OpenAI CEO Sam Altman downplayed the significance of a new artificial intelligence (AI) model released by Chinese startup DeepSeek on Thursday, saying it did a “couple of nice things” but has been “wildly overstated.”
DeepSeek’s R1 model, launched last week, has shaken up the American AI industry. The startup has claimed it cost just $5.6 million to train the model with a couple thousand reduced-capability chips, raising questions about the level of investment needed to develop AI.
“They did a couple of really nice things, but on the whole, that is wildly overstated,” Altman said during an event in Washington, D.C. “Again, this is a model at a capability level that we had quite some time ago.”
“We’ve got a sense of what it takes to run a cost-efficient model,” he continued. “So again, I don’t want to diminish the work. I think this is a great work, and it’s, in some sense, it’s like the first surprise or new competitor to OpenAI in quite a while.”
DeepSeek’s rise has threatened to upend the current consensus on AI, which assumed massive investment in chips and data centers was necessary to build new AI models. Leading American AI companies, like OpenAI, have spent billions of dollars on infrastructure in recent years.
Emblematic of this consensus is the Trump administration’s new Stargate project, which seeks to invest up to $500 billion in AI infrastructure over the next four years, with OpenAI, Oracle and SoftBank taking the lead.
Altman argued Thursday that greater compute and infrastructure is still needed to build frontier models and keep up with the growing adoption of AI.
“The cost to do a frontier model, to be at the very edge, that goes up — the cost of training, compute, test time, compute, data, whatever you want,” he said, adding, “The amount of demand there is to use a given level of AI, as you bring that price down, goes up by much more than the decrease in price.”
He also appeared cavalier about the potential for distillation, a technique used to transfer the knowledge of a large model to a smaller model.
OpenAI has said it is investigating whether DeepSeek “improperly distilled” data from its models. The company’s terms of service bar users from utilizing distillation to build competing models.
“People will distill our models,” Altman said. “Whatever, that’s fine. “
The OpenAI CEO touted DeepSeek’s R1 model earlier this week as “impressive,” although he didn’t appear as concerned as some in and around the tech sector.
“deepseek’s r1 is an impressive model, particularly around what they’re able to deliver for the price,” Altman wrote Monday night on social platform X. “we will obviously deliver much better models and also it’s legit invigorating to have a new competitor! we will pull up some releases.”
Others have taken a more sensational approach. President Trump called DeepSeek a “wake-up call” for the American AI industry, while former Google CEO Eric Schmidt said it could represent a “turning point” in the global AI race toward open-source models.
