DeepSeek, a Chinese company founded just over a year ago, recently released a new, open-source large language model (LLM) that uses less power and costs less to run than many current models, and delivers strong performance across a variety of benchmarks.
DeepSeek’s R1 model is notable both because it is open source, meaning that others can access and modify its underlying code, and because it is a highly competitive model developed outside the United States. However, it is less capable than current frontier models and not as efficient as some recent lighter-weight offerings. Rather than a paradigm shift, we see DeepSeek and future lighter-weight models as a natural evolution of a trend we have been tracking closely: LLM and non-language generative AI (GenAI) models have been getting more efficient and less costly and will likely continue to do so.
But what does that mean for those investing in technology businesses? What about “picks-and-shovels” investors who anticipated that an AI boom would trigger a significant increase in demand for both data centers and power infrastructure? Here’s our view.
How Does DeepSeek Influence Our Private Equity Team’s View of the Technology Landscape?
We are confident that GenAI will keep getting less expensive, model competition will continue to grow, and there will be an increasingly wide selection of open-source options. As lower-cost models emerge, we see an exciting opportunity to put the productivity-enhancing power of GenAI in the hands of more portfolio companies.
Over the last few years, our private equity portfolio companies have made their businesses more effective and efficient by integrating artificial intelligence across customer support, engineering, sales, and other areas. Several of our portfolio companies have harnessed GenAI to develop products that better serve customers. OutSystems, which helps companies more efficiently build enterprise-grade, mission-critical software and improve developer productivity, created a GenAI-based technology that allows developers to build the first iteration of an app entirely from conversational prompts, rather than coding. Development that used to take 3-6 months now occurs in a matter of minutes.
As tools become more accessible, we expect more companies will be able to automate tasks, gain insights from data, create new products and services, and provide more value to customers. We also think GenAI has the potential to make employees’ jobs more meaningful by reducing or speeding up rote, low-value tasks.
DeepSeek and other low-cost and open-source GenAI models pose a risk to companies that build and train general GenAI models, making their work more commoditized. Our private equity teams have not to date invested in such companies, but we have been acutely conscious of the disruptive potential GenAI poses to business models in a variety of industries since the launch of ChatGPT made sophisticated generative AI tools widely available several years ago. When our private equity teams look at potential investments, they incorporate a detailed analysis of both the potential disruptive threats and potential opportunities GenAI creates for a particular company’s business model.
Will Demand for Data Centers and Power Continue to Drive Infrastructure Investment?
The amount of data generated each year has been increasing for decades, and along with it, the need for the computing (processing power and memory), storage, and networking capabilities housed in data centers. The global shift to cloud computing has been a particularly strong driver.
We have been investing heavily behind data centers since 2019, with a focus on “cloud regions,” or locations with clusters of data centers, where demand is rising quickly. These regions are close to population centers and therefore offer low latency (the time it takes for data to travel between the user and the data center) and proximity to key customer networks and existing collections of data, with high barriers to entry due to a scarcity of available land and power. To date, we have executed on this thesis with investments in five data center platforms across the globe comprised of more than 110 facilities and $190 billion in total enterprise value, including a development pipeline representing dozens of gigawatts.
The overall demand for data centers has accelerated with the evolution of GenAI, which requires vast amounts of computing power for training models and “inferencing,” or responding to prompts. Cloud regions have absorbed much of this demand, though constraints on power have led some customers and developers to seek out non-traditional markets for training AI models, for which latency is less important.
Developing more efficient systems, as DeepSeek did, has happened throughout the history of computing. Importantly, however, aggregate demand for computing, storage, and networking has outpaced efficiency gains, leading to long-term growth in the volume of required data center infrastructure. Over the last five years, data center capacity leases have increased 5-7.5 times faster than the historical trendline would have predicted, and 2024 was the largest year on record.1
The technology environment is changing in ways we believe will continue to benefit cloud-region and close-proximity data centers, which is where we continue to focus. We expect the ratio of computing power needed for training (where proximity doesn’t matter as much) vs. inferencing (where proximity matters a lot) to flip from 60/40 to 20/80 in the next 3-5 years as adoption of AI increases and new reasoning models that can “think” and draw conclusions evolve. As Nvidia noted on its Feb. 26 earnings call, reasoning models need 100 times as much computing resources as traditional models.
Finally, in building new data centers, we take a conservative approach by focusing on the highest-quality markets and only expending capital when a long-term customer contract is secured.
We also expect to continue investing in power infrastructure amid broad-based growth in electric load not only from data centers, but from the energy transition and the reshoring of manufacturing. Overall electricity demand in the United States is expected to grow 2.4% a year, up from essentially zero in the last 20 years, according to our Global Macro & Asset Allocation team. In addition to additional power generation capacity, infrastructure capital is needed to modernize, upgrade, and replace transmission and distribution grids, some of which are over 40 years old in the United States.
Our most recent power investment, a stake in a portfolio of electric transmission companies in the American Midwest, was attractive for a number of reasons, including a transparent revenue framework underpinned by federal regulation and partnership with a market-leading utility. One of the most important reasons, however, is that it offered a diversified load growth profile across numerous end markets, including data centers and industrial and manufacturing activity.
Investing Wisely Today to Stay Ahead of Tomorrow
DeepSeek’s model surprised much of the market, but we have expected the cost and power required for GenAI to decline over time and invested accordingly. We continue to believe that GenAI will create attractive opportunities in both private equity and infrastructure, but we invest with a clear-eyed view about the risks of disruption, the sources of potential opportunity, and a critical eye toward overly optimistic projections about future demand.
