Colocation is space and power for hardware you own. You bring the servers. The site provides the floor, the electricity, the cooling, the physical security, and the hands that can touch your equipment when you can't be there.
What you bring and what the site provides
You bring the compute: GPU servers, storage, and any networking gear that connects them. Everything else is the site's job.
What the site provides:
- Floor space and racks. Your servers live in steel rack cabinets, either shared with other tenants (a "cage" or a row of racks) or in a private suite. The unit of space is the rack, measured in rack units, but you order in rack count, not rack units.
- Power. The site takes utility power, conditions it, backs it up with generators and batteries, and delivers it to your racks. You order power in kilowatts (kW) or megawatts (MW). One kilowatt is one thousand watts; one megawatt is one thousand kilowatts.
- Cooling. Servers produce heat. The site removes it, either by pushing cold air through the hall or by running water to the rack. The method matters because modern GPU servers can produce far more heat per rack than a traditional air-cooled hall is built to handle.
- Physical security. Locked cages, badge access, cameras, and sometimes biometrics. Your hardware is inside a controlled facility, not a warehouse.
- Remote hands. Site staff who can swap a cable, reseat a drive, or cycle power on your instruction. This is the human layer that makes remote management practical.
- Network access. The site connects to carriers and internet exchanges. You choose your own network provider and order a cross-connect, which is a physical cable run inside the building from your rack to the carrier's equipment.
How a server count becomes a rack-and-kW order
An operator needs two numbers from you: how many racks and how many kilowatts. You get there from your server count.
Start with power. A standard 8-GPU server running H100 or H200 draws about 10 kW. A server running B200 GPUs draws about 14 kW. Those are planning figures per 8-GPU server.
A concrete example: say you are deploying 50 H100 servers. At 10 kW each, that is 500 kW, or 0.5 MW, of IT load, the power your servers themselves consume, before the site adds its own cooling and distribution overhead. The site's own cooling and power distribution adds overhead on top of that, which the site will account for in its design, but 500 kW is the number you bring to the conversation.
For rack count, a typical air-cooled rack holds two or three of these servers before it hits the rack's power or airflow limit. At two servers per rack, 50 servers fills roughly 25 racks. At three, about 17. The exact number depends on the site's per-rack power limit, which varies. Ask the operator what their standard rack power cap is and work backward from your per-server draw.
If you are running newer liquid-cooled hardware, the numbers change significantly. A GB300 NVL72 rack, for example, draws up to 142 kW for a single rack. That kind of density requires a site that can deliver water to the rack, not just cold air to the room. The liquid vs. air cooling explainer covers that in detail.
Tier ratings and what they mean for you
Data centers are rated by the Uptime Institute on a Tier I to Tier IV scale. Tier III is the floor for serious AI workloads. It means the facility can be maintained without taking your systems offline, and there is no single component whose failure brings the building down. If an operator cannot confirm Tier III design or equivalent, that is a risk to your uptime.
What to tell an operator when you ask for space and power
Before you get on a call, have these five things ready:
- Market. Which city or region. Power availability and lead times vary enormously by location.
- Power requirement. Your total IT load in kW or MW, derived from your server count and per-server draw.
- Cooling method. Air-cooled, rear-door heat exchanger (a water-cooled coil fitted to the back of a rack that captures heat before it reaches the room), or direct liquid to the rack. This determines which sites can physically support your hardware.
- Timing. When you need the space ready. Build-out and power provisioning take time, and the gap between signing and go-live is often longer than teams expect.
- Term. Know the term you can commit to before you ask for pricing; XIRR brokers colocation on terms of 36 to 60 months. Know the term you can commit to before you ask for pricing.
If you are still deciding between running your own hardware in colocation and contracting reserved GPU capacity from a neocloud operator, a company that owns GPU clusters and sells access to them by the month or year rather than by the hour, the colocation vs. reserved capacity guide walks through that decision.
Sources
Per-server power figures for H100, H200, and B200 8-GPU servers are planning figures used across the industry. The GB300 NVL72 rack power figure of 142 kW maximum is from NVIDIA's reference architecture. Tier III design standards are defined by the Uptime Institute. The liquid cooling background draws on general data center engineering practice.
Share your requirements. We'll canvas the market.
Tell us what you need (for colocation space and power: market, kW or MW, cooling and timing; for reserved GPU capacity: region, GPU type, cluster size and timing) and we'll canvas colocation and neocloud operators in parallel. Shortlist in 48 hours. We don't sell hardware. We broker colocation space and power, and reserved GPU compute capacity.
Earlier conversations get better terms. When you engage early, we have time to negotiate with vendors before you need to commit. We're paid only when a deal closes, and the fee structure is agreed in writing before any introduction.