Shares of AI-related hardware companies plunged on Monday, including Taiwan Semiconductor Manufacturing (TSM -13.33%), Marvell Technology (MRVL -19.10%), and Arista Networks (ANET -22.35%), which fell 14.4%, 18%, and 21.9%, respectively, as of 1:11 PM EDT.

The obvious reason for the sell-off is last week’s release of the AI reasoning model DeepSeek V3 R1. Over the weekend, it became apparent that DeepSeek, a little-known Chinese AI lab, had built a leading frontier AI model with significantly less computing power than had been thought necessary.

That led most AI hardware names, many of which were trading at high valuations, to sell off hard. But how much of a risk does this new model really pose?

DeepSeek rocks the AI world

DeepSeek is a Chinese AI lab that has been in operation since 2023. It was spun out of a quantitative hedge fund called High-Flyer, founded in 2015 by entrepreneur Liang Wenfeng. DeepSeek released the large language model V3 in December, but on January 20, it released the “reasoning” version of the model, called R1. By several metrics, R1 met or exceeded the latest OpenAI reasoning model, called o1.

And DeepSeek was able to do it at just a fraction of the cost of o1, with computing power limited by 2022 and 2023 U.S. trade restrictions on leading AI GPUs. We are talking about one-tenth to one-thirtieth of the cost of models built by OpenAI and other “Magnificent Seven” names. DeepSeek’s public paper on R1 explains how the company cleverly used a variety of hacks to derive tremendous results from limited hardware. Over the weekend, its solutions withstood scrutiny by experts.

In response, major AI hardware stocks are selling off. The thinking is that companies will now be able to spend less on AI infrastructure to build leading-edge models. Given the high valuations at which some AI hardware stocks were trading on the expectation of future growth, it’s no surprise these stocks are selling off today.

This could be bad news for Taiwan Semiconductor, which has a virtual monopoly on the production of leading-edge processors for AI. Taiwan Semi just posted blowout earnings and raised expectations for capital investment in the year ahead, and its valuation had risen to over 30 times earnings heading into today.

Meanwhile, Marvell makes a variety of chips in AI infrastructure, including custom AI chips for the in-house designs of major cloud computing customers and a variety of networking and optical chips for data center communications. Marvell’s stock rocketed 83% in 2024 and came into today trading at a whopping 46 times this year’s adjusted earnings estimates.

Arista is also a leading data center networking company that builds high-performance switches. It too had a fine 2024, up 88%, and came into the day trading at 62 times earnings on the expectation of massive future AI data center buildouts.

Given these companies’ high valuations, it’s no surprise that the DeepSeek model and the doubts it brings to AI hardware spending caused a sharp correction.

Image source: Getty Images.

Is the fear overblown?

Some Wall Street analysts think the pullback is a buying opportunity for AI-focused hardware stocks. The analysts cite two major reasons. Some believe that DeepSeek’s advertised cost is below the actual cost of running the company. Others believe DeepSeek may have unauthorized access to Nvidia (NVDA -16.97%) H100 chips.

However, many in the industry believe DeepSeek’s numbers and the claims outlined in its paper. But if the cost to train and run AI models comes down materially, some say that will trigger a windfall of AI usage and demand — meaning more profit for AI hardware stocks, not less.

But it’s also possible that the rise in demand could be for inferencing hardware, which uses AI models in the real world. Up until now, hardware for training AI models has dominated spending. Therefore, investors may need to shift their focus.

And as always, high valuations increase risk should anything go wrong. Investors are certainly learning a lesson in the importance of valuation today, even for high-performing companies.