DeepSeek – what’s the big deal?

The recent release of DeepSeek’s open-source AI models, V3 and R1, sent shockwaves through the investment world, triggering a selloff in U.S. semiconductor stocks. The Chinese AI lab’s claim of training its models at a fraction of the cost of leading competitors—just $5.6 mn for V3—sparked concerns about AI infrastructure spending and competitive dynamics.

“The market is worried that DeepSeek’s cost efficiency threatens the assumed dominance and valuation premium that leading U.S. semiconductor and infrastructure companies have enjoyed,” says J.P. Morgan Wealth Management in a recent note.

The launch of the new AI model from China caused shares of companies benefiting from the artificial intelligence boom to plummet worldwide earlier in the week. “It was a day of soul-searching in capital markets on day 2 of the full reckoning that AI models might be much cheaper and consume much less hardware and electricity to train and use than originally thought,” wrote DWS in a note on Wednesday. “It is unclear whether DeepSeek will be the ultimate beneficiary of the event or other companies, but the Magnificent 7 were up 2.7%, with Nvidia regaining some 40% of the market value it lost on Monday.”

However, according to Dominic Rizzo, Portfolio Manager, Global Technology Equity at T. Rowe Price, while DeepSeek’s breakthroughs in efficiency are noteworthy, “the reported training cost… may not be realistic, as it excludes several actual, very real expenses, such as experimentation.” He also highlights that DeepSeek reportedly owns 50,000 H100 GPUs, a potential $1.5 bn investment, raising questions about the true cost advantages.

And not all is bad. “There is also a positive perspective for continued large-scale training,” Rizzo points out. “DeepSeek has demonstrated that reinforcement learning is effective, and it is reasonable to assume that it could improve with increased computational power and more data. This would suggest that AI laboratories might benefit, somewhat ironically, from increasing their spending.”

Looking ahead, the AI landscape is likely to feature a mix of both large and small models, Rizzo predicts. While more powerful models may be necessary for certain consumer applications and could come at a higher cost, the emergence of lower-cost alternatives could accelerate AI adoption. Smaller models, in particular, stand to benefit from ongoing advancements in software, hardware, and the insights gained from larger-scale developments.