• Committed power: 10–80kW per rack, billed at 80–100% utilization; A/B 3‑phase 208–415V feeds.
• Cabinet or cage: 42–52U, with appropriate depth and weight rating for GPU/HPC chassis.

• Standard cooling:

1. Air / hot‑aisle containment up to ~20–30kW per rack.
2. Rear‑door heat exchangers in some facilities to push to ~30–40kW.

• Network port: 1–10Gbps included; sometimes 10G commit or burst‑able.

• Cross‑connects: $100–$400 per month per x‑connect (to cloud on‑ramps, carriers, IX).
• Remote hands: $150–$300 per hour for GPU swaps, cable tracing, etc.
• Install & turn‑up: $500–$3,000 for rack/stack, cabling, and initial testing.
• HD cooling upgrades: RDHx or “liquid‑ready” manifolds adding $500–$2,000 per rack per month at 40kW+.
• Burst bandwidth / 95th percentile: Overage charges if you go beyond your commit on heavy data movement.

​

Real‑world difference between the “marketing” quote and the actual monthly run‑rate often lands 20–60% higher, depending on cross‑connect count, support usage, and cooling path.

Option A: Google / GPT + vendor sites

“Liquid cooled colocation” searches surface mostly coolant vendors and hyperscale pilots, not colo for 20–100kW. No visibility into live CDU capacity.

​Option B: Traditional colo sales

“HD” pages cap at 15–20kW air; “liquid ready” often means no actual hardware. You cycle through sales teams with no real liquid experience.

​Option C: Coolant / hardware vendors

They push their ecosystem but miss optimal colo pricing, markets and operators for your specific kW and network needs.

Option D: QuoteColo (liquid broker)

Share kW, cooling type and markets. We check 500+ sites for deployed liquid and send 3–5 options with $/kW, design details and lead times. Free to you; typically 10% savings.

Rear Door Heat Exchangers (RDHx)

RDHx replaces the rear rack door with a liquid‑cooled heat exchanger that removes heat directly from server exhaust air. Modern active RDHx systems can handle 15–200+ kW per rack, while passive variants are ideal up to ~20kW with no extra fan power.

• Best for: 20–40kW racks where you want to keep standard air‑cooled servers.
• Pros: Minimal changes to IT gear, retrofits into existing rows, good stepping stone to full liquid.
• Cons: Still air‑dependent; very high densities or uneven airflow can be tricky.

Direct‑to‑Chip (D2C) Liquid Cooling

Direct‑to‑chip runs coolant through cold plates attached to CPUs/GPUs, with a CDU managing a closed loop and rejecting heat to facility water. As rack densities exceed 30kW, D2C hits TCO parity with air and becomes 20–35% cheaper than optimized air above 50kW per rack when you factor in PUE and space.

• Best for: 30–60kW racks with standardized server platforms (e.g., HGX reference designs).
• Pros: High thermal efficiency, good fit for modern AI racks, straightforward PUE gains.
• Cons: Requires compatible hardware, CDUs, and careful leak detection and maintenance.

Immersion Cooling

Immersion places entire servers in a dielectric fluid, eliminating most server fans and enabling extremely high rack‑equivalent densities with lower CAPEX and OPEX. One 10MW TCO study found immersion cooling cut 10‑year total cost by ~39% vs air, with 30–40% savings in annual OPEX alone.

• Best for: Very high density clusters and greenfield designs where you control hardware and firmware.
• Pros: Highest density, lower mechanical complexity, large energy savings, quiet.

Cons: Hardware compatibility, operational learning curve, and fewer colos offering production immersion today.

• The global liquid cooling for AI data centers market is growing at 30%+ CAGR, driven by AI and accelerator‑heavy deployments.
• Above 30kW per rack, liquid cooling matches or beats air TCO; above 50kW, it can be 20–35% cheaper than air on a full rack basis due to lower PUE and higher density.

• Immersion cooling on 10MW‑scale builds has shown ~40% TCO savings over 10 years compared to air, including both CAPEX and OPEX.

• Colocation operators are racing to add liquid capabilities, but not all projects are live or open to smaller footprints—many early capacities are reserved for hyperscale.