Plumbers in the Data Centre: The AI Skills Gap Nobody Budgeted For
A business case for a rack-scale AI system usually has lines for hardware, hosting, power and support. It rarely has a line for the person who knows what to do when a coolant sensor trips at 3 a.m.
That person is hard to find. The data centre industry was short of skilled staff before AI arrived. Liquid cooling and very dense power have added requirements that barely existed in most facilities five years ago. The skills gap is the constraint that does not show up on a quotation.
What the surveys say
| Finding (Uptime Institute, 2026) | Share of operators |
|---|---|
| π Difficulty finding qualified candidates | 53%, up from 46% a year earlier |
| β‘ Skills gap in electrical roles | 38% |
| π§βπ§ Skills gap in junior and mid-level operations | 38% |
| π Skills gap in operations management | 35% |
| π§ Skills gap in mechanical roles | 34% |
Sources: Uptime Institute Global Data Center Survey 2026, figures as summarised in Second Talent's Asia data centre workforce outlook. These are global figures; Uptime does not publish a separate Asian breakdown.
One recruiter's reading of the same survey adds a telling detail: for the first time in the survey's history, meeting AI infrastructure staffing needs entered operators' top concerns.
Source: Spencer Ogden
Why AI racks make it worse
A traditional data centre technician works with air handlers, power distribution and cabling. A rack-scale AI system adds whole new subjects.
| Skill | π¨ Traditional hall | π§ AI rack hall |
|---|---|---|
| Air cooling and airflow | β Core | β Still needed |
| Coolant loops, pumps and CDUs | β Rare | β Core |
| Coolant chemistry and quality testing | β Not needed | β Routine |
| Leak detection and response | β Not needed | β Essential |
| Busbars and very high rack power | β Rare | β Core |
| Servicing with coolant connections | β Not needed | β Every hardware swap |
| GPU cluster software and scheduling | β Rare | β Core |
An industry executive summed up the new profile this year: technicians who can work with direct-to-chip liquid cooling and rack densities above 100 kW, and specialists who understand fluid dynamics, coolant chemistry and high-voltage power together. That combination used to belong to supercomputing centres. Now every organisation deploying a liquid-cooled rack needs access to it.
Source: Digitalisation World
The evidence closer to home
Singapore's own liquid cooling standard, SS 726:2026, was written partly to address what its announcement describes as the industry's limited experience with data-centre liquid cooling. A national standard acknowledging an experience gap is about as clear a signal as you will get.
There are efforts to close it. Microsoft, for example, partnered with Singapore's Institute of Technical Education to set up an academy teaching applied data centre skills. Training pipelines take years to deliver experienced people, though, and AI racks are arriving now.
Source: JLL on the data centre skills gap
Three uncomfortable observations
1. "The vendor will handle it" has limits
Hardware vendors do offer services for these systems. HPE's list for its NVL72 racks includes site readiness assessment, coolant fill, CDU commissioning, leak detection and thermal monitoring. That covers installation and scheduled support. It does not put an engineer in your data hall at the moment something goes wrong. Somebody on site has to recognise a problem and take the first safe action.
2. "The data centre will handle it" has limits too
In colocation, the operator runs the building and you own the rack. Liquid cooling blurs that line. The coolant loop crosses from their plant into your hardware. If responsibilities are not written down, a leak becomes an argument about whose problem it is, held while the rack is offline.
3. The same few people are wanted by everyone
Hyperscalers, colocation operators, system integrators and vendors are all recruiting from one small pool. Surveys report that staff being poached by competitors has become a retention problem in its own right. A plan that depends on hiring one experienced liquid-cooling engineer is a plan with a single point of failure.
How to plan around it
| Step | What to do |
|---|---|
| 1οΈβ£ Map responsibilities | For each task below, write down one named party: vendor, operator, integrator or you |
| 2οΈβ£ Budget for it | Add operations as a line in the business case, not an assumption |
| 3οΈβ£ Buy the service contract deliberately | Check response times and what is excluded |
| 4οΈβ£ Train your own people | At least two staff should understand the system's alarms and first response |
| 5οΈβ£ Rehearse | Walk through a leak alarm and a cooling failure before the rack carries production work |
The responsibility map
| Task | Who? (fill in before signing) |
|---|---|
| π§ Coolant fill and commissioning | β |
| π§ͺ Routine coolant quality testing | β |
| π¨ First response to a leak or cooling alarm | β |
| π§ Hardware replacement involving coolant lines | β |
| π¦ CDU maintenance | β |
| π» Cluster software, scheduling and monitoring | β |
| π Out-of-hours escalation | β |
If any row has no name against it, you have found your skills gap before it found you.
The quieter alternative
There is another way to deal with a shortage of liquid-cooling skills: need fewer of them. An air-cooled 8-GPU server asks for power and airflow, both of which existing teams understand. For organisations without access to specialist operations staff, that is a legitimate reason to choose a tier down, and one that rarely appears in a technical comparison. We make the case in Why Air Cooling Is Still More Popular Than Liquid Cooling.




