Bob Walicki, Corporate Scientist at Ecolab, explores why clearer commissioning criteria and better fluid-quality validation are needed as data centres scale increasingly powerful GPU clusters.
The handover from construction to operations has become one of the highest-risk moments in a data centre’s lifecycle, and the industry is finally starting to treat it that way. A GPU cluster built for AI training or inference can represent tens of millions of dollars in compute silicon before a single workload ever runs on it. Increasingly, that equipment is cooled not by air but by liquid moving through cold plates with internal passages measured in microns. The result is extreme capital concentration alongside mechanical tolerances an order of magnitude tighter than those found in a legacy chilled-water loop.
The reality is that, for most of the last two years, direct-to-chip (D2C) liquid cooling has scaled faster than the commissioning practices meant to support it. Every operator has been tackling the same question, largely in isolation: how do you know a loop is actually clean, chemically stable and ready to carry heat away from some of the most expensive silicon ever put into production? Acceptance criteria vary by site, by contractor and sometimes by shift.
That inconsistency is now getting real scrutiny. Organisations including the Open Compute Project (OCP) and ASHRAE have published pre-commissioning and fluid-quality guidance for liquid-cooled IT equipment, and hardware makers are working directly with cooling and fluid-science partners to specify how pipework should be cleaned and qualified before it ever touches a cold plate. There is not yet a universal standard, but the direction is clear: cleanliness and validation are moving from varied practices towards documented, auditable criteria that support reliable performance.
The reason cleanliness deserves this level of rigour is straightforward once you look inside a cold-plate microchannel. Passages that narrow to a few hundred microns or less don’t tolerate the same debris that a facility water loop shrugs off. Weld slag, thread sealant, construction dust and biofilm from stagnant hydrotest water are all common in newly built pipework, and every one of them is small enough to lodge in a channel and start degrading thermal performance long before it causes an outright failure. According to coolant testing conducted by Ecolab across operating data centre sites globally, roughly 32% of glycol samples fell outside the specification ranges set out in OCP and ASHRAE guidance – an indication of the fluid-quality challenges that can remain in systems that were, on paper, commissioned and running.
Addressing these challenges doesn’t necessarily require new technology. It requires sequencing and documentation that treats the liquid loop the way a mechanical engineer would treat any other critical process system, rather than as plumbing to be signed off and forgotten.
Start with a written pre-commissioning plan, developed before construction finishes, not improvised once the pipework is installed. That plan should specify make-up water quality, cleaning chemistry compatible with the system’s metallurgy and elastomers, and where wastewater will go – a question worth raising with local regulators early, since acceptable discharge routes vary by jurisdiction and can otherwise stall a schedule at the worst possible moment.
Flushing itself should be staged, rather than completed in a single pass. Coarse filtration removes the larger debris first; fine filtration, down into the low single-digit micron range, comes only once the bulk contamination is gone. Bringing ultra-fine filters into service too early just clogs them. Flow velocity matters as much as filter rating: most cleaning specifications call for 0.9-1.5 m/s (3–5 ft/s) to keep sediment suspended and moving towards the filter rather than settling back onto the pipe wall. And once a hydrotest is complete, that water needs to start moving again immediately. Stagnant test water is exactly the environment biofilm and sediment need to establish themselves, so the gap between hydrotest and active circulation should be measured in hours, not days.
Instrumentation is what turns this from a one-off cleaning exercise into a documented acceptance process. Calibrated flow meters, pressure and differential-pressure gauges, and inline conductivity, pH and turbidity monitoring should all be in place before flushing starts, not added afterwards to tick a box. Capturing that data at baseline – during the initial fill, before the system ever runs a real workload – gives commissioning teams and operators a reference point they can audit against for the life of the system, and gives facilities teams an early-warning trend line instead of a surprise months later.
That baseline should also mark the handover from commissioning to operations, rather than the end of the process. A loop that passed acceptance criteria on day one can still drift: coolant chemistry degrades, particulates accumulate, and seasonal or workload changes shift flow and thermal demand. The same instrumentation installed for commissioning can be extended into continuous, real-time telemetry, tracking chemistry, particulate levels and flow performance over the system’s operating life so that drift shows up as a trend engineers can act on, not a failure they have to diagnose after the fact.
For operators, commissioning teams and facilities managers approaching a D2C handover, a few requirements are worth insisting on, regardless of which vendors are involved. Ask for a written, documented commissioning plan rather than verbal assurance that a system was ‘flushed and ready’. Require records of flushing velocity, fluid specifications, filtration stages and cycle duration, not just a final pass/fail result. Confirm cleaning chemistry was matched to the system’s actual materials, with a plan for handling any waste it generates. And make sure the instrumentation used to validate the system at handover stays in place afterwards, feeding the same monitoring programme that will run the loop for years, not just the weeks around start-up.
None of this is exotic engineering. It’s the same discipline mechanical and process engineers have long applied to other critical fluid systems, adapted to tolerances that are new to this industry. As GPU cluster densities keep climbing and the standards bodies continue converging on shared guidance, the operators who build this rigour into their handover process now – rather than waiting for a single codified standard to arrive – are the ones best positioned to scale quickly and consistently.
