From Centralized UPS to Rack-Level BBU: Rethinking Power Continuity and Protection in AI Data Centers

• Author : Rolan Lee｜LITEON Cloud Infrastructure Platform & Solution

• Written & Interviewed by : LITEON Editorial Team (Corporate Brand Value Development Center)

• Technical review :LITEON Center of Core Competence

In an AI data center, power continuity is more than just an uptime metric. It is the lifeblood that keeps thousands of accelerators, memory states, checkpoints, and inference queues flowing together as one system. A small electrical disturbance during large-scale training can break a synchronized workload, leading to loss in the form of idle GPU fleets, corrupted or rolled-back checkpoints, delayed model delivery, and weeks of wasted compute expense. Not only that, but the same instability can break real-time service commitments during inference.

According to Uptime Institute's 2024 survey, 54% of respondents said the last major outage they experienced cost more than US$100,000, and 20% said it was more than US$1 million.

That is why backup power architecture is moving away from the old model of centralized UPS rooms protecting the facility as a whole to protection much closer to the load. Inside the high-density rack itself, where AI power behavior is the fastest, steepest, and least forgiving. With AI workloads driving power swings and peak demand, distributed rack level energy storage becomes the more proximal layer to ensure compute continuity. This is the moment when the Battery Backup Unit (BBU) is a critical building block of next generation AI infrastructure.

Architecture Transformation and Technical Definition: Why Is BBU the Best Line of Defense for AI Infrastructure?

From Facility Backup to Rack-Embedded Reserve

A Battery Backup Unit (BBU) is a distributed energy-storage module embedded within the server rack, positioned adjacent to the power shelf, and connected to the rack's DC busbar. It provides support to the compute payload at the electrical layer where GPU, accelerator, and high-current server loads draw power. This is important because AI infrastructure is not a stable IT load. The sharp transient demand from ultra-high power chips requires the protection layer be physically and electrically closer to the rack, not just in a remote electrical room.

Why DC-Coupled Backup Changes the Equation

Architecturally, the difference is the power path. A typical online UPS will rectify incoming AC to DC to charge the battery and then invert it back to AC for downstream distribution. Server PSUs will convert it again for IT electronics. In contrast, the rack BBU is DC-coupled, so that the backup energy can be directly injected into the rack power plane with fewer conversion steps, lower accumulated loss, and better energy utilization under high-density operation.

Two Minutes Is an Engineering Window, Not Just Runtime

For LITEON, the key design target is not "long backup" in the old facility sense but controlled continuity at the rack. Its BBU can keep power for 2 minutes, which gives the system enough time to complete critical data movement, protect volatile operations, or perform an orderly shutdown sequence before the event propagates into a compute failure. Within this framing, backup power is a precision buffer between electrical instability and loss of workload.

Power Protection Also Means Thermal Protection

This is the bit a lot of people miss. The compute stack is not the only load that needs support during an interruption. The Cooling Distribution Unit, or CDU, also has to keep coolant flowing, so that residual heat from dense AI hardware does not accumulate faster than the system can get rid of it. By keeping rack level power on long enough for IT and cooling subsystems to remain controlled, BBU is a pragmatic line of defense for electrical continuity and thermal safety.

UPS vs. BBU

Technical Challenges Beyond Physical Limits: Achieving Extreme Power Density in Limited Space

The Power Curve Has Outgrown the Envelope

AI computing has changed the power demand over the last four years from a board-level design problem to a mechanical, thermal, and packaging problem. Single-unit loads that used to be a bit above 100W are now approaching the 1kW class, which implies almost 10x the electrical output with no requirement for the rack to grow 10x. Meanwhile, industry data agrees and says that modern AI infrastructure is pushing far beyond traditional CPU-era rack densities and that sophisticated accelerators are already reaching hundreds of watts per device.

The 1U Constraint Is Where Engineering Gets Real

The hard part is not just to build a more powerful module. It is having that capability integrated into the customer's existing 1U form factor as cells, busbars, connectors, control boards, insulation spacing, structural parts, and airflow channels are already fighting for every millimeter of space. LITEON's solution is to treat the Battery Backup Unit (BBU) as a compact power density platform, by employing component miniaturization and controlled thermal airflow such that higher output does not result in a commensurate increase in footprint or heat concentration.

Average server rack densities are still below 8 kW, and the majority of facilities still do not run racks above 30 kW. It shows how exceptional AI rack-density requirements have become.

Fivefold Density Gains Require System-Level Packaging

LITEON's power density is approximately 5x higher in the same physical volume than design references five or six years ago. No doubt, it is a jump that is less about finding one better part and more about the accumulation of many small optimizations, such as smaller magnetic components, denser interconnects, lower-loss current paths, improved cell arrangement, more disciplined heat extraction, and mechanical layouts that reduce dead space. That is to say, the engineering problem is no longer only electrical. However, it is about the art of making power, temperature, safety clearance, and serviceability co-exist in a very narrow chassis.

Configuration Flexibility Follows the Compute Scale

Not all AI deployments have the same number of accelerators, backup targets, or redundancy expectations. LITEON can configure BBU plans based on the compute scale and protection level needed, as opposed to one fixed ratio across all racks. Note that existing market configurations already surpass 1:0.7. Thus, such flexibility enables operators to better fine-tune backup capacity from lighter protection for moderate workloads to denser reserve architectures for larger AI clusters. The available rack space must be converted into usable and scenario-specific resilience.

LITEON's Core Technology: PSU and BBU Firmware Integration Enables Zero-Delay Online Backup Switching

Always Backup Starts Before the Fault

LITEON's online backup design considers backup readiness as a live operating state rather than an emergency action in delayed time. The Battery Backup Unit (BBU) is electrically ready to support the load. The backup energy can be delivered at the microsecond level, even if the PSU side goes to an abnormal condition. It is like a "Always Backup" architecture where the backup source matches the operating logic of the system before a workload-level failure occurs.

Microsecond Switching Protects the Compute State

The risk to AI servers is not simply if backup power is available, but if the transition is long enough to upset accelerator boards, memory subsystems, storage transactions, or networked training synchronization. LITEON's approach is meant to make that transfer electrically seamless so that the rack does not experience a meaningful interruption while backup support takes over. It allows active workloads to remain stable instead of being forced into recovery behavior.

Firmware Becomes the Real-Time Power Dispatcher

The intelligence is in digital control optimized in firmware. Voltage, current, PSU health, reserve readiness, and load behavior can be continuously monitored. When the control loop detects an abnormal event, it schedules backup output with sufficient precision to fill the deficit immediately. Here is where power protection becomes active coordination because the system is also trying to see how to keep the electrical operating envelope of the rack in real-time.

LITEON's Advantage Is Owning the PSU–Backup Handshake

LITEON's ability to develop both the PSU and the backup module also allows it to tune switching logic, communication protocol, fault thresholds, timing sequence, and protection behavior at the firmware level. It results in a tighter collaboration than architectures assembled from suppliers specializing in only one side of the chain. Such a system-level advantage is backed by LITEON's AI infrastructure portfolio, which includes its PSUs, power shelves, high-efficiency backup units, power management, thermal design, and system integration.

Quality Defense and Safety Practice: Passing Strict Safety Certification Requirements

Certification Is the Starting Gate, Not the Finish Line

Before we even talk about performance, safety validation of a rack backup module supporting AI infrastructure must start first. A single weak cell, connector, insulation gap, or protection circuit can turn a localized fault into a rack-level risk. LITEON's Battery Backup Unit (BBU) has qualified to global safety certification requirements. This is important as recognized battery standards assess normal operation as well as abnormal electrical, thermal, and mechanical stress conditions that define whether the product is trustworthy in high density environments.

Fire Propagation Testing Defines Real Physical Safety

The battery fire propagation test is the critical checkpoint and asks a very simple question. If a cell or module is driven into an extreme failure condition, does heat, flame, or gas release propagate beyond the intended safety boundary? UL 9540A is used for assessing thermal runaway and fire propagation behavior in battery energy storage systems. Passing this kind of validation shows that safety prevention is designed by choice of materials, enclosure design, spacing, sensing, protection logic, and controlled failure containment.

Low Defect Rate Becomes a Technical Barrier

In this market, low cost without disciplined quality control can create hidden risks such as voltage drift, abnormal heating, cell imbalance, connector degradation, or thermal runaway under repeated charge and discharge stress. When it comes to LITEON, reliability is taken as an engineering indicator, going above and beyond a manufacturing target, and therefore the pursuit of low defect rate becomes part of the product architecture itself. By using tighter screening of components, process control, aging verification, traceability, and failure-mode analysis, we move away from a price driven commodity design to a reliable industrial solution.

Power is the No. 1 reason for significant data center outages, according to the Uptime Institute's 2025 outage analysis. So, backup-system quality control should be a core reliability discipline.

Field Data Turns Safety Into Repeatable Reliability

What bolsters LITEON's stance is lab validation, along with long-term B2B deployment experience with CSP customers. Charge profiles, discharge behavior, temperature response, protection thresholds, and aging patterns are validated against operating data. LITEON leverages leading charge/discharge management technologies and field-proven know-how to assist AI infrastructure operators in creating power continuity systems that are not only theoretically compliant but also robust enough for large-scale and continuous production environments.

Building a Power Resilience Firewall for the AI Era Through Distributed Energy Storage and Professional Expertise

Since AI facilities are using tightly coupled compute, power, and cooling systems, resilience should be closer to the rack. Faster in reaction, more distributed in placement, and more deeply coordinated with the infrastructure that keeps accelerator clusters available under stress. Hence, battery backup unit technology seems to be a core layer of AI-era protection, as rack-level storage is a way to manage fast load fluctuations near the computing hardware.

For LITEON, such a "power resilience firewall" is created through a full-stack rack design capability. High power density in constrained space, 2-minute ride-through for controlled continuity, online backup switching, PSU-BBU firmware coordination, rigorous safety validation, and long-term reliability data from demanding CSP environments.

For operators seeking steady, scalable, and resilient AI data center growth, LITEON's rack-level BBU solutions might turn backup power into an active infrastructure advantage.