Turn-up of a hyperscale AI training cluster
Deployed and commissioned a multi-megawatt, liquid-cooled GPU training cluster in the largest US data-center market — from white space to first job.
A hyperscale AI provider needed accelerator capacity online fast in the most competitive data-center market in the country. The workload demanded rack densities far beyond what air cooling could remove, an InfiniBand fabric sized for collective operations, and a commissioning plan that could take the hall from bare white space to a first training job in weeks — not quarters.
We delivered the high-density, liquid-cooled deployment end to end: direct-to-chip cooling manifolds and CDUs, 130 kW rack power distribution, the InfiniBand fabric, and structured cabling — then ran integrated testing and commissioning against the provider's acceptance criteria. The cluster was handed over feeding jobs and fully instrumented, not just powered on.
What we delivered
High-density racks
130 kW liquid-cooled racks engineered for the latest accelerator platforms.
Direct-to-chip cooling
CDUs, manifolds, and leak detection for the full liquid-cooling loop.
InfiniBand fabric
Non-blocking fabric sized for large-scale collective operations.
Testing & commissioning
Integrated load testing and handover against acceptance criteria.
The result
Planning something similar?
Tell us about your environment, scope, and timeline. We'll outline a clear delivery plan and the team to execute it.