Time to Think Differently

“When the facts change, I change my mind. What do you do sir?” –John Maynard Keynes  

Introduction  

The introduction of DeepSeek¹ has catalyzed significant shifts in the artificial intelligence (AI) sector, challenging existing assumptions and reshaping the investment landscape. This transformation has prompted investors to reconsider the future of AI dominance, accessibility and the strategic deployment of capital.

Rethinking AI investment opportunities 

Investments in AI are heavily influenced by advancements in computing power, algorithmic efficiency and data utilization. These core components of the AI value chain, the “picks and shovels,” have been the primary beneficiaries of investor interest. Investment has concentrated heavily in a select group of companies, often overlooking the diverse potential outcomes for these stocks. DeepSeek’s innovative algorithms and open-source approach are redirecting focus from traditional hardware enhancements to more nuanced areas like algorithm optimization and data exploitation. While algorithm improvement has been continuous, DeepSeek employs novel techniques. This evolution not only accelerates technological advancements but also diversifies investment opportunities beyond hardware-centric firms and data center beneficiaries. 

Financial implications of commoditized AI models

The narrowing gap between leading AI models and their fast followers, driven by open-source contributions and model distillation, is commoditizing foundational AI layers. While we have held this view for some time, the trend has accelerated and it’s no longer those with the deepest pockets that necessarily win the race for the best models. This shifts the value proposition towards application-specific AI, benefiting sectors like SaaS and prompting a potential wave of innovative AI applications. For investors, this may suggest a potential shift in focus from investing in underlying AI technologies to targeting companies that excel in applying these technologies across various industries. We think that those that are best positioned will combine unique data with AI features and customer workflows.

This is also positive for IT service providers as enterprise adoption accelerates the need for support in implementing AI increases. Lastly, the need for security continues to rise, due to increased points of attack and free access to malware and phishing schemes. Overall, we believe these developments provide significant scope for software. 

Data, data and more data 

They say the key to real estate is location, location, location. For AI it is data. When it comes to training, Large Language Models (LLM’s) look to be reaching their scaling limits as performance improvements slow. As such, those who have proprietary data sets continue to have an advantage, while those who aggregate public data may be negatively impacted. This also supports our view that selective data platforms and service providers stand to benefit as corporates move to make sure of their data readiness in support of new AI-driven applications. 

Persistent demand for compute and investment in AI infrastructure  

Persistent questions about AI capital expenditures and the revenues needed to justify their ROI have become more pronounced with recent developments. The continuous need for advanced hardware to manage complex AI tasks ensures ongoing investment in AI-related infrastructure. Although we are transitioning from a training-intensive phase to one dominated by inference, which has seen rapidly declining costs, current compute needs show no signs of overcapacity. However, technological advancements, such as Nvidia’s new Blackwell GPU — which is four times more powerful and less energy-intensive than current models — promise significant improvements in computing power and efficiency.

In our view, the quest for Artificial General Intelligence and cutting-edge LLMs will likely sustain demand for infrastructure and high-end compute. However, an AI model designed for health care to analyze all proteins in the world is significantly smaller than an LLM that processes vast data volumes. We believe we are moving to a world of smaller models and ongoing demand for compute, but as long-term investors, we are cautious about robust growth projections for 2027 and beyond influencing valuations.

We feel hyperscalers are well positioned despite potential short-term constraints such as software, bureaucracy or corporate implementation delays. While capital commitments for 2025 are expected to remain high, the outlook becomes less certain further out. The need for capital expenditures will likely continue to rise, but the intensity of these investments might decrease, as indicated by Microsoft’s recent pause in spending on long-lived assets. This trend to us suggests declining capital intensity, which is favorable for hyperscalers but may impact other data center ecosystem players like those in power and cooling.

The trend towards smaller models could hasten the shift to edge computing, where these models operate on personal devices like phones or laptops, instead of servers. We remain vigilant about a potential refresh cycle that could embed more powerful chips in consumer electronics, anticipating that “AI-capable” devices might command premium pricing. 

Strategic implications  

In our view, companies integrating AI effectively to enhance operational efficiencies and customer experiences have the potential to outperform. Productivity improvements have the potential to be broadly positive for margins where companies have pricing power and can hold onto these gains. Combined with falling AI costs, this could also present as a deflationary impulse. 

Additionally, geopolitical shifts in technology access and regulation could influence global investment. We note that Mistral launched its new Small 3 model, which they claim is smaller, faster and cheaper. Do these developments provide an opportunity for European companies to catch up?

During the DeepSeek selloff, several stocks related to the “A1 data center trade” uniformly fell. This is puzzling, as they range from networking and connectors to power generators, electrical equipment providers and cooling systems. While linked, each has different exposures to data centers with different consequences, which begs a broader question on market structure and basket trades.

Conclusion

As DeepSeek redefines the boundaries and capabilities of AI, investors may need to consider adopting a dynamic approach to try and capitalize on this evolving market. We feel the focus should be on companies and sectors that not only develop AI technologies but also effectively implement them to drive business value and competitive advantages. Navigating this shifting landscape will require a keen understanding of both technological trends and their broader economic and geopolitical impacts.

Hyperscalers remain well placed, software stands to benefit while deeper introspection is required over the duration of growth rates. Change in this space will not abate. As active managers, we will try to take advantage of the market action, remain nimble and position for AI going forward rather than remain exposed to what has worked for the last two years. 

¹ DeepSeek is a relatively new Chinese based company building AI models. 

The views expressed herein are those of the MFS Strategy and Insights Group within the MFS distribution unit and may differ from those of MFS portfolio managers and research analysts. These views are subject to change at any time and should not be construed as the Advisor’s investment advice, as securities recommendations, or as an indication of trading intent on behalf of MFS. No forecasts can be guaranteed.

Diversification does not guarantee a profi t or protect against a loss. Past performance is no guarantee of future results.