Engineered integration between battery management systems (BMS) and uninterruptible power supply (UPS) technology is shifting data center operations. Behind it: accurate state of charge (SOC) and state of health (SOH) reporting powers input power smoothing (IPS), protecting AI data center uptime, battery lifecycle, and total cost of ownership (TCO).
The question facing data center operators today is no longer whether AI workloads are difficult to manage (they are). The real question is whether the infrastructure supporting them is built to respond, not react.
Modern AI workloads — from large language model (LLM) training to real-time inference — generate rapid, large-burst power demand patterns that cycle between idle and peak draw in fractions of a second, propagating upstream and stressing generators, transformers, switchgear, and utility connections. For years, operators have relied on imperfect workarounds: oversizing infrastructure, capping graphics processing unit (GPU) power, or managing instability reactively. These approaches reduce exposure but don't eliminate it, and come at the cost of higher capital expenditure, reduced compute utilization, and longer job completion times.
Managing consequences ≠ prevention
GPU clusters don't ramp up gradually. They step — hard and fast — jumping between idle and peak power in milliseconds, repeatedly and across the duration of a workload. Every one of those transitions sends a ripple through the power train. Without a mechanism to absorb those ripples, they reach generators, transformers, and utility connections with full force.
The traditional response has been to build bigger: larger generators sized for worst-case peaks, more switchgear, more headroom. Some facilities go further, imposing software-based power caps on GPU clusters to artificially limit peak draw.
The result? Compute utilization drops. Job completion times increase. Capital expenditure climbs. But the underlying volatility problem remains unsolved.
Input power smoothing (IPS): A smarter architecture
Input power smoothing (IPS) takes a fundamentally different approach. Rather than sizing infrastructure around worst-case peaks, IPS uses integrated lithium battery cabinets working in tight coordination with Vertiv power converters to act as an active energy buffer between AI loads and upstream power sources.
Here's how it works: whenever the AI load fluctuates, regardless of magnitude, the UPS charges and discharges the battery to compensate, so upstream sources see a stable, controlled demand profile. The battery acts as a short-duration energy buffer, absorbing the volatility before it reaches the grid. Smoothing operates within defined SOC thresholds, so reserve capacity for backup is always protected. The result is a power train that responds to AI workloads without being destabilized by them.
But IPS is only as reliable as the data driving it. And that's where battery state precision becomes mission-critical.
State of charge: The metric that governs every decision
State of charge (SOC) represents the percentage of usable energy currently available in a battery; the battery equivalent of a fuel gauge. In a conventional backup role, SOC is a status indicator. In an IPS architecture, it is something far more consequential: an active input to UPS control logic that governs when the battery charges, discharges, and holds.
When SOC is accurate, the uninterruptible power supply (UPS) makes the right call, every time. It knows when to engage the battery, when to protect reserve capacity, and when to taper smoothing activity as thresholds are approached.
When SOC is inaccurate, the consequences compound quickly. If the actual charge is lower than what the battery management system (BMS) reports, the UPS may continue smoothing and over-discharge the battery. This consumes the reserve capacity needed for an actual outage. If the actual charge is higher than reported, the UPS may prematurely stop smoothing, exposing upstream infrastructure to the very load swings IPS was designed to prevent. In either case, the system behaves unpredictably, and unpredictability in a power train is a liability.
State of health: The metric that protects your investment
State of health (SOH) measures a battery's actual usable capacity relative to its original nameplate rating; a measure of how much the battery has aged. A battery at 100% SOH delivers its full rated energy. As batteries degrade through cycling, SOH declines, and both available runtime and smoothing capacity shrink.
In AI power smoothing applications, batteries cycle far more frequently than in conventional backup roles. That makes accurate SOH tracking essential for understanding what the system can deliver today, and for planning what it will deliver tomorrow.
Accurate SOH enables operators to move from reactive maintenance to predictive maintenance. Instead of replacing batteries based on uncertainty or discovering degradation only when it becomes operationally visible, facilities can plan service actions around actual asset condition. The financial impact is real: batteries used fully and safely throughout their intended lifecycle reduce the total cost of ownership (TCO), while inaccurate SOH readings lead to premature replacement or over-cycling, both of which increase cost and risk.
Integration is what makes it work
Precise SOC and SOH reporting doesn't happen by accident. The BMS uses battery- and cell-model-specific algorithms to calculate both metrics with the accuracy and real-time responsiveness that the UPS cannot replicate on its own. High-speed, accurate communication between the BMS and UPS enables control algorithms to respond correctly under millisecond-level AI load changes.
For operators running AI workloads at scale, the compounding benefits of IPS with accurate SOC and SOH reporting extend well beyond individual events. In the near term, IPS reduces the need to oversize upstream infrastructure. In the medium term, accurate SOH tracking extends battery service life. Over the long term, facilities that manage power volatility at the battery-UPS layer gain the flexibility to scale AI compute capacity without proportional infrastructure investment.
The shift from managing consequences to preventing them starts with knowing exactly what your battery can do — and building a system precise enough to act on it.
*In compliance with the EU AI Act’s transparency requirements, this article included the use of AI during content organization and refinement. Writers and technical subject matter experts (SMEs) further reviewed, refined, and finalized the published version.
