AI data centers are scaling to unprecedented gigawatt levels, creating extreme dynamic loads driven by GPU synchronization and rapid power ramping. Unlike traditional workloads, AI training clusters generate unpredictable millisecond-level power spikes that destabilize grids and generators, turning energy management into a core architectural challenge.

AI training clusters don’t sip power – they draw it in sudden bursts. When hundreds of megawatts of GPUs spin up simultaneously, they create intense “pulse loads” that can swing site demand dramatically within milliseconds. Generators cannot ramp fast enough to respond, and utilities cannot tolerate infrastructure repeatedly surging and dropping load.

This shift is redefining the role of batteries in AI infrastructure. Rather than passive backup systems, next-generation lithium-ion batteries are becoming active stabilizers that absorb pulse loads, smooth demand, and enable real-time load balancing.

At DCD>Connect | New York 2026, Ampace Technology explored this challenge in its showcase, “Powering gigascale AI - How advanced batteries stabilize extreme training loads.”

“You can’t have a gigawatt site loading up utility and then dropping out continuously – it wreaks havoc on all those different devices,” said Aaron Schott, UPS sales manager at Ampace.

Schott outlined how Ampace is re‑engineering lithium‑ion batteries – from cell chemistry to control software – to absorb AI power fluctuations, while maintaining safe, reliable operation over a 10-15 year lifecycle.

From micromobility to gigascale AI

While Ampace is increasingly focused on digital infrastructure, the company’s expertise extends far beyond the data center sector. As one of China’s largest lithium-ion battery manufacturers, the company develops end-to-end solutions spanning electric micromobility, high-power applications such as drones and power tools, and behind-the-meter energy storage.

That breadth allows Ampace to control the full battery stack, from cells and modules to complete integrated systems. According to Schott:

“As an end-to-end solution provider, we’ve been talking to all the major hyperscalers and working with different UPS partners, developing different solutions to support the load, no matter where it comes into the data center.”

For AI data centers, Ampace focuses on three layers of infrastructure: in-rack battery backup units (BBUs) and 800V DC sidecars, UPS-connected battery cabinets, and grid-scale battery energy storage systems (BESS).

Across all three, the goal is the same: smoothing the irregular pulse loads generated by AI GPUs before they reach generators or utilities.

The AI pulse load problem

Traditional data center loads are relatively stable and predictable. AI clusters, by contrast, generate synchronized bursts of activity that create rapid oscillations in power demand across entire facilities.

At the rack level, batteries may endure tens of millions of microcycles involving repeated 50-millisecond charge and discharge events. At the site level, synchronized GPU clusters can trigger coordinated ramp-up and ramp-down events across 100-300MW deployments, especially where gas turbines are involved.

“You have a big problem because the gas turbine can’t adjust to that frequency of change,” Schott said. “So we have to look at how we support the microcycles, and how we support all this ramp-up, ramp-down functionality.”

As a result, batteries are becoming the front line for stabilizing AI infrastructure – smoothing microcycles, buffering load swings, and supporting high charge and discharge rates over a projected 10-15 year lifespan.

An end‑to‑end battery architecture

To manage AI volatility, Ampace has developed a multi-layered battery architecture spanning the full data center power chain, from in-rack systems to grid-scale energy storage.

At the rack level, Ampace has developed the SP25, a custom high-power cylindrical lithium iron phosphate (LFP) cell for BBUs (battery back-up units) and 800V DC sidecars, derived from its power tool battery line, but redesigned for AI data centers.

“This SP25 is capable of over 200 watts per cell – it’s extremely high power – but we have to be able to support that power in the BBU as safely as possible,” Schott explained.

Higher up the power stack, Ampace’s PU 100 and PU 200 UPS battery cabinets use LFP pouch-cell technology to support a range of runtime and discharge profiles, from short-duration 10C systems to longer backup applications.

The systems incorporate forced-air cooling and optimized thermal management to maintain stability during aggressive cycling. For a typical 1.25MW UPS deployment, Schott says configurations can deliver between five and seven minutes of end-of-life runtime while reducing the number of cabinets required – a critical consideration as AI campuses scale toward gigawatt capacity.

At the grid edge, Ampace is deploying 5MWh containerized LFP BESS, designed to buffer interactions between AI campuses, utilities, and on-site generation. Together, these systems create an end-to-end architecture in which batteries absorb and smooth AI-driven volatility before it can disrupt upstream infrastructure.

Smarter batteries for AI loads

For Ampace, solving AI power volatility requires more than advanced battery hardware alone – it also demands a fundamentally different control architecture.

Traditional data center batteries discharge infrequently, often just once per day. AI workloads, however, create continuous high-frequency cycling that demands faster monitoring and more adaptive battery management systems (BMS).

To address this, Schott presents Ampace’s increased BMS sampling rates to track state of charge (SoC), temperature, and current under rapidly changing conditions. The company has also introduced adaptive load balancing across multi-cabinet UPS deployments, allowing systems with 10 or more battery cabinets to distribute microcycling evenly and maintain a consistent state of health across the fleet over time.

SoC management is especially critical. Batteries must be able to absorb excess power when GPUs pause and gas turbines cannot ramp down quickly enough, while still retaining enough capacity to discharge during pulse events or grid disturbances.

The result is a more software-defined, dynamically-managed battery platform designed to adapt in real time to the rapidly evolving load patterns of AI infrastructure, rather than functioning as a conventional static UPS backup system.

“You have to feed it to the battery,” Schott explained. In this way, batteries become the front line for absorbing and reshaping that power.

Making lithium‑ion safer

As lithium-ion batteries take on a more active role in AI power systems, safety becomes a central engineering challenge. Ampace’s response centers on semi-solid state battery technology.

The primary safety concern in conventional lithium-ion batteries is the liquid electrolyte, which is both flammable and electrically conductive. In thermal runaway events, the electrolyte can accelerate off-gassing, fire propagation, and cascading system failures.

“Lithium‑ion batteries have always been given a bad name because of that liquid electrolyte,” Schott noted

Ampace’s semi-solid state design reduces electrolyte content while maintaining enough ionic conductivity to support the high-power demands of AI and UPS workloads. The result is a substantially lowered fire and leakage risk, reduced smoke generation, and limited off-gassing during thermal runaway events.

According to Schott, Ampace tested scenarios where all 32 cells in a module were driven into thermal runaway simultaneously, with no thermal propagation between adjacent modules and no cascading failures across the cabinet. The reduced electrolyte content also cut smoke generation to roughly one-quarter to one-fifth of that seen in conventional lithium-ion systems.

Beyond cell chemistry, Ampace also designs around UL 9540A-style thermal propagation standards and incorporates fire suppression at both the module and cabinet level, helping contain incidents even in the event of an external fire.

The broader goal, Schott said, is to provide data center operators with greater confidence that battery failures – even in the worst-case scenarios – remain isolated, controlled, and non-catastrophic.

Preparing for the next wave of GPUs

As AI infrastructure rapidly evolves toward denser racks, higher GPU power draw, 800V DC distribution, and solid-state transformer (SST) architectures, Ampace is working closely with hyperscalers, UPS manufacturers, and infrastructure partners to anticipate how future AI workloads will behave.

This includes active collaboration with firms such as DG Matrix on SST-based power architectures, alongside the development of custom battery cells tailored for next-generation 800V DC battery backup systems and emerging rack-level power designs.

As AI campuses scale toward gigawatt capacity, deployment speed is becoming as important as performance. To streamline installation and reduce commissioning complexity, Ampace ships its UPS battery cabinets fully pre-assembled with modules already installed, minimizing on-site integration work across large deployments.

For Schott, continuous validation is central to the approach. “Let’s test as much as we can. Test, test, test,” he said, underscoring that reliable AI power infrastructure will depend on iterative engineering, rigorous stress testing, and close collaboration across the data center power ecosystem.

To find out more, visit Ampace.com.