DeepSeek Won't Sink U.S. AI Titans

Necessity might be the mother of all invention, but sparking the mother of all selloffs seemed like a stretch. That wasn't the case Monday morning, though, as U.S. markets opened to fresh fears about DeepSeek.

- Dan Gallagher

The Chinese artificial-intelligence startup announced a significant breakthrough late last week with AI models that perform nearly on par with advanced U.S.-born technology.

The rub is that DeepSeek claims to have trained one of its latest models for $5.6 million in computing costs—a fraction of what is currently spent on this side of the Pacific on the same activity. OpenAI's GPT-4 model that was launched in late 2023 cost more than $100 million to train, according to Chief Executive Sam Altman. In a podcast last year, Anthropic CEO Dario Amodei said the cost to train some models is approaching $1 billion.

Such blowout costs have been great news for companies such as Nvidia, Broadcom and Marvell, which have seen their market values surge because of explosive demand for AI chips and services. The high entry price of AI—and sanctions from the U.S. government limiting the sale of advanced AI chips to Chinese companies—also serve as a competitive moat for tech titans such as Microsoft, Amazon, Google and Meta Platforms. They are among the few companies with enough capital to build out expensive AI networks on a large scale.