It feels like this day was always going to come - when all the artificial-intelligence names collapse - but one might have imagined it would come from, say, a profit warning from Nvidia. Instead, it comes from what appears to be an Nvidia consumer, Chinese AI service provider DeepSeek.
It works well, jumping to number-one in Apple's U.S. App store, but the issue is less the performance and more the training behind it. There are differing claims as to the number of Nvidia chips used by the company - 10,000 or 50,000, the latter roughly the size of the cluster on which OpenAI is training GPT-5 - but what is rocking markets is how cheap it was. DeepSeek's own research paper puts the training costs of one of its models at less than $6 million.
Not everyone is convinced. "While DeepSeek's achievement could be groundbreaking, we question the notion that its feats were done without the use of advanced GPUs to fine tune it and/or build the underlying [large language models] the final model is based on through the distillation technique," said Atif Malik, a Citi analyst. What is real is that DeepSeek is charging end users significantly less - 14 cents per million input tokens, vs. $15 for OpenAI's o1 model.
So people are selling first and asking questions later. The chip stocks and related companies are getting hammered: Nvidia $(NVDA)$, for sure, but also ARM $(ARM)$, Super Micro $(SMCI)$, Taiwan Semi $(TSM)$, ASML $(ASML)$, Advanced Micro Devices $(AMD)$ and Broadcom $(AVGO)$, whose new status in the $1 trillion market cap club is suddenly in jeopardy. President Donald Trump only last week held a press conference announcing up to $500 billion in AI infrastructure in a deal with SoftBank (JP:9984), Oracle $(ORCL)$ and OpenAI. Will all of that infrastructure still be needed?
As much as the tech stocks are getting crushed, at least there's a use for them in a DeepSeek-fueled future. What about the companies literally powering the AI revolution, the utilities and their suppliers? Morgan Stanley last year forecast that AI power consumption could increase from 3% of electricity consumption in the U.S. in 2023 to 10% by the end of the decade. Will that still be the case? The DeepSeek model is open source, so if their claims withstand scrutiny, the other AI models will copy their techniques and become vastly more energy efficient.
And in the power space, the group of stocks that have been most tied to the AI revolution are the nuclear plays. Shares of Oklo $(OKLO)$, a designer of fast reactors, have gained 97% this year - yes, just in January - and 289% over the last 52 weeks. Shares of Constellation Energy $(CEG)$, which thanks to a Microsoft deal is restarting Three Mile Island, is up 186% over the last 52 weeks; NuScale Power $(SMR.AU)$, a designer of small modular reactors, is up a stunning 917% over 52 weeks; and Vistra $(VST)$, an operator of nuclear plants, is up 369% over the last 52 weeks. That nuclear melt-up may soon become a meltdown.