Darren H. S. Tan. Chief Executive Officer at UNIGRID.
The rapid expansion of AI data centers has sparked a growing conversation about electricity demand. Analysts have warned that training and running LLMs will require enormous amounts of power, prompting utilities and infrastructure providers to race toward new generation capacity. But the biggest infrastructure challenge may not be how much total electricity AI consumes but how that electricity is consumed.
Unlike many traditional enterprise workloads, AI training and inference can create sudden, high-intensity swings in power demand. As organizations build the next generation of AI infrastructure, they may need to rethink not only how they generate and store electricity but how they manage sudden surges in power demand.
Creating A Different Kind Of Power Problem
Data centers have long relied on battery systems to provide backup power during outages and help manage energy use alongside the electrical grid. AI is expanding that role by creating workloads that require energy storage to respond to rapid changes in power demand, not just emergency events.
Model training creates massive power demands that ramp up almost instantly and drop just as quickly when training ends. Inference introduces a different challenge, with demand fluctuating continuously as query volumes and utilization rise and fall throughout the day. As a result, the real challenge for AI data centers isn’t simply obtaining sufficient electrical capacity but also designing storage infrastructure that can adapt to big swings in power demand happening in minutes and seconds instead of hours.
Looking Beyond Energy Capacity
Historically, buyers have compared battery systems largely on energy capacity and cost per unit of stored energy (such as $/kWh). While these remain important metrics for many applications, AI workloads are shifting attention toward a battery’s power-to-energy ratio, or the amount of instantaneous power it can deliver relative to the amount of energy it stores (such as $/kW and kW/kWh).
For lithium-ion systems not specifically designed for repeated high-rate operation, frequent high-power cycling can generate more heat and accelerate battery degradation. These systems are energy-dense but power-constrained. To compensate, AI data center operators often oversize battery systems (i.e., installing more energy capacity than the application actually requires so each battery operates under less demanding conditions while still delivering the necessary power output).
The trade-off is higher costs, a larger physical footprint and excess energy capacity that may never be fully utilized. Consider water storage as a point of comparison. Lithium-ion batteries have a large tank (high energy) but a small tap to access the water (lower power). Sodium-ion batteries have a smaller tank but a large tap to access the water quickly. Data centers need access to water quickly and frequently. This could either be achieved by having more lithium-ion batteries with numerous smaller taps, or one single sodium-ion battery with a large tap.
Matching Battery Technologies To Different Workloads
No battery chemistry is universally the right fit for every application. Lithium-ion batteries have become the industry standard because their high energy density and mature manufacturing ecosystem have made them well-suited for everything from electric vehicles to consumer electronics and long-duration energy storage.
But alternative battery chemistries might make more sense for supporting AI data centers. For example, certain sodium-ion chemistries are being engineered to support high-power, short-duration cycling. Additionally, they perform better at extreme temperatures (hot and cold), which can eliminate the risk of thermal runaway needed with lithium-ion, which requires massive, expensive active cooling systems just to keep the batteries stable.
Battery Selection: Becoming A Facility Design Decision
For data center operators, battery selection is no longer just an equipment decision. It is increasingly a facility design decision.
Every square foot allocated to batteries, cooling systems, fire protection or related infrastructure is space that cannot be used for additional compute capacity. As AI workloads increase both power density and demand for backup energy, the physical requirements of battery systems are becoming more closely tied to data center economics.
Lithium-ion remains a proven and widely used option, but some deployments require dedicated cooling, fire-suppression systems, physical barriers or outdoor placement. Those requirements can influence where batteries are installed and how much supporting infrastructure a facility must accommodate.
Alternative chemistries may create new design options. Certain sodium-ion systems, for example, are engineered to maintain stable thermal performance during high-power operation, which may reduce cooling or safety-infrastructure requirements depending on the application.
That flexibility can allow storage to be located closer to the electrical load. In turn, operators may be able to reduce cabling, simplify parts of the power architecture and preserve more floor space for revenue-generating compute.
This does not mean lithium-ion is becoming obsolete. The more important shift is that AI infrastructure is expanding the range of battery requirements. No single chemistry will be ideal for every environment.
As data centers become larger, denser and more power-intensive, battery selection will increasingly shape facility layout, operating costs, safety strategies and long-term scalability. The question is no longer simply which battery performs best but which battery is best suited to the demands of the infrastructure it supports.
Energy Storage: Becoming Active Infrastructure
Battery systems once viewed primarily as emergency backup are becoming active participants in managing increasingly dynamic AI workloads. As new data centers are designed, operators should evaluate energy storage through a broader lens than backup runtime alone, considering factors like:
Power-To-Energy Ratio: How much energy can a battery deliver instantaneously to power AI workloads’ demand without requiring excess energy capacity?
Thermal Characteristics: How much cooling infrastructure is required, and how does it affect facility design, operating costs and battery placement?
Cycle Life: Can the battery withstand tens of thousands of rapid, short-duration charge and discharge cycles without significant degradation?
Rather than searching for a one-size-fits-all battery, AI data center operators will be better served by matching storage technologies to the operational demands of their workload. As AI continues to reshape the grid, the winning storage technologies won’t simply store the most energy—they’ll deliver the right power, at the right time, most reliably and efficiently.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
