AI and GPU Colocation
High Density Racks, Liquid Cooling and the Data Centers That Can Take Them

Placing GPU servers you own in a data center that can power and cool them. For current rack-scale AI systems that means 120 kW or more per rack and direct-to-chip liquid cooling, which most data center space cannot provide.

Data center deployments at about 25 kW per rack or more, usually driven by GPU and AI workloads. Current AI deployments run from about 35 kW for air-cooled GPU servers to 140 kW and above for rack-scale liquid-cooled systems, while most legacy facilities support 5 to 10 kW.

A GB300 NVL72 draws about 140 to 160 kW and weighs roughly 1.5 tons. The Vera Rubin VR200 NVL72, shipping in volume from the second half of 2026, draws about 190 to 230 kW. Rubin Ultra on the Kyber rack is specified at around 600 kW for 2027. All require direct-to-chip liquid cooling.

A small share. Density-capable capacity concentrates in Northern Virginia and Silicon Valley, with faster availability in Phoenix, Columbus, Atlanta and Reno, and in Canada at QScale near Quebec City and eStruxture in Vancouver. We run live searches across operators at no cost.

No. The deciding question is whether a facility water loop reaches the white space today or has to be built, because retrofitting one into an operating hall is a construction project. A provider supporting liquid cooling and a specific hall having a loop today are different answers.

Air cooling tops out around 30 to 35 kW per rack. Rear-door heat exchangers extend that to about 50 to 70 kW. Direct-to-chip liquid cooling puts cold plates on the processors and supports 200 kW and above, and it is required for current rack-scale GPU systems.

For steady workloads, usually. In our 64-GPU example, owning in colocation runs about 75 percent below hyperscaler on-demand pricing, about half a GPU cloud, and about 25 percent below a 12-month reservation, before power and financing. For bursty or short-term work, renting is usually better.

Renting suits training bursts, early-stage models and capacity needed quickly. Reserved GPU servers and clusters can be live in days or weeks, while owned hardware in colocation takes months. We place both and compare them for your workload.

Generally no. Below that level the work of buying hardware, contracting space and running infrastructure usually outweighs the savings, and staying on rented capacity until workloads stabilize is the better answer.

It depends on size. Up to about 200 kW can often go live within weeks where space is built and powered. Between 200 kW and 1 MW usually takes one to six months. From 1 to 5 MW, most capacity is reserved six to twelve months ahead for 2027 delivery, and 5 MW and up typically means twelve to twenty-four months. Hardware lead times can add to any of these.

Two main ways. Smaller deployments are often quoted at an all-in rate per kW of committed capacity, with electricity included up to that commitment. Larger blocks, often 100 kW and up, are usually billed for space and committed capacity separately, with the electricity you actually use metered and a cooling overhead applied. Comparing quotes means converting both to the same monthly total.

Power usage effectiveness averages about 1.54 across the industry. Air-cooled colocation commonly runs in the 1.4 to 1.6 range, while purpose-built liquid-cooled halls can run well below that, around 1.1 to 1.25. On metered billing, PUE directly raises your power cost, so it belongs in any comparison.

More as the commitment grows. Small deployments need a standard financial review. From around 1 MW, operators expect strong financials or security such as a deposit, letter of credit or prepayment, and multi-megawatt commitments usually require investment-grade credit or substantial security. Many operators will not quote without it, so it is the first thing to prepare.

Outside the coastal hubs. Phoenix, Columbus, Atlanta, Reno and Dallas typically run 20 to 40 percent below Northern Virginia for workloads without a coastal latency requirement, and Quebec lower still on hydro power for training.

The maximum sustained density for the specific hall, the documented path and cost to increase it, cooling service levels separate from power, and density escalation rights giving you a claim on future capacity. Standard contracts contain none of this.

Yes, as paid consulting: what density a site can support, which operators and tenants would take it, power and land, diligence and valuation. Placement for buyers is free; consulting for developers, operators and investors is priced as a project.

Yes, across North America, through formal channel partnerships with Digital Realty, Equinix, DataBank, CoreSite, Cologix and others, and with operators outside our channel when the requirement needs them. Send the requirement and you'll hear back within 24 hours. The operator you choose pays us.