Executive Summary
- The industry is moving from long-standing in-rack conventions toward higher-voltage DC distribution designed for megawatt-class racks.
- When it comes to measured power, the power loop is no longer the steady, predictable part of the data center.
- Nick Onyett at Data Center Solutions makes a short list of questions worth asking before the design locks to help operators come out ahead.
Power monitoring was built for a slower data center
Data centers have measured power at the rack for years. Metered and intelligent PDUs report voltage, current, and energy, often down to the individual outlet, and pass it upstream to management software. Visibility, in that sense, is not new.
But that instrumentation was built for a slower problem: accounting for consumption, planning capacity, allocating cost, catching a breaker before it trips. Those are steady-state questions, and steady-state questions were the right ones when rack behavior was relatively stable. The data was collected on the cadence those jobs required, polled in seconds or minutes and read after the fact.
AI changed the cadence of the load. A training rack swings from near-idle to full draw and back in seconds as jobs start, checkpoint, and stop. The measurement is there, but it was designed to tell you what happened, not what is happening. And you cannot coordinate, in real time, against a picture that resolves a step too late.
Nameplate is the wrong number to plan from
Consider how power gets allocated. Capacity is planned from nameplate ratings, the maximum a device could draw, printed on its label. Breakers, circuits, and PDUs are sized against that figure, with headroom reserved on top. It is a static number, fixed at design time, describing a worst case that may never arrive.
Real draw looks nothing like the label. It moves constantly, and at AI densities it moves violently, swinging with every job that starts and stops. Planning a dynamic system from a static number forces a bad trade-off. Reserve enough headroom to stay safe against a spike you cannot see coming, and you strand capacity behind it. Run closer to the limit to reclaim that capacity, and you take on the risk of a transient you have no way to catch in time.
Most operators choose safety. It is the right call given what they can see, but the cost is real, and in this market it is steep. Power is a gating constraint on AI expansion, which means stranded capacity is not an accounting abstraction. It is compute you have already built and paid for, sitting dark behind reserved breaker space.
Nameplate tells you what a rack could draw. The number that matters now is what it is drawing, moment to moment.
Turning the meter into an instrument
The gap between nameplate and real draw only closes if you can see the real draw as fast as it moves. That is the shift underway in power distribution: from metering built to report consumption to metering built to catch behavior.
The difference is resolution and delivery. Conventional monitoring polls the PDU on a fixed interval, typically every few seconds, and reports through protocols designed for periodic status checks rather than live signal. That cadence is fine for a meter reading. It is far too slow to register a swing that begins and ends in milliseconds. Next-generation power distribution closes that gap by capturing power data at millisecond resolution and streaming it continuously, instead of waiting for the next poll. For the first time, the measurement moves as fast as the load.
Speed matters here because the load does not swing at random. In large training clusters, thousands of GPUs ramp up and down in lockstep as jobs compute, synchronize, and checkpoint together. Those synchronized swings are large, fast, and repeating, and they are invisible to any system sampling in seconds. Catch them at millisecond resolution and the operator can finally plan capacity against what a rack actually does, not the worst case its label allows.
This is where the PDU stops being a passive distributor and becomes the sensory layer of the facility. Placed at the rack, exactly where the volatility lives, it is the natural point to capture the real electrical signature and hand it to whatever needs to act on it, from capacity planning to real-time coordination with cooling.
Why this gets harder from here
None of this eases as the industry scales. It intensifies. Rack density is climbing from tens of kilowatts toward a hundred and beyond, and the power architecture feeding those racks is being rebuilt to keep pace, with the industry moving from long-standing in-rack conventions toward higher-voltage DC distribution designed for megawatt-class racks. Denser racks and thinner margins make every uncaught transient more consequential, not less.
The swings scale too, and this is the part that turns a rack problem into a facility problem. Because thousands of GPUs move in lockstep, their individual swings do not average out. They stack. Aggregated across a training hall, synchronized load steps can reach tens of megawatts, fast enough and large enough to register at the utility interconnect and raise real grid-stability concerns. What begins as a single rack drawing unevenly becomes, at scale, a signature the grid can feel.
That is why the resolution of your power data is not a detail. It is the foundation everything above it depends on. Capacity planning, fault response, and any hope of coordinating power with cooling all inherit the speed and fidelity of the measurement beneath them. Build that layer for a slower era, and everything above it inherits the lag.
What to decide before the next build
For anyone specifying infrastructure now, this reduces to a short list of questions worth asking before the design locks.
First, at what speed does your power data actually resolve? Not whether you meter, but whether the measurement is fast enough to catch a millisecond swing or only quick enough to log an average. Second, are you planning capacity from nameplate ratings or from measured, dynamic draw? The gap between the two is stranded compute, and at current power prices that is expensive ground to cede. Third, where does your power telemetry go? If it flows into a dashboard to be read later, it is documentation. If it flows, fast, into something that can act on it, it is control.
The through-line is simple. The power loop is no longer the steady, predictable part of the data center. It is among the fastest-moving, and the instruments watching it have to keep the same pace. The operators who come out ahead on capacity and stability will be the ones who stop treating power data as a meter to read after the fact and start treating it as a live signal to act on.
