High-Density GPU Colocation Pricing and Available Capacity

GPU rack density bands from 10 to 100+ kW

A facility's maximum rack density is not the same as a deployable design. Ask how many racks can operate at that density in one row or suite, what diversity is assumed and how the cooling capacity is measured.

Air, containment, RDHx and direct-to-chip

Usable power is not allocated power

A quoted 100 kW block can describe reserved utility capacity, UPS output, breaker capacity or usable IT load. Normalize every proposal to continuous usable IT kW at the cabinets, with the redundancy design and any operating headroom stated. Then verify whether the power can be concentrated at your required rack density.

Rack weight and floor loading

GPU systems, manifolds, CDUs and rear-door heat exchangers can push rack weights beyond a conventional enterprise cabinet. Provide the final loaded weight, dimensions, center of gravity and service clearances. Confirm slab loading, point loads, anchoring, delivery route, freight elevator capacity and equipment handling before ordering.

Typical high-density pricing by market

These monthly rates are planning bands for roughly 10 to 40 kW deployments. Larger clusters may price differently. Cooling infrastructure, density concentration, network, tax and one-time engineering costs are not automatically included.

Planning bands synthesize public 2026 high-density market estimates. They are not provider quotes or evidence of live capacity. Pricing is normalized only after the deployment design is known.

GPU colocation deployment readiness checklist

• Initial and full-scale IT kW, rack count and ramp dates
• Server model, loaded rack weight and power-cord configuration
• Expected sustained and peak kW per rack
• Cooling approach, facility-water temperatures and CDU responsibility
• A/B distribution design and acceptable maintenance conditions
• Fabric topology, carrier bandwidth, cloud access and cross-connect count
• Staging, burn-in, remote hands and spare-parts requirements
• Future adjacent expansion and any required contractual rights

What current AI hardware actually demands

Density planning used to be abstract. It is not any more, because the reference designs now in market have published power and cooling requirements that decide which buildings can host them. The current generation is the clearest example.

Source: NVIDIA's GB300 NVL72 enterprise reference architecture, which specifies 72 GPUs, direct liquid cooling and a full-rack requirement of up to 142 kW. Actual draw depends on configuration and workload behaviour. Confirm the specific SKU and power profile with your hardware vendor, then hand those numbers to the facility rather than a density band.

Why most facilities cannot take it

The gap between what current hardware needs and what the installed base can deliver is the central fact of this market, and it is wider than most buyers expect.

Source: Uptime Institute's analysis of its 2024 Global Data Center Survey, which reports that about 1% of operators had racks above 100 kW. CDU capacity limits are design-specific. Both are directional: the installed base is moving quickly, and a facility's answer today may differ from its answer six months ago.

This is why "liquid cooled" on a provider's capability list is not an answer to a density question. The useful questions are narrower. What is the maximum sustained kW per rack this hall supports today, with the cooling loop that is actually installed? What is the coolant supply temperature and flow rate at the manifold? What is the CDU capacity per rack, and is it shared or dedicated? Is there a secondary loop, and who owns it up to the rack? A provider who can answer those quickly has done this before.

What high density actually costs

High-density colocation is priced in two or three layers, and comparing a single headline number against another provider's bundled number is where most cost surprises originate.

Ranges are from colocation broker and market reporting rather than operator disclosures, so treat them as directional. The Ashburn air and liquid figures are market reported for 100 kW-plus densities. For audited figures on a consistent basis, the pricing benchmarks page carries quarterly rates reported by the two largest operators.

The questions that separate a real GPU quote from a brochure

1. Sustained and peak kW per rack the hall supports today, with installed cooling, not the design ceiling.
2. Coolant supply temperature, flow rate and manifold type at the rack, and whether they match your hardware's specification.
3. CDU capacity per rack, whether it is dedicated or shared, and what happens to your density if a neighbouring tenant ramps.
4. Loop ownership boundary. Who owns and maintains the secondary loop, and where does responsibility transfer?
5. Floor loading in pounds per square foot against the loaded rack weight, which for these designs is substantial.
6. Power basis and excursion handling. Allocated or metered, and what occurs above committed draw.
7. Redundancy for the cooling loop, not just the electrical path. A liquid deployment has a new single point of failure.
8. Delivery date for the specific hall, and whether the liquid infrastructure is installed or planned.

The last one matters most. A great deal of announced liquid-cooled capacity is not yet built. See how we verify capacity for what we confirm before an option reaches a shortlist.