(Credit: Andrey Rudakov / Bloomberg / Getty Images)
News of an open-source large language model from China’s DeepSeek is blowing up social media and the stock market this week.
The company's R1 model reportedly runs on lower-capacity and less expensive Nvidia H800 chips. Its AI chatbot is now the App Store's most downloaded app, having bumped off ChatGPT. It can even run locally on a user’s computer.
Some reports indicate that DeepSeek was developed for $6 million. However, as USA Today notes, a research paper accompanying DeepSeek's launch "noted that this cost referred specifically to chip usage on its final training run, not the entire cost of development."
Still, that pales in comparison to what US tech companies are set to spend on AI in the near future. DeepSeek arrived just days before executives from OpenAI, Oracle, and Softbank showed up at the White House and pledged to invest $500 billion over the next four years in AI data centers, an effort they dubbed Project Stargate.
At Meta, the metaverse will take a backseat to $65 billion in AI spending. Microsoft plans to spend $80 billion on AI-enabled data centers in FY 2025. DeepMind CEO Demis Hassabis estimates it could cost Google $100 billion to develop the next generation of AI tech. OpenAI CEO Sam Altman reportedly wants trillions.
How We Got Here
For an industry that is supposed to be a step ahead, ChatGPT's arrival in 2022 shocked US tech giants, but they were quick to go all-in, replacing journalists with AI-generated news, firing entire customer service teams, and deciding that computers, not humans, could design video games.
They are now scrambling to respond to DeepSeek. Venture capitalists who spent nearly half their cash on AI last year hope DeepSeek doesn't bring about what Axios terms an extinction-level event.
While the tech industry is always on edge, anticipating another bubble like the one that burst in 2000, a feeling of too big to fail has permeated the top Silicon Valley companies when it comes to AI, particularly now that they’ve aligned themselves with the Trump administration, which can deflect blame and consequences like no other in history. Ironically, it is them being too big in both size and spending that gave DeepSeek an advantage.
DeepSeek was operating at a disadvantage, since an export ban prevented it from having access to the same powerful Nvidia chips that US companies use. And yet it strung together a much cheaper solution with far fewer and lower-powered chips for results that are as good or better than those competitors. (Despite the fact that it runs Nvidia chips, DeepSeek's launch resulted in Nvidia losing almost $600 billion in one day, the worst a company has ever fared in the market.)
It's All Too Familiar
While DeepSeek has caused a shakeup, it is not shocking. The empty promises of crypto, NFTs, and the metaverse are not far behind us. Silicon Valley has turned valuation into the only product on which it reliably delivers.
Despite the best efforts of the AI hucksters, the technology itself has not lived up to the hype. AI is notoriously unreliable for the average person, mucking up internet searches with no end of erroneous results in Google’s AI Overviews, hallucinating answers to chatbot queries, and reporting fake headlines in news alerts. (DeepSeek might have a propaganda problem of its own.)
And despite Oracle's Larry Ellison promising at the White House that Project Stargate will fuel cancer research, Donald Trump's own actions may make that more difficult. It reminds us of IBM's Watson going from promises of tailored cancer treatment to working at call centers.
AI's biggest value so far has been to shareholders, and if it can’t deliver on that, the companies that have backed it will execute their most popular move: the pivot.