Market experts take a deep dive into DeepSeek

Chinese artificial intelligence (AI) start-up DeepSeek shocked markets with its release of an AI model that it said had been developed for a fraction of the cost of the industry-leading models in the US. The market’s volatile reaction was exemplified by tech giant Nvidia.

On Monday, 27 January, Nvidia made history by losing $589 billion from its market cap, the largest single-day loss in value by a company in history.¹ The next day, it gained back $260 billion — the second-biggest single-day value gain.¹

Investors are grappling with the questions raised by these events. To help sort through the wide variety of issues, we spoke with a range of Invesco experts across markets and asset classes. Here’s what they had to say.

Global markets

Kristina Hooper, Chief Global Market Strategist

It was reported that DeepSeek has found a way to create an AI model that does not require the kind of sophisticated chips that thus far were believed to be necessary. It was reported that the model demonstrated performance comparable to some of the best AI models on the market today, though there are some notable weaknesses (such as in code generation). Skeptics have highlighted that the training data used in producing the model benefited from prior work done by other AI companies. Nevertheless, DeepSeek’s announcement sent US AI-related stocks plummeting on Monday, 27 January.

This illustrates the dangers of high valuations. When priced for perfection or near perfection, a stock is more vulnerable to a significant sell-off – even based on news flow with little in the way of details. I must underscore that this is not a time for knee-jerk reactions since we know so little at this juncture. We’ll need to learn a lot more about this before we can decide whether this is a serious threat to US AI companies – and whether it could present a significant opportunity to make AI more accessible at a far lower cost. We do think it could prompt more scrutiny on companies’ AI investment spending.

Emerging markets

Justin Leverenz, Chief Investment Officer, Developing Markets Equities, US

AI “scaling” is a function of three variables — models, training data, and compute (the computational resources used to train AI). The market’s reaction to the news about DeepSeek was a function of the anxiety around what appears to be a massive improvement in model optimisation, which could drive down the cost of edge AI. While this is good news for demand and the adoption of AI, it has raised questions about the continued need for the massive datacenter capex spending that has facilitated brute force compute in the early phase of the AI revolution. The concern is that, like the fallout from the excessive investments in bandwidth during the first Internet bubble, these model efficiency gains could potentially create a major air pocket for the “picks and shovels” first-wave beneficiaries.

The sell-off is clearly about the sustainability of near-term growth in AI semiconductor manufacturers, given the huge performance/cost gains of DeepSeek. And that leads to concerns about foundry/memory/semi-cap equipment/datacenter hardware companies. While there is still a lot of uncertainty, and disagreement, about whether the hyperscalers will moderate capex in response, it is eminently clear, in our view, that lower inference costs will lead to a variety of benefits for underlying companies that actually use AI. 

We believe that the intensity of the market’s reaction to DeepSeek’s news is a reflection of the massive expectations, high valuations and market concentration in US tech stocks.  We don’t intend to imply that Asian hardware companies are not going to come under pressure. However, we believe that valuations remain relatively attractive.

Tech stocks

Nick Kalivas, Head of Factor and Core Product Strategies, Invesco ETFs, US

Skeptics are questioning DeepSeek’s cited expenditure of $6 million, warning that there’s no way to verify the total amount spent, including upfront costs that could be measured in billions. However, the success of the model has prompted the idea that the costs for AI have the potential to be lower than previously thought.

DeepSeek’s model is entirely open-source, and large language model developers in the US will look to implement its efficiency gains into their models. This would lower the need for future capital expenditures to continue growing at the same rates.

DeepSeek’s narrower model has prompted the idea that companies do not need extremely large computation resources and that AI may be very effective on smaller and more targeted tasks, opening up uses on smartphones and computers.

The Jevons Paradox says that an increase in efficiency in the use of a resource will generate an increase (rather than a decrease) in resource consumption. In terms of DeepSeek, this means that US companies can repurpose their AI capex sooner, moving away from massive investments in training generative AI models to more complex AI use cases.

We saw a short-term bearish impact from the DeepSeek news, but the market has been sorting out potential winners and losers and generated some rotation in the technology space.  The revelations from DeepSeek have the potential to benefit companies that were once thought to need a large spend on graphic processing units (GPU), and those that can bolster and secure AI workloads on smartphones and computers.

Ash Shah, Senior Portfolio Manager, Senior Research Analyst, Discovery Growth team, US

In any technology, when costs come down, the volumes go up. Think about personal computers, for example. They used to be really heavy and cost around $4000. But then prices went down, and you went from one personal computer in a household to everyone in the family having one. We’ve seen this same pattern with a variety of technologies.

So, this is going to be no different, in my opinion. The prices come down, demand will likely explode. And the more the prices come down, the better it is because then all sorts of companies can utilise AI for a lot less. I think this is really great for software companies because they can run their software on top of the AI models.

In the short run, I think people are going to be a little skeptical because they saw stocks get hammered immediately on the DeepSeek news. But then a rebound came the next day. So, I think people are going to come to realise that this is an evolutionary thing, but it’s not a revolution.

Diversification

Bradley Smith, Director, International ETF Specialist, US

The reaction to DeepSeek illustrates the need for diversification. The S&P 500 Equal Weight Index outperformed the market-cap-weighted S&P 500 Index (which is top-heavy with tech stocks) by 1.48% on 27 January, when the DeepSeek news first hit the market.²

It's not just that the weight in the top 10 names in the S&P 500 is near a record level, it's that many of these large names are doing similar things in the tech space. This means that the sources of return driving the S&P 500 Index have greatly diminished. Before COVID, the S&P 500’s composition was drastically different. With the rise of mega-cap technology names, this has changed — and in my view, it underscores the need to add alternative sources of return to a portfolio, which could include an S&P 500 Equal Weight strategy, commodities, non-core fixed income, private markets, and more.

Fixed income

Matt Brill, Head of North America Investment Grade, US

We view the impact of DeepSeek to be more impactful to economic growth and equities than credit. The five large-cap names gathering most investors’ attention this week were Apple, Microsoft, Nvidia, Amazon, and META, which have a combined weight of around 40% of the NASDAQ Composite Index, yet only around 2% of the US investment grade market.³

In addition, these companies are highly rated credits with an average rating of AA and all have negative net debt, which means they have more cash than debt. The impact of the potential fallout from DeepSeek on credit metrics for large-cap technology firms should be immaterial, in our view.

Drilling down into corporate fundamentals, in the near term, certain hardware and semiconductor names could experience growth headwinds if AI-related capital expenditures get trimmed. However, US investment grade issuers that fall into that category generally benefit from some combination of low leverage and diversified business lines, so credit profiles should be insulated from even significant drops in AI capital expenditure. Additionally, lower costs to run AI models and lower capital expenditure — if they materialise — should directly benefit the largest AI spenders.

Macro view

Ben Jones, Director of Macro Research, Multi-Asset Strategies, UK

Those who are saying that the DeepSeek-related stock sell-off changes the outlook for growth, inflation, and rates are overblowing the story, in my view. No doubt, the rally in US equities in recent years has meant looser financial conditions and generated a wealth effect in the US, which supported consumer spending. If US stock indices fall heavily, conditions would tighten, and the wealth effect could reverse, undermining the consumer growth story in the US and possibly leading to recession. But there is little reason to think that should be the base case now. The sell-off was narrow. Nvidia lost nearly $600 billion on Monday, 27 January, but 70% of the S&P 500 Index put in a positive return, including three of the Magnificent 7 names.⁴

That said, the AI story is shifting, and US stocks face a higher bar to deliver on earnings given stretched multiples. I therefore have more confidence that we will see markets outside the US outperform this year. I think calling the top in the US dollar now is a little too early, largely because I think there is room for the gap between US yields and the rest of the world to widen. Within the US, repositioning portfolios to lighten up on US mega-cap stocks and reduce concentration risk through greater diversification seems very sensible to me today. But it did before the DeepSeek news, too.