Disclaimer: Views in this blog do not promote, and are not directly connected to any Legal & General Investment Management (LGIM) product or service. Views are from a range of LGIM investment professionals and do not necessarily reflect the views of LGIM. For investment professionals only.

28 Jan 2025
5 min read

Who’s afraid of DeepSeek?

The Chinese firm’s groundbreaking R1 AI model challenges many assumptions on the state of the technology.

Deep Seek

Since the AI boom got rolling in US stocks, there have been a few generally accepted truths about how the technology would play out.

Firstly, you needed large volumes of computational power to train models classed as state-of-the-art. Second, you needed large financial resources to pay for that computational power. Third, it was the US that was setting the pace, with other countries unable to keep up with Silicon Valley’s deep pockets and engineering talent. 

All three assumptions are now in doubt after Chinese AI company DeepSeek* launched its R1 model (the latest in a class of ‘reasoning’ models to handle more complex tasks). The R1 model is not only comparable in terms of performance to Open AI’s o1 model (scoring similarly in coding and maths tests), it is also vastly cheaper, charging $0.14 per million tokens (pieces of sentences) compared to OpenAI’s $7.50. 

The model also writes out its thought process (so-called Chain-of-Thought) in ways users find helpful. DeepSeek’s relatively small team of researchers, based in China’s tech heartland of Hangzhou, have also published the model’s source code to all, thrilling researchers. To top it all, DeepSeek reports that costs were minor compared to most AI models to train, at just under $6m. 

Seeing market vertigo?

DeepSeek’s advances have US equities feeling vertigo after rising so high in the past four years. In particular, DeepSeek’s claim that they produced their model without using an expanded amount of high-end GPUs (graphics processing units) of the kind Nvidia* produces has chipmaking stocks (such Nvidia and Taiwan Semiconductor Manufacturing Company*) swooning. Large-cap equities had been pricing in a substantial build-out of these cutting-edge chips as the AI boom steadily spread.

Does DeepSeek threaten those assumptions, and in so doing, threaten the US AI complex? While it’s too early to say, it wouldn’t be the first time. The dot.com bubble saw tech hegemons like Oracle* and Cisco* soar in value, until it was clear their monopoly on high-end servers was too expensive for the typical user, only then to crash in a market rout. Could Nvidia await the same fate?

What’s DeepSeek’s secret recipe?

In light of DeepSeek’s remarkable engineering feat, researchers have been poring over the firm’s technical report to explain how it achieved such incredible results at low cost. As with any breakthrough, there is scepticism of its reliability. This scepticism takes three main forms.

Firstly, on cost. DeepSeek’s reported minimal model cost of just $5.5m likely refers to training only, not the cost of expenditure on computation, staffing or running the model, all of which will be substantial (albeit perhaps covered by local government). 

Secondly, on chips. Detailed research finds that despite Biden-era controls on China’s access to AI chips, many thousands of GPUs will have entered China anyway, casting doubt on claims DeepSeek had no additional chip capacity. 

Lastly, on design. DeepSeek’s prior models (like its V3 model) have referenced OpenAI explicitly when asked some queries, leading people to think the firm’s R1 model is piggy-backing on OpenAI’s technology, rather than developing new reasoning capabilities.

While the numerous efforts to raise efficiency throughout DeepSeek’s process have undoubtedly wrung substantial efficiencies out of the model, it’s unclear how much of their success is attributed to these unmentioned factors.

The Trump factor 

With US-China relations in the spotlight, Cold War comparisons were quick to come by. Marc Andreessen, Silicon Valley luminary, described DeepSeek’s R1 model as a “Sputnik moment” for AI. 

Those familiar with the Cold War Space race will know that Sputnik’s launch was a foundational moment for NASA and the US government’s quest to put a man on the moon. But it also led to fears (proven baseless) that there was a “missile gap” between the US and USSR.

Today, given geopolitical tensions, it’s possible national security aims could supersede economic ones. We think Trump’s new administration might find any Chinese tech advantage challenging – and so devote government resources at winning the AI race. That could present potential upside for many US-based AI firms, in our view.

Portfolio implications

Suspicions aside, we believe that DeepSeek’s groundbreaking method to raise performance could augur well for AI adoption. A cheaper product could encourage more users to harness the technology for profitable use; although Cisco’s servers built in the dot.com era proved loss-making for some time, they are part of why you are reading this on the webpage you are now. 

There could be clear benefits to the entire corporate sector should a productivity-enhancing technology become available at a significantly lower price. The MSCI EAFE and even the S&P equally-weighted index were up slightly on Monday as the market digested this potentially game-changing news.

Nevertheless, for the AI-supercharged US equity index, heightened valuations and a sudden injection of uncertainty might be problematic. While we don’t have a tactical view on US equities today our process has always emphasised diversification**, leading us to hold a smaller amount of US megacaps than a pure market-capitalisation-based approach. Regardless of what comes next, this episode is a useful reminder of the potential benefits of spreading our bets.

 

*For illustrative purposes only. Reference to a particular security is on a historic basis and does not mean that the security is currently held or will be held within an LGIM portfolio. The above information does not constitute a recommendation to buy or sell any security.

**It should be noted that diversification is no guarantee against a loss in a declining market.

United States Technology Asia China
Generic author image

Matthew Rodger

Assistant Economist

Matthew is an economist covering emerging markets. He uses countries’ historical experience, alongside fresh economic data and quantitative methods, to recognise new investment opportunities. Prior…

More about Matthew

Recommended content for you

Learn more about our business

Legal & General Investment Management is one of the world's largest asset managers, with capabilities across asset classes to meet our clients' objectives and a longstanding commitment to responsible investing.

Image of London skyscrapers

Sign up for blog email alerts

Receive the latest articles in a weekly digest by registering via the email preference centre