Managing power volatility in AI data centers: The critical role of energy storage

By Isaac Chan, Spencer Gore, Daniel Kubis, Benjamin Lowe and Kyle Gordon

As AI data centers scale, extreme and volatile power demands are redefining how facilities manage reliability and efficiency. Battery energy storage is emerging as a critical solution – particularly for AI training workloads – helping stabilize power systems and enable resilient, future‑ready data center designs.

The rapid expansion of AI data centers is reshaping the power generation equipment landscape. Often publicized are the capacity expansions of turbine or reciprocating engine OEMs looking to supply data center operators’ seemingly insatiable appetite for power. Sometimes lost in these headlines is the importance of this segment to the battery sector.

Among stationary storge applications, data center has emerged as the second largest stationary market for batteries globally. At Roland Berger, we expect that data center demand for batteries will grow from ~20 GWh globally today to over ~70 GWh of demand by 2030. This ~26% growth rate is among the highest of battery segments and is the reason why this has become an important growth vector (and margin drivers) for players in the battery value chain looking to wait out electric vehicle (EV) winter .

Energy storage use cases driven by data center power challenges

Once limited to the role of acting as the bridging source of power supply between the grid and backup generation sources (e.g., diesel reciprocating engines) during grid outages, the use cases for batteries at data centers is growing. Now batteries are used to increase time-to-power by firming readily available renewable power generation sources (e.g., solar, wind). Data center operators are also using batteries to reduce their carbon footprint, for example, by replacing diesel-fired back up gensets with batteries (“backup” use case).

The most interesting, and least understood use case, for batteries is the role that they play in smoothing fluctuations in load at AI training data centers. As AI algorithm training grows and supply constraints force data center operators to explore on-site generation sources for power, the challenges associated with managing the workloads at these data centers and the dangers that their fluctuations in load cause if unchecked are of paramount importance to data center operators, grid agencies and equipment suppliers.

The new power paradigm: AI training and data center infrastructure

AI training workloads present a fundamentally different power profile compared with traditional data center operations, resulting in two key challenges: harmoic frequencies and ramp up challenges.

Harmonic frequencies: The synchronized nature of deep learning tasks creates sharp and frequent fluctuations in power demand as hundreds of thousands of GPUs “flip” from compute runs to communcation intervals all at the same time. These rapid swings in demand can reach hundreds or even thousands of megawatts within milliseconds and introduce real operational risks. These swings in load create harmonic frequencies that ressonate at the natural frequencies of power generation equipment. If left unchecked they can accelerate mechanical fatigue in generators and turbines and create voltage instability and frequency regulation challenges for utilities .

Ramp up challenges: Conventional power architectures were designed for relatively stable and predictable loads. As a result, they struggle to accommodate the dynamic requirements of AI training. The rapid ramp up at the start of training runs often exceeds the response capabilities of on site generation sources. Unaided, on-site generation equipment will leverage mechanical energy (or torque) to match the power demands of AI training loads. This can lead to mechanical stress and accelerated failure of power generation equipment, however, its impact varies by power generaton setup.

It should be noted, although ramp up challenges are more acute for data centers with prime power sourced from on-site power generaiton equipment, these ramp up challenges become an issue for any data center that uses turbomachinery (i.e., diesel gensets) for backup power during a grid outage.

Energy storage as a strategic solution

There are many different potential solutions for transient load mangement at AI data centers. Software based solutions and adjustments at the GPU level can provide partial mitigation by introducing secondary workloads or controlling ramp rates. However, these approaches often come at the cost of reduced efficiency or increased hardware stress. Energy storage technologies, including lithium-ion batteries and supercapacitors, offer a more comprehensive and sustainable solution, meeting strict operational requirements without unnecessary energy loss, however, at increased costs.

In the end, a combination of software, firmware and hardware (i.e., energy storage) solutions will be required to effecitvely and efficiently manage transient loads at AI data centers.

Tailoring battery strategies to data center needs

There is no single blueprint for deploying energy storage in AI data centers. Requirements vary widely depending on the type of compute, ranging from traditional workloads to AI inference and large scale training, as well as the characteristics of the power supply (i.e.,, grid connected vs. non-grid connected, source of on-site prime power generation). The frequency and intensity of training cycles, checkpoint behavior, and backup system design all shape the optimal storage configuration.

The role of Roland Berger

To navigate this complexity, Roland Berger has developed proprietary HYPERSPACE models, which simulates how AI workloads interact with different power generation and storage setups. By capturing real world operating conditions, including grid instability, equipment stress, and workload driven demand patterns, the model enables a rigorous assessment of both technical performance and economic outcomes across energy storage strategies.

This analytical foundation allows operators and investors to move beyond simplified assumptions and make informed decisions on system design. It supports the identification of the appropriate storage scale, the evaluation of trade offs across reliability and cost, and the design of resilient power architectures tailored to specific operating environments.

Building on these insights, Roland Berger supports end-to-end decision making, from defining energy strategies and prioritizing investments to guiding implementation and operational optimization. This ensures that energy storage is positioned as a core enabler of reliable and efficient power delivery, aligned with the evolving demands of AI driven data centers.

Charting the path forward

The rise of AI data centers represents a defining moment for the battery industry. As the fastest growing segment within stationary energy storage, these facilities are setting new standards for performance, resilience, and operational complexity. Addressing these demands requires both technological innovation and a deep understanding of the relationship between computing, power systems, and storage solutions.