GPU colocation in Canada: what AI teams need from a high-density data centre
09.30.2026
GPU colocation means placing your own GPU servers in a data centre built for dense AI hardware. You own the compute; the facility provides the power, cooling, network and security around it. In Canada, Hut 8 has 5 MW of high-density capacity available now across two Tier III designed data centres, in Kelowna, BC and Mississauga, Ontario, at 135 kW and up per rack.
This guide covers what to check before you place a GPU cluster:
- power per rack
- cooling
- resilience
- connectivity
- data residency and time to deploy.
What makes a data centre ready for GPU colocation?
A GPU cluster packs a lot of power into a small footprint. A facility built for general IT spreads that load across a floor. A high-density facility delivers it to the cabinet.
Three things decide whether a site can host a GPU cluster:
Power per rack
in kilowatts, delivered to a single cabinet.
Cooling
that can remove that heat, including liquid cooling as densities rise.
Resilience
in power and cooling, so a single failure doesn't stop training or inference.
How much power per rack does a GPU cluster need?
It depends on your hardware, so ask for the per-rack figure, not a facility total. Hut 8's Kelowna data centre supports up to 135 kW per rack in 50U cabinets. The Mississauga data centre supports 135+ kW per rack in 42U cabinets, one of the highest densities in the market.
At 135 kW and up, a full GPU node deploys into a single cabinet instead of being spread across a floor.
Does GPU colocation need liquid cooling?
As rack power rises, air alone does less of the work. Both Hut 8 Canada high-density sites are liquid-cooling capable. Kelowna runs a chilled-water closed loop plus air, with N+1 cooling. See power and cooling for how each site is built.
What resilience should an AI data centre have?
Look for the design standard, the power path and the generator runtime:
- Kelowna (KEL1): Tier III designed, 2N (A+B) power, N+1 cooling, 2 × 1,500 kW generators with more than 29 hours at full load.
- Mississauga (GTA1): Tier III designed, 2N power and 2N UPS, 10 × 700 kW generators (N+1) with more than 24 hours of runtime.
Both sites carry a 99.99% uptime SLA and are monitored by a 24/7 network operations centre.
Kelowna or Mississauga: which site fits your GPU cluster?
Download Now
Get the high-density one-pager.
Available capacity, per-rack density and site specs for Kelowna and Mississauga on one page.