A liquid-cooled hall puts alarms from the coolant distribution units, the leak detection, the power monitoring and the building management system in front of one shift, and a loss of flow gives that shift minutes rather than hours. VARCOTT keeps those alarms in one list under your alarm matrix, judges each on its value and on the quality of its data, and shows beside it the readings that answer whether it is real, whether it is one rack or the whole loop, and what the system has already done. After an event, it gives you the timeline and the record the grid operator asks for.
The building management system, the power monitoring, the coolant distribution unit controllers and the leak detection each keep their own alarm list with their own priorities. Many of their thresholds were set at the maker's defaults or at engineering estimates during design and never revisited. The shift can face hundreds of alarms across those systems with little context, and has to decide which ones mean something.
Liquid cooling changed how much that costs. A loss of flow to a 135 kilowatt rack reaches consequence in minutes. A leak alarm that arrives without its evidence starts a discussion about whose problem it is: one maker of cooling units describes a coolant unit alarm at 2 a.m. that waited about 45 minutes because the IT team lacked the authority to act and the facilities team was not watching that system. A flow meter that has drifted reads adequate flow through a blocked circuit, so the alarm that should have warned never fires. And Texas is moving to require large loads to ride through grid disturbances, while the national reliability body has asked for the processes and procedures behind it, so the record of how a hall rode through an event is becoming something the shift has to be able to produce.
Today the shift answers these with the alarm response procedure, with the senior engineer who knows which alarms matter, and by calling people until someone with authority picks up. That is enough when an event gives hours. A loss of flow gives minutes.
Every alarm from the building management system, the power monitoring, the coolant distribution units and the leak detection lands in one list under your alarm matrix, with the same priorities, the same states and the same operator action text whoever made the device. The states and actions follow the ISA-18.2 alarm lifecycle: acknowledge, shelve, suppress, return to normal, with escalation when nobody responds. The shift works that list instead of four.
An alarm fires on what a reading says. VARCOTT also judges what the reading is worth. A sensor that stops reporting raises its own alarm rather than passing for a calm plant. A value that disagrees with what the physics of the loop says it should be is flagged before it is trusted. A feed that has gone stale is shown as stale. The drifted flow meter is caught because the model of the circuit it sits on expects a different number.
When an alarm fires, the readings that answer the shift's three questions are on the same screen: whether it is real, whether it is one rack or the whole loop, and what the system has already done. A leak alarm carries its zone, the readings around it, its severity and the valve state. The operator answers from the screen instead of from four systems and a phone call, and the answer is on record.
Every alarm, reading, state change and operator action of an event, from every system, goes on one timeline with what the model expected beside what was measured. The report for an operations review, a tenant or the grid operator is produced from that record in days rather than weeks: AI agents reason over the timeline, the connection record and the replayed physics models to reconstruct what happened in what order, and every statement they make is checked, by a deterministic step with no language model in it, against the readings it cites before it is shown, with the evidence kept beside the report. Where your relays and power-quality meters record it, the ride-through record the grid operator asks for is assembled from those records and the historian, aligned on the same timeline and checked against the model of the power distribution. In Texas, a large computational load that fails to ride through a disturbance must report its root cause to ERCOT within 90 days of ERCOT's request, under operating guide revision NOGRR282, in effect since August 2026; that report is written from the same timeline. The engineering page explains how the agents are checked.
An alarm list is only as good as the matrix behind it: for each alarm, its cause, its consequence, the operator's action and its priority. In a process plant, writing that matrix takes a team of operators and engineers a season of workshops. In a liquid-cooled AI hall the equipment repeats, so most of that work can be done once.
The coolant distribution unit, the manifolds and leak zones, the rack power shelves, the panels, the busway and the UPS of one hall are the same classes as the next hall's, so their failure modes, and most of their alarms, can be written once per class and fitted per hall. VARCOTT writes the matrix per equipment class and per platform generation, then fits it to your hall: your redundancy, your controls, your plant upstream, your procedures. What you receive is the matrix for every class in the hall with priorities and operator action text, the test evidence per alarm that commissioning guidance for liquid-cooled halls asks owners to require at handover, the matrix delivered in the form your building management and power monitoring integrator loads rather than as a separate tool, and the matrix kept current as racks are added, thresholds change and the record of the hall changes with them.
Four things are true before the shift picks up the phone.
VARCOTT monitors; it does not control. The leak detection, the coolant distribution unit and the building management system keep every automatic action they have today, and no automatic action is added. A false leak alarm interrupts cooling only if your existing systems were set to do that, and VARCOTT shows the shift that they did.
The alarm carries its zone, the readings around it and what the model says those readings should be: the manifold pressure, the branch flows, the return temperatures of the racks on that row. A real leak shows in the loop's pressure and reservoir level; condensation on a cold line and a wet sensor do not, and those readings are on the screen beside the alarm.
The operator action text is yours, from your alarm matrix. VARCOTT shows it and records what was done; it does not invent a procedure. Escalation follows your chain, and the record shows who was called and when.
It runs on your own hardware, inside your network. It reads the building management system, the power monitoring, the coolant unit controllers, the leak detection and the GPU telemetry, and writes to none of them. The record belongs to you.
Each device is modelled from its maker's data, the coolant distribution unit with its heat exchanger, pump, valve and controls, the manifolds, the racks' cold plates, the panels, the breakers, the busway, and the models are solved together on one clock, so a change in the power distribution shows up in the loop and in the racks in the same step. Every reading is compared with what that coupled model says it should be. That comparison is what makes the evidence beside an alarm more than the alarm's own value, and it is the same model that computes the hall's usable megawatts and its hours, so a hall that starts with the alarm list has part of the capacity study already in hand.
The alarm engine follows the ISA-18.2 alarm lifecycle and runs on live data. The cooling-loop model is our own, verified test by test against the Lawrence Berkeley National Laboratory's Modelica Buildings library; the power-distribution model is built the same way, each device from the standards and the maker's curves. The matrix, the procedures and the record stay yours, in your systems.
VARCOTT is founded by David Perez, an electrical engineer who has written power software for mission-critical facilities since 2013: the white-space power management of Fortune 500 data centers, and one of the first live electrical digital twins in data centers.
The first step is the alarm matrix and one list for one hall, scoped in one call. It starts from your alarm export, your points lists and your last event.