In the wake of DeepSeek's release, it may be efficiency gains and lower costs that drive developments in the AI market.
More than two years after the world witnessed the launch of ChatGPT, artificial intelligence (AI) technology continues to evolve at a rapid pace, fuelled by substantial investor interest and funding for development and infrastructure. While AI is a general-purpose technology that looks set to transform the economy and society in the long term, the road will be bumpy as innovation re-shapes the competitive landscape and shifts the beneficiaries over time.
We recently proposed a framework for investors to use when considering the development of new technologies and how to adjust their portfolios to take advantage of these. The LGT PEARL framework has five categories of companies that should benefit from the different stages of new technology development:
At the time we identified the kinds of companies that are already operating in the first three categories in the AI marketplace. The Pioneers include developers of AI models; the Enablers are semiconductor companies and data centre suppliers and operators, among others; and the Accelerators are software companies and IT consulting firms. Because the AI business is still so new, the Reformers and Laggards are yet to be identified.
There are risks associated with investing in the AI theme. The primary concern is that innovative techniques could reduce the need for computing power and the associated data centre infrastructure. At the same time, however, efficiency gains should drive down the cost of the technology, making it available to a wider audience, thus accelerating adoption and driving additional innovation.
In keeping with Jevons Paradox, computing power requirements are likely to remain high, though increasingly shift from model development to real-world application. We expect Enablers like semiconductor companies to benefit from this; we also expect a shift in investor interest to companies in the Accelerator category, such as software firms.
In January 2025, the Chinese start-up DeepSeek made headlines by releasing a reasoning model with similar capabilities to OpenAI's o1 model, but reportedly at a fraction of the cost of development. While the reported outlay of just USD 5.6 million may not actually include all the development expenses, DeepSeek's performance has been impressive, earning endorsements from major US tech leaders.
This achievement raises questions about the necessity for large infrastructure investments to develop and refine these types of applications, and the effectiveness of US export controls to prevent China's AI progress. In addition, it makes it difficult to assess the monetisation potential of AI models and related activities.
Prior to the release of DeepSeek, it was widely agreed that training large-scale AI models is computationally intensive and requires the most advanced IT infrastructure, typically housed in data centres. As a result, companies such as the hyperscalers invested heavily in this infrastructure. While not all of this infrastructure is dedicated solely to AI, the development of these models is widely acknowledged as the driving force behind this investment.
DeepSeek puts AI infrastructure spending under the microscope.
The question now is: How much of this infrastructure will really be needed? That said, no company appears to be cutting back on this capital expenditure as yet. We believe that the arrival of DeepSeek will put these investments under greater scrutiny and markets will demand more proof of return on investment to justify such large expenditures.
While the market is concerned that innovative techniques like DeepSeek may reduce the need for computing power and thus data centres, the opposite may ultimately be true. As the technology becomes more efficient and cost effective, demand may shift from pre-training AI models before launch to post-training and inference workloads that improve the models through advanced reasoning techniques, but also require a lot of computing power.
Innovative AI techniques like DeepSeek may drive down costs, which with the growing adoption of open-source models is likely to democratise the use of AI technology. Accelerator companies will benefit from this affordability by integrating AI into their core offerings and building applications that make AI available to more users.
Many companies don't have the internal expertise or financial resources to develop AI applications in-house. However they already use external software solutions in areas such as CRM (customer relationship management) or ERP (enterprise resource planning).
So software companies integrate AI into their offering. Examples include AI assistants that support employees and AI agents that perform tasks autonomously. Companies can use these software solutions to drive productivity gains and leverage their valuable data.
While software vendors have successfully launched AI solutions that resonate well with customers, they have struggled to find optimal pricing models. As the underlying technology becomes more affordable, more value may trickle down to the application layer.
This shift could enable companies to offer more competitive and flexible pricing strategies, which in turn would make AI solutions accessible to a wider range of businesses. As a result, customers may experience faster AI adoption and a more compelling return on investment, as reduced costs and improved pricing models make it easier to integrate AI into their operations.
As for our PEARL framework, we continue find Enabler companies like semiconductor manufacturers attractive, which are essential to the production of computing power. However, as efficiency gains speed up the adoption of AI, we expect a shift in investor interest to Accelerator companies, such as software companies.
LGT's experts analyze global economic and market trends on an ongoing basis. Our publications on international financial markets, sectors and companies help you make informed investment decisions.
