AI’s next bottleneck isn’t chips — It’s power!
Data center capacity is expanding at an unprecedented pace, driven by the rapid deployment of artificial intelligence infrastructure. Global data center capacity increased from approximately 36 GW at the end of 2022 to 42 GW in 2023, rose further to 51 GW in 2024, and is projected to reach 62 GW in 2025.
This acceleration reflects a structural shift in capital allocation following the launch of generative AI applications, as hyperscalers such as Microsoft, Amazon, Google, and Meta Platforms transition from general-purpose cloud infrastructure to AI-optimized compute clusters.
The expansion is heavily concentrated in the United States, where data center capacity is expected to nearly double between 2022 and 2025. As of late 2025, the U.S. accounts for approximately 73% of the global data center construction pipeline, significantly outpacing Asia and Europe.
This concentration is driving a corresponding surge in electricity demand, with AI workloads requiring significantly higher power densities than traditional cloud infrastructure. Rack-level power, which historically ranged from 10–30 kW, is now scaling toward 100 kW, 200 kW, and in some cases beyond 500 kW, with next-generation systems targeting megawatt-class configurations.
At the same time, power infrastructure investment is accelerating in parallel. In the United States, energy storage system deployments exceeded 20 GW of new capacity over the 2024–2025 period, while new generation capacity continues to expand to meet rising demand.
However, despite this rapid buildout, these investments are primarily focused on grid-level supply and long-duration storage. They do not address the most critical requirement of AI data centers: the ability to deliver stable, uninterrupted power at the point of compute on millisecond time scales.
In this article, I focus on two power companies, Vistra (VRT) and NRG Energy (NRG) that I believe are positioned to benefit significantly from the growth of AI-driven data centers. The drivers are not immediately obvious, and understanding them requires a closer look at how power is delivered and managed within these systems.
For readers of my work, this is by design. My analysis is built on deep familiarity with specific technologies and markets, rather than the broad, rotating coverage typical of platforms where contributors write across unrelated sectors without developing true domain expertise.
Data center power architecture
According to Chart 1, a data center’s power architecture is structured as a sequence of tightly integrated layers that progressively refine electricity before it reaches the compute load. Power enters from the grid, passes through transfer and redundancy systems, and then flows into the uninterruptible power supply (UPS), where it is conditioned and stabilized before being distributed to racks and consumed by AI processors.
The critical insight is that energy storage is not deployed as a standalone system. It is embedded within the UPS layer, which sits directly between the grid and the IT load. This positioning makes it the only part of the architecture capable of responding instantaneously to disturbances, ensuring uninterrupted operation.
Power is delivered across multiple layers, but continuity is established at a single point within the system.
According to the table, the defining characteristic of the data center power system is not its physical structure, but the speed at which each function must respond to maintain system stability. Grid supply operates on time scales measured in minutes to hours and is designed to provide bulk electricity rather than precision.
Transfer systems improve reliability by switching between sources in seconds, but they still cannot prevent short-duration interruptions.
The critical distinction emerges at the UPS layer. Unlike all other segments, the uninterruptible power supply operates on millisecond time scales and incorporates embedded energy storage, typically in the form of lithium-ion batteries. This allows it to deliver conditioned power with zero interruption, bridging disturbances instantaneously while upstream systems respond.
Downstream, power distribution and rack-level conversion maintain high stability but do not provide energy buffering, instead relying on the UPS to supply continuous input. At the final stage, AI chips operate on nanosecond time scales and require absolute power continuity, making them entirely dependent on upstream infrastructure.
— Dr. Robert Castellano, Semiconductor Deep Dive Newsletter, USA.
