It didn’t matter if that meant burning more fossil fuels, threatening the power grid, or our ability to even dream of hitting our emissions targets. Nothing could stand in the way of the AI future.

Just last week, OpenAI CEO Sam Altman was at the White House alongside president Donald Trump, Oracle CEO Larry Ellison, and Softbank CEO Masayoshi Son to announce a major new initiative: the corporate titans would direct as much as $500 billion into AI infrastructure as part of their Stargate project. Microsoft, Meta, and other major US tech companies were quick to remind investors they had tens of billions of dollars of data center investment of their own planned. Now that might all be up in the air.

Chinese company DeepSeek — an offshoot of a hedge fund — has released a new AI reasoning model called R1 that is challenging the very core of the narrative Altman and his peer have been trying to sell us. As OpenAI, Anthropic, and Google have been struggling to make their models better even with greater scale, DeepSeek has produced a model that’s competitive with the supposed American leaders at a mere fraction of the training cost — an estimated $5.6 million, compared to about $100 million for OpenAI’s GPT-4.