Tech stocks slide on DeepSeek concerns
CIO Daily Updates
From the studio:
From the studio:
Thought of the day
Thought of the day
What happened?
US tech shares fell on Monday amid investor concerns about the potential disruption caused by Chinese artificial intelligence startup DeepSeek to the AI-driven US stock rally over the past two years.
On 20 January, DeepSeek, which was founded by hedge fund manager Liang Wenfeng, released its latest open-source AI model together with a paper explaining how it was built at a fraction of the training and inferencing costs compared to leading models developed by better-resourced US companies. DeepSeek's prior releases, such as v3 launched on 26 December 2024, have also been computationally efficient and outperformed established peers.
The tech-heavy Nasdaq fell 3%, while the S&P 500 closed 1.5% lower on Monday. Among the big US tech firms, chipmaker NVIDIA was down 17%. In a risk-off move, the yield on 10-year US Treasuries dropped 8 basis points. While the intraday moves were large, it is also important to keep them in context; both the S&P 500 and the Nasdaq Composite are up year-to-date. At the time of writing, Nasdaq futures are up 0.5%, while S&P 500 futures are 0.3% higher.
What do we think?
The market reaction reflects concerns about potential AI price wars, the sustainability of capex at major US tech firms, and the challenges of navigating a shifting investment across the enabling, intelligence, and application layers of the AI ecosystem.
Without taking any single-name views, at this stage we would make the following observations:
Low-cost models are positive for AI adoption. A lower-cost model could help speed AI adoption, which in turn increases the potential demand for services in the intelligence and applications layer (those using AI), as well as potentially enhancing productivity gains for the wider economy. It may even lead to higher demand for AI infrastructure, even if it is less compute-intensive. Because of lower cost, the lower demand for AI infrastructure compute intensity could be more than offset by wider adoption.
Experience from the mobile phone industry shows that the introduction of more cost-effective smartphones has led to wider adoption globally, while industry leaders have remained dominant in the high-end segment.
Continued capex is required to produce innovative AI models as the technology advances. The human brain, the inspiration behind the development of LLMs, remains computationally far more efficient than LLMs for comparable tasks. While advances in LLMs have been impressive, especially OpenAI’s o3 model released in December, which shows significant performance improvements versus major benchmarks, we have not yet reached a state of artificial general intelligence (AGI). As long as AGI has not been reached, cutting-edge frontier AI labs (those producing the most advanced models pushing the boundaries of AI capabilities) will continue to spend on compute to advance LLM performance. At the same time, we expect that state-of-the-art models will be continuously approximated by more efficient algorithms, as we have seen in the course of the last year.
Big tech's AI investments are also not limited to developing text-only models: They include development of multi-modal models including the generation of audio and video, and integration across applications.
On Friday, even after the release of DeepSeek's model, Meta indicated it plans to spend as much as USD 65bn this year to expand its AI infrastructure, almost USD 10bn above market expectations. While more details remain to be seen, we believe this suggests capex commitments from leading US tech companies are intact.
Where value accrues in the AI chain will shift over time. Of course, developing technology and evolving competitive dynamics in different parts of the value chain will mean that investors may need to adjust their focus between the enabling, intelligence, and application layers over time.
Analogous to the development of the internet, we have long thought that value creation would shift over time from those developing the infrastructure (the enabling layer) toward those using it (the application layer and beneficiaries). At the same time, we caution against jumping to firm conclusions about the impact of DeepSeek on industry dynamics at this early stage.
How do we invest?
AI is here to stay, and if anything, DeepSeek reinforces that. However, the latest developments do also show that investment approaches that are too concentrated or overly passive can be risky, as value can quickly shift within the AI ecosystem. An active and diversified approach is a better way to gain exposure to AI, in our view. We recommend investing into our Transformational Innovation Opportunity (TRIO) themes of AI and Power and resources, as well as the US equity market.
Further clarity on DeepSeek's impact, big tech's capex plans, and AI monetization should emerge in the fourth-quarter earnings season in the coming weeks. For now, our first conclusion is that lower computational costs should lead to greater demand for AI that will increase the aggregate spend on compute. Investors can access this opportunity through our AI TRIO. We also remain constructive on China’s tech sector, which could be a particular beneficiary of DeepSeek’s open-source model, given the increased trade and technological barriers between the US and China.
During the past decade, there has been at least one 10% valuation reset in global tech every year (except in 2017, which experienced a strong bull market). Barring the 30% or so reset in 2022 (which was driven by rising rates and the Russia-Ukraine war), over the past decade tech indexes have rebounded strongly over the subsequent 12 months after a 10% valuation reset.
Second, we also continue to believe that wider use of AI may further reinforce trends toward higher electricity consumption, which is supportive of companies exposed to power and resources. As noted above, we expect large frontier labs to continue to spend on energy-intensive training of models, even if there are models available that are more focused on less energy-intensive inference. Beyond AI, our Power and resources TRIO is further supported by the trend toward increased electrification globally and the anticipated power demands from innovation relating to quantum computing, climate, and blockchain among others.
Third, we expect the greater efficiency from new lower cost algorithms to lead to increased economic productivity, which is supportive of the broader equity market. In addition to these potential productivity gains we believe the combination of resilient US economic activity, solid earnings growth, lower borrowing costs, and the potential for greater capital market activity will lead stocks higher over the balance of 2025. In our base case, we see the S&P 500 reaching 6,600 by the end of the year and believe it could even hit 7,000 in the event of a strong growth scenario.
We recommend investors closely monitor upcoming tech results, take advantage of any extreme volatility (including through structured strategies), and build positions for longer-term growth in the AI ecosystem ("AI" portfolio), power and resources ("Power and resources" portfolio), and equity markets more broadly.
