DeepSeek Shows There Is No Longer One “AI Trade,” but Many

Key Takeaways

• The AI technology boom is still in its early stages, but DeepSeek’s model marks a new phase for stock investors.
• Fund managers say that among big AI players, there will be greater divergence between the leaders and laggards.
• The changing mix of chips used for AI could benefit Broadcom and Marvell. However, a growing market could still be bullish for Nvidia, and Taiwan Semiconductor could benefit no matter which chips are used.
• Thus far, Meta has been one of the few firms to show major concrete returns for AI usage.
• The outlook for utilities depends on whether falling prices expand demand more than efficiency gains lower it.

Over the past two years, the AI trade has lifted a wide swath of technology stocks and other names. However, recent developments from Chinese company DeepSeek likely signal a new chapter in this boom. In late January, stocks that had been big winners thanks to AI were sent tumbling by the news that DeepSeek’s R1 model can solve complex problems using fewer expensive chips than models from its larger rivals. While the exact savings are unclear, this looks like a shift in AI investing, and observers think the space will see more distinct winners and losers.

“After an environment in which a rising tide lifted all boats, we expect there to be much more stock-specific differentiation,” says Felise Agranoff, one of the portfolio managers of the $22 billion JPMorgan Growth Advantage Fund JGVVX. Electricity provider stocks have also been boosted by deals with AI firms and increased energy demand for data centers. Rising AI efficiency could be bearish or bullish for utilities, depending on how falling costs affect demand for new applications.

The Importance of DeepSeek

Two major processes are involved in creating AI models: training, wherein they are given data to analyze, and inference, when they are given problems to solve using their training. Thus far, training has been largely done using GPUs—a type of chip Nvidia NVDA has a virtual monopoly on in this space. Nvidia GPUs also hold most of the market share for inference applications. A minority of chips used for inference are ASICs, which are made by other firms, with Broadcom AVGO and Marvell Technology MRVL being top names.

While not a major break with established trends, DeepSeek helped shift public attention to inference because of its large efficiency gains. “I think it brought awareness to several things happening with models that investors weren’t focused on,” says David Cross, portfolio manager at American Century Investments. Importantly, it showed how different techniques can significantly lower the costs of creating models and highlighted the importance of a different process.

“As the cost of training and inference comes down, the number of applications and the demand and adoption of those applications will accelerate,” says Tom Marsico, chief executive officer of Marisco Capital and portfolio manager on the $1.1 billion Marisco Focus Fund MIFOX. The fund is heavily invested in AI-related plays, with allocations to Nvidia, Taiwan Semiconductor Manufacturing TSM, Meta Platforms META, Alphabet GOOGL/GOOG, and Microsoft MSFT making up about a third of its assets. He argues that cheaper models will lead to an increase in demand for AI services and the chips to power them, and that this will more than make up for the lower demand due to efficiency gains.

Nvidia: Mixed Views

The declining cost of developing AI models could be good or bad news for Nvidia, depending on how falling costs affect AI usage. Agranoff is somewhat less bullish, pointing out that while companies continue to invest enormous amounts in AI, they may be more selective in deploying that money, which may favor ASIC chips. “I think there’s been more pressure to be more efficient from a cost perspective within capital spending budgets,” she says. “I think ASICs are a lot more capital-efficient and give you diversity of supply.”

As a result, the JPMorgan Growth Advantage fund is underweight in Nvidia compared with its benchmark, the Russell 3000 Growth Index. However, even with a relative underweight, Nvidia is still its largest holding, with a 9.1% weight. As to whether the rise in demand from falling prices will offset the fall from increased efficiency, Agranoff says, “It’s more of an art than a science.”

Utilities: Falling Intensiveness and Rising Demand

Utility stocks are in a similar place to Nvidia. While Cross says ASICs are less power-hungry than GPUs and could lower how much electricity each AI application uses, he predicts that the increase in AI usage from falling prices should more than offset the fall in demand from increased efficiency. He adds that while he is bullish on those stocks, investors should keep an eye on them amid any shift to ASICs.

Broadcom and Marvell: Potential Beneficiaries From a Changing Mix of Chips

Fund managers say a rise in ASIC chip demand would have two major beneficiaries: Broadcom and Marvell. Agranoff points out that Broadcom is a major supplier to Alphabet, which has continued to boost AI spending, while Amazon AMZN has partnered with Marvell. “What we’re seeing is companies such as Meta invest more in ASICs, so they need to rely more on the Broadcoms of the world,” explains Agranoff. “And so I think Broadcom has a sustainable competitive advantage.” An expansion of ASIC demand won’t happen overnight, and she thinks it would be “an evolution, not a revolution,” but that the mix of chips will change considerably over the next several years. The JPMorgan fund is overweight in Broadcom relative to its benchmark, with a 4.1% allocation.

Marsico disagrees that this trend makes Nvidia less attractive. While he concurs that ASIC demand will likely rise, he thinks the firm’s dominant position means it can better exploit the growth in AI usage caused by falling prices.

Meta: Showing Real AI Gains

While falling costs of developing and implementing models could help the AI industry grow, growth also depends on the technology demonstrating concrete returns on the huge investments in the field. While those have yet to appear for many firms, Meta has been an exception. Agranoff and Marsico pointed to Meta making gains in its ad business thanks to using AI to better target and generate engagement.

“Facebook is seeing a tremendous return from the development of their AI system on the back of how much better ad targeting is getting, the conversion rate, and thus the pricing they’re realizing,” says Marisco.

Alphabet: Potential Cannibalization of Search

While AI has benefited Meta, it may eat into Alphabet’s core search business. Marsico says it may suffer in margin and growth rate because users may increasingly rely on AI to answer questions. However, he remains bullish on the stock overall. This means Alphabet is in the unenviable position of needing to develop technology that could partially supplant its core search business.

“It’s more defensive than offensive spending, because they’re the ones most at risk, so they need to make sure to disrupt themselves,” says Agranoff. However, she cautions against being too bearish on the firm. “I wouldn’t count Google out. They’ve definitely been slower to show returns so far, which is why we have some skepticism … but their Gemini model stacks up well from a performance standpoint.”

All AI Roads Lead to Taiwan Semiconductor

Taiwan Semiconductor may benefit from rising demand for chips, regardless of whether they are GPUs or ASICs. Most major semiconductor firms don’t actually make the chips; rather, they design the chips, while specialized contract manufacturers handle the fabrication.

Taiwan Semiconductor has a majority market share in this business, making chips designed by Nvidia, Broadcom, and a plethora of other firms. “No matter what, Taiwan Semiconductor wins, whether it’s GPUs or ASICs. They’re in a really advantageous position,” says Agranoff.