Don't Call Me BESS: Why AI data centers need an active power layer

On one side: the electric grid, refined across a century to serve predictable industrial loads that draw power smoothly and recover gracefully when something goes wrong. On the other side: AI data centers, engineered for maximum compute, drawing power in ways the grid was never designed to absorb, with zero tolerance for disturbance. The grid and the AI factory need an active power layer between them that can absorb volatility in both directions.

I have lost count of how many times someone has called our AI UPS™ a BESS. It’s an honest mistake. Battery storage is well understood—the market leaders have made their case, and the data center industry has internalized it as a foundational layer. So when someone hears we build a battery-backed system that sits between the grid and a hyperscale AI campus, the mental model arrives pre-loaded: big batteries, behind the meter, energy management.

That model is wrong. The gap between the assumption and reality is precisely where the hardest unsolved problem in AI infrastructure sits.

What’s actually happening inside an AI data center

GPU clusters fire up and ramp down in milliseconds, throwing power swings of plus or minus 70% of a single unit's rated capacity. These are not occasional spikes; they are continuous, high-frequency oscillations that existing grid infrastructure wasn’t designed to absorb.

The problem runs in both directions. The data center is exposed to the grid: voltage dips, frequency excursions, momentary faults. But the data center has also become a threat to the grid. GPU load oscillations excite resonances in the transmission system itself, setting up instabilities that propagate well past the fence line.

When a voltage event sweeps through, hyperscale campuses protect themselves by tripping offline. At gigawatt scale, the simultaneous loss of that much load is exactly what triggers cascading failures and regional blackouts.

This is a two-front war, and while most solutions address one front, we address both.

Giving the battery its due

Battery storage is valuable technology. We have built custom BESS solutions for commercial, industrial, and critical infrastructure applications for nearly a decade, and we own and operate our own Independent Power Producer sites. BESS can store energy, shift load, and participate in markets. It has a real role. But it has a fixed position beside the power path, not in it.

When the grid hiccups, a BESS system has to detect the problem, respond, and then transfer. Even a fast static transfer switch takes as little as 4 milliseconds, and often longer. That might not sound like much but in power electronics, it’s an eternity. In that window, voltage has already sagged and GPU hardware has already seen a disturbance. In systems running continuous AI training jobs, that disturbance has consequences.

Moving beyond speed, the deeper problem is architectural. BESS is a parallel system, meaning it sits beside the power path and reacts to what it sees. It is optimized for energy management: storing, shifting, and dispatching power over minutes and hours. That is the problem it was designed to solve, and it solves it well. Power quality conditioning at the microsecond level is a different problem entirely, and no amount of control software changes the fundamental constraint: a system that is not continuously in the circuit cannot provide continuous protection.

This is why most data center developers who start with BESS quickly realize they also need a UPS. This is a separate system, inside the building, to protect the load from disturbances the BESS cannot catch fast enough.

So now you are running two systems: BESS at the meter to manage grid-facing behavior, and distributed low-voltage UPS units inside to protect the servers. When you have two vendors, two maintenance contracts, two commissioning processes, you’ve introduced coordination layer with it’s own failure modes.

And yet neither system, individually or together, reliably satisfies the voltage ride-through and ramp rate requirements that ERCOT is mandating for large loads (and that NERC, CAISO, MISO, and SPP are moving toward).

You're left with two systems, and the full cost and complexity that come with both, but the core regulatory requirement is still unmet.

What inline really means

AI UPS doesn’t sit beside the power path. It is the power path. Every watt flowing from the grid to the data center passes through our system continuously (not just during a fault).

This is what fully inline, double-conversion architecture means. Incoming grid power is converted from alternating current to direct current (AC to DC), stored momentarily in the battery, and then converted back to clean, tightly controlled AC by the output inverter before it reaches the facility. The load doesn’t encounter raw grid power directly.

Because the system is always in the circuit, its transfer time isn’t just fast, it’s zero. There is nothing to transfer because the power is continuously conditioned.

We operate at medium voltage: 13 to 35 kilovolts, the distribution voltage level at which most large facilities interconnect today. Operating at medium voltage instead of low voltage does two things: 1) It pulls the equipment out of the building, freeing indoor space for additional compute and cooling. 2) It also puts protection at the boundary, covering the whole site in both directions. Everything inside the fence is shielded from the grid. One system covers the entire electrical plant (IT load, cooling, pumps, and mechanical equipment), not just the server racks.

Both fronts, simultaneously

I mentioned the two-front war. Here is what fighting both sides at once looks like.

Protecting the data center from the grid. Because every watt passes through the system, the load is fully decoupled from the grid. Whatever happens upstream (voltage dips, momentary faults, the instabilities that ripple through a stressed transmission system) we filter out before it reaches the load. The data center never sees it. GPUs keep running and training jobs stay intact, no matter what the grid is doing. We have run this against live AI load profiles at full capacity and grid faults drop to zero, and the load keeps running. We have measured data using real hardware, under real grid conditions. The system meets every proposed voltage ride-through requirement, including the strictest standards regulators are imposing on large loads

Protecting the grid from the data center. When thousands of GPUs spin up at once, there is a sudden, large demand spike on the grid. ON.energy’s system absorbs that spike internally with integrated storage and presents a smooth profile outward. In testing, we put swings of more than 70% of system capacity on the input, and the grid saw a flat line. And when the grid faults, the data center does not trip offline and dump gigawatts of load at once—it rides through. We have tested full voltage ride-through against the most stringent requirements now imposed on large loads, while running live AI workloads. AI factories can pull as chaotically as the workload demands. The grid sees a smooth, stable load profile.

What we are

Some have stapled BESS and UPS together and called it a solution. We understand why. When the right tool doesn't exist, you work with what you have.

We built the right tool and deployed it at scale: one system, inline, medium voltage, fully proven in the field. The technology has been reviewed, approved, and adopted by the hyperscalers and regulatory bodies. It solves 100% of the problem at a lower cost than the two half-measures it replaces. The market has spent years waiting for something that actually works. And we’re here.

But don’t call us BESS. We’re AI UPS.