Build stage 5

GPU clusters

GPU clusters are the last stage of an AI data center build. Once the building is commissioned and the operations team has taken over, thousands of GPU servers are delivered, racked, cabled together on a high-speed network and tested as one system before the customer starts training or running AI models on them.

By Czarina Tabayoyong, Founder, Critical Path Hiring. Former Amazon, AWS and Google recruiter.

Last reviewed October 20, 2026 · First published October 20, 2026

Reviewed as terms, pay and markets change.

Building ready for IT, then racks delivered and installed with liquid cooling connected, network fabric cabled, burn-in and acceptance testing, and handover to the customer.

Work

Rack installation

GPU servers often arrive already installed in racks from the manufacturer or an integrator. Technicians move them into the data hall, set them in place, connect power and, for liquid-cooled racks, connect the coolant lines from the coolant distribution units (CDUs) built during construction.

Network fabric

GPUs in a cluster have to share data with each other constantly, so they are connected by a separate high-speed network called the fabric, built on InfiniBand or high-speed Ethernet. That takes a very large amount of fiber cabling, and every run is labeled and tested before burn-in.

Burn-in and acceptance

Every server, GPU and network link is tested under heavy load for days to catch parts that fail early. Bad parts are swapped and retested. The cluster is accepted when it passes the customer's performance tests.

Cluster operations

After handover, cluster engineers and site reliability engineers keep the cluster running: replacing failed GPUs and cables, updating firmware, monitoring performance and scheduling the jobs that run on it.

People

Data center technicians who rack, cable and swap hardware
Network engineers who design and bring up the InfiniBand or Ethernet fabric
Fiber and structured cabling technicians
HPC and GPU cluster engineers who bring up and test the cluster
Site reliability and infrastructure engineers who run it after handover
Deployment program managers who coordinate deliveries, vendors and schedules

Earlier stages

GPU racks draw far more power and give off far more heat than ordinary servers, which is why the power, cooling and commissioning work in the earlier stages has to be planned around them. Commissioning teams test the liquid cooling systems and the power to the racks before the GPUs arrive, and operations teams keep those systems running once the cluster is live.

Hiring impact

These roles sit between the facility side and the computing side, and many teams need people who understand both. Candidates who have deployed GPU clusters can name the hardware generation, the cluster size, the network fabric and their part in bring-up or testing. Many come from high-performance computing at national labs and universities, from cloud providers and from network engineering.

For when to open each search against the build schedule, see the hiring timeline by build stage.

When to hire by build stage

Which searches should be open before each milestone, from design to first rack.

Data center salary guide

Pay bands by role, level and market, with the method stated.

Data center jobs by track

Roles on the build, commissioning, operations and compute tracks, with entry routes.

Czarina Tabayoyong

Founder and principal recruiter, Critical Path Hiring. Czarina recruits commissioning, MEP, power and critical facilities people for data center builds. She previously recruited cloud and data center architecture talent at AWS Professional Services, and PMO and project delivery talent for hyperscale data center campuses in Texas and Virginia.

More about Czarina · LinkedIn