Powering the AI boom: deficits create investment opportunities in self-generation solutions

The AI boom has led to skyrocketing demand for new data centers across the US. However, US electric utilities simply lack the necessary generation or transmission capacity to power the planned data centers. This insufficient supply of electricity is the largest factor hindering the industry and was the dominant theme at the Data Center Dynamics (DCD) trade show recently held in New York. Robert Berner, Vontobel Quality Growth investigative analyst, attended this event that has brought data center operators and their equipment providers together for the past 25 years. 

Companies benefitting from power constraints  

More data center operators are considering self-generation to provide “bridge” power until they can connect to a utility (which can take years) or permanently power their centers altogether. In industry jargon, this solution is referred to as “BYOP” (bring your own power) or “behind the meter.” 

This could be a potential boon for generator manufacturers and providers of alternative power sources, such as hydrogen fuel cells and small nuclear reactors. Schneider Electric, a holding in Vontobel Quality Growth’s Global strategy, is benefitting from the complications navigating BYOP . The French-based energy management and industrial automation company has established a strong business that supplies a broad array of products to data centers, including transformers, switching gear, uninterruptible power supplies, cooling equipment and data center management software. 

Schneider Electric is now offering a turnkey solution where it will set up a complete behind-the-meter system for data centers, including other vendors’ generation equipment. It then works out a power purchase agreement for the data center, much as if it were buying electricity from a utility. It is ahead of its key rivals, Vertiv and Eaton, which have yet to launch a similar service. 

Vendors and data center operators view data center demand as buoyant due to contracts already procured for AI workloads. While large language models remain crucial to this demand, we believe inference will drive the next wave of AI compute in support of real-time AI applications. According to Schneider Electric, 20% of its sales are tied to data centers. We believe our investment in the company gives us a foothold in this growth story; however, if data center demand should falter, the remainder of Schneider’s power management and control products have wide applications across other end markets, including power utilities, commercial and residential buildings and factories. Schneider also has other growth drivers, such as grid modernization and digitalization, which gives us downside protection. We believe this is worth having because, as the stock market has recently shown, AI sentiment is unlikely to rise in a straight line.

As we explore power constraints more broadly, we invested in UK electric-transmission company National Grid. The UK is facing power shortages due to the growth of data centers, electric vehicles, and heat pumps, as well as aging grid infrastructure that can’t fully integrate renewable-power sources. In response, National Grid has embarked on a major capital expenditure plan through 2029. We expect the doubling of capex should accelerate the company’s earnings growth compared to the previous five-year period. The investment added a new growth driver to our International Equity strategy and strengthened its defensive characteristics at a time of increased market volatility.

Challenges of the power deficit: extended wait times and exacerbated demands from AI workloads

The time it takes for a new data center to receive power delivery has more than doubled since 2022, averaging approximately five years nationally. In crowded markets, such as Data Center Alley in Northeast Virginia, it can now take seven years or more to secure power. These wait times are expected to increase as data center electricity usage is projected to rise from 4% of the US grid today to 10% to 12% of the grid by 2030. 

The power density of AI workloads has exacerbated these demands. While traditional CPU server racks use about 20 kilowatts of electricity, the latest GPU servers use about 60 kilowatts. Nvidia’s next generation GPUs will use 600 kilowatts per rack, and the company has indicated that the 1-megawatt rack is not far off. Additionally, stronger cooling systems are required to cool GPUs, further increasing electricity demand. 

This has led to the construction of larger data centers, with the typical AI factory now using 200 megawatts of power, up from 30 megawatts just a few years ago. These load requirements make it even more challenging for data centers to secure power in a timely manner. Data center operators can seek more remote sites with available power, but it can still take years for utilities to establish transmission lines to deliver power to the facility. Often, utilities only provide data centers with a fraction of the power they need, with agreements to scale up to full load over several years. 

Data center operators and their customers, meanwhile, can’t afford to wait. After investing $30K or more per GPU, they are eager to put their capital to work. The competition among hyperscalers is also intense, with none wanting to fall behind due to power constraints. 

Seizing Opportunities

The AI boom is driving an unprecedented demand for data centers, leading to significant power supply challenges. We believe the extended wait times for power delivery, increased power density of AI workloads, and the need for stronger cooling systems are exacerbating these issues. However, in our opinion, these challenges are also opening up new investment opportunities in self-generation power solutions, with companies like Schneider Electric and National Grid leading the way. As the industry navigates these hurdles, innovative solutions and strategic investments may be key to sustaining the growth and evolution of AI and data center operations.