The CDU touchscreen says supply temperature at setpoint and differential pressure in band, and it is describing a moment that has already gone. On the numbers in the DataCenterDynamics piece, a PT100 probe in a standard pocket mount takes three to eight seconds to register a change [9], the PLC scan cycle adds one to five seconds before the controller acts on it [10], and coolant moving at 0.2 to 0.5 metres per second through a 10 to 25 litre secondary loop needs 15 to 45 seconds to carry corrected supply temperature to the cold plate [8]. Take the best case at each stage and the chain runs 19 seconds; take the worst and it runs 58 [15]. Valve actuation is on top of that.
A facility team believes it bought junction temperature control. What a PID loop tuned at commissioning actually holds is a loop setpoint, sampled seconds late, at a point 15 to 45 seconds of fluid travel from the silicon it is protecting [7].
The other timescale is the accelerator's. An inference burst reaches 60 to 80 percent of peak TDP in 30 to 50 milliseconds as the device crosses from prefill to decode [6]. On a part dissipating 700 to 1,000W under sustained load, that is roughly 420 to 800W of swing per device [18]. Set the fastest link in the control chain against the fastest event it is meant to answer and 19 seconds is about 380 times the duration of the 50-millisecond burst that triggered it [16]. A controller polling once per second takes one reading for every 20 to 33 of those bursts [17], so the burst goes essentially unrecorded.
The pitch on the accelerator side is time-to-compute: ramp governors that existed on earlier hardware have been taken out so a cluster can go from standby to full load in tens of milliseconds [5][4]. That is a genuine win for the training team, and the thermal transient it creates lands on a facility ledger where, by the same author's account, it does not appear in monitoring, reliability calculations or risk frameworks [13]. With rack-level draw above 100kW [3], that entry carries real weight on the ledger.
Where the evidence stops matters as much as where it points. These durations are first-principles arithmetic from loop geometry and published component specifications [11], in an opinion column [19]. Measured junction traces, throttling counts, and failure rates are absent from it. It establishes a floor on response time, which is a different thing from a demonstrated population of failures.
The measurement worth doing uses numbers already on your drawings. Divide secondary loop volume per rack row by design flow rate and you have your own propagation delay, no vendor required. Then compare your telemetry sample interval with 50 milliseconds. Sort racks on those two axes: workload that steps versus workload that ramps, and sub-second junction telemetry versus second-scale loop telemetry. Three of the four cells are defensible on Friday. The fourth, stepping workload watched by second-scale loop sensors, is where a commissioning-tuned PID rack running LLM training sits today.
