AI Data Centers: Power and Cooling Skills Employers Need Now

AI data center skills are the electrical, mechanical, controls and commissioning skills needed to run racks that draw ten times or more the power of a typical enterprise rack and are cooled with liquid instead of air.

In practice that means medium-voltage and high-density power distribution, direct-to-chip liquid cooling loops, CDU and leak detection controls, and commissioning that tests water as well as watts.

Few engineers have done all of this at scale, which is why data center liquid cooling jobs are some of the hardest roles to fill right now. Many of the AI campuses on our US data center map are being built for exactly these high-density loads.

Why are AI data centers changing the skills employers need?

Demand is growing fast and it is concentrated in AI. The International Energy Agency’s Key Questions on Energy and AI (April 2026) puts global data center electricity use at about 485 TWh in 2025, rising to about 950 TWh by 2030. Electricity use by AI-focused data centers grew about 50% in 2025 and is set to triple by 2030, according to the same report.

The money behind it is large. The IEA reports that capital spending by the five biggest tech companies topped $400 billion in 2025 and is expected to rise about 75% in 2026. For a wider set of numbers on capacity, construction and jobs, see our data center statistics roundup.

What matters for hiring is where that power goes. AI training and inference servers pack far more power into each rack than the servers most facility teams grew up with. That single change touches every discipline in the building.

How much power does an AI rack draw compared to a normal rack?

The gap is wide. The Uptime Institute Global Data Center Survey 2025 found the most common average rack density was almost 9 kW, up from 8.3 kW in 2024. Uptime reports growing adoption of 10 to 30 kW racks, with few operators running racks above 30 kW. About one third of operators said they run AI training or inference.

A single AI rack can be an order of magnitude beyond that. Supermicro’s GB200 NVL72 datasheet (February 2025) lists operating power of 125 kW to 135 kW for one liquid-cooled rack of 72 NVIDIA Blackwell GPUs. The cooling options include an in-rack coolant distribution unit (CDU) rated at 250 kW and an in-row CDU rated at 1.3 MW.

The roadmap goes higher. NVIDIA’s 800 VDC architecture post (May 2025, updated January 2026) says the company is planning for 1 MW IT racks and beyond starting in 2027.

Other surveys show the same shift. AFCOM’s State of the Data Center 2026 executive summary reports average rack density of 27 kW, up from 16 kW the year before. AFCOM and Uptime survey different groups and measure density differently, so the numbers are not directly comparable. Both point the same way.

What is data center liquid cooling, and why does it matter for jobs?

Air cooling runs out of room as racks get denser. In the Uptime Institute Cooling Systems Survey 2025 (fielded April to June 2025), about three quarters of respondents put the point where air cooling becomes too costly, and direct liquid cooling becomes necessary, at 20 kW per rack or more. The most common answer was 40 to 49 kW (25%). Twenty-two percent of organizations said they use direct liquid cooling in their data centers.

Adoption is moving. AFCOM’s 2026 executive summary says 36% of respondents have liquid cooling deployed and another 28% plan to adopt it in the next 12 to 24 months.

Direct-to-chip vs. immersion

Direct-to-chip (DLC) pumps coolant through cold plates mounted on the processors. The rest of the server may still be cooled by air. This is the approach used in rack-scale systems like the GB200 NVL72.

Immersion submerges whole servers in a dielectric fluid inside a tank. It removes more heat per rack but changes how technicians service hardware, handle fluid and manage floor loading.

The loops engineers now have to own

Most direct-to-chip designs use two loops. A secondary loop, often called the technology cooling system, carries treated coolant from the CDU to the cold plates. A primary loop, the facility water system, carries heat from the CDU to chillers, dry coolers or cooling towers. The CDU sits between them, controlling flow, pressure and temperature.

ASHRAE Technical Committee 9.9 classifies liquid cooling by facility water temperature, with classes W17, W27, W32, W40, W45 and W+ in its 2021 liquid cooling white paper. Warmer water classes let sites run chillers less or skip them, which changes how mechanical engineers design and operate plants.

What power skills do AI data centers need?

Electrical teams see the biggest jump in consequence per rack. A fault on a 130 kW rack affects far more compute than a fault on an 8 kW rack. Employers look for:

• Medium-voltage experience. Switchgear, transformers and utility interconnects, since large AI campuses take power at medium or high voltage.
• Higher-density distribution. Busway, 415/480V distribution to the rack and larger power whips, plus familiarity with the DC distribution NVIDIA and others are planning.
• UPS and battery energy storage (BESS). Sizing and maintaining UPS for large, fast-changing AI loads, and lithium-ion battery systems.
• Arc flash and NFPA 70E. Higher available fault current means stricter energized-work practices, labeling and PPE programs.
• Power quality and studies. Short-circuit, coordination and arc flash studies, and an understanding of how synchronized GPU loads swing.

The copper math shows why this is a new field. NVIDIA’s 800 VDC post says a 1 MW rack on traditional 54 VDC in-rack distribution would need up to 200 kg of copper busbar, and that 800 VDC can carry 85% more power through the same conductor size than 415 VAC. Engineers who understand DC protection and higher-voltage safety are rare today.

For more on staffing the power side, read our guide to building high-performing teams for data center power.

Recruiter tip: When a resume says “data center electrical,” ask what voltage the candidate worked on and what the largest single rack or row load was. Someone who has only worked downstream of the PDU is a different hire than someone who has operated medium-voltage switchgear.

What mechanical and cooling skills do AI data centers need?

Mechanical teams go from managing air to managing fluid. That is a different job. The skills that come up most often in data center liquid cooling jobs:

• CDU operation and maintenance. Pumps, heat exchangers, filters, redundancy and failover.
• Secondary and primary loop design. Hydraulic balancing, pressure drop, flow rates and approach temperatures across the TCS and FWS.
• Water treatment and fluid chemistry. Coolant selection, corrosion and biological control, filtration and wetted-material compatibility. Cold plates have very small channels, so dirty coolant becomes downtime.
• Leak detection and response. Sensor placement, automatic isolation valves and written procedures for a leak on a live rack.
• Chiller plants at warmer temperatures. Running chillers, dry coolers and economizers at higher supply temperatures, and heat reuse where it applies.

Uptime’s 2025 cooling survey found the top barriers to direct liquid cooling were lack of standardization across systems (39%), cost (38%) and reliability concerns (35%). Experienced engineers are how operators get past the reliability worry.

What controls, BMS and commissioning skills are needed?

Liquid cooling adds fast-moving systems to monitor. A pump failure or leak on a liquid loop needs a quick, automatic response, and the building management system has to act on it. Employers want controls engineers who can:

• Integrate CDU telemetry (flow, pressure, supply and return temperature, leak alarms) into the BMS and electrical power monitoring system.
• Write sequences of operation that coordinate CDUs, chillers and IT load.
• Work with Modbus, BACnet and SNMP, and with DCIM tools that pull data from both facility and IT sides.

Commissioning changes too. Integrated systems testing (IST) for a liquid-cooled hall means flushing and filling loops, testing with load banks that can heat liquid, proving leak detection and isolation, and running failover with water in the pipes. Commissioning agents who have done this on a live AI hall are few.

Recruiter tip: Ask commissioning candidates to walk through a liquid cooling IST script step by step. People who have done it will talk about flushing, fluid sampling and failure scenarios without prompting.

Which roles need which AI data center skills?

Why are AI data center engineers so hard to find?

Three reasons stand out. First, the experience base is small. Uptime’s 2025 cooling survey puts direct liquid cooling at 22% of organizations, so most facility engineers have never run it. Second, the same people are wanted by hyperscalers, colocation providers, neoclouds, contractors and equipment makers at once. Third, general staffing is already tight: Uptime’s 2025 global survey found 46% of operators struggle to find qualified candidates, and AFCOM’s 2026 summary says 87% report workforce gaps are affecting operations.

Demand for power talent is especially high. AFCOM’s 2026 summary lists power systems and energy engineers among the roles respondents expect to need, named by 36%.

Where can employers find people with AI data center skills?

Look beyond other data centers. The underlying skills exist in several industries:

• Semiconductor fabs. Process cooling water, ultrapure water, high-reliability power and strict change control.
• HPC and national labs. Many supercomputers have used direct liquid cooling for years.
• Industrial process cooling. Chemical, pharmaceutical and food plants run large chilled-water and heat-exchange systems.
• The U.S. Navy and nuclear power. Electricians, electronics technicians and machinist’s mates trained on shipboard power, pumps and cooling systems, with strong procedure discipline.
• Hospitals and critical healthcare. Engineers used to running power and HVAC that cannot go down.
• Utilities and substations. Medium- and high-voltage experience for the electrical side.

Hire for the fundamentals and train the data center specifics. A fab process engineer can learn CDU sequences faster than a generalist can learn fluid chemistry. For more on the role itself, see what a data center engineer does, and for help with electrical searches, our electrical engineering recruiters page.

Frequently asked questions