GPU Colocation: What It Is and What to Check Before You Sign

Most pages about GPU colocation are written by providers selling it. That is fine, but it leaves out the parts a buyer needs most: where the power comes from, how firm it is, and how the lease lines up with the life of your hardware.

This guide covers what a colocation data center is, what changes when the racks are full of GPUs, how colocation compares with renting GPUs in a cloud and with building your own site, and a checklist to put to any provider. It ends with an honest note on where Accretion fits, which is not everywhere.

What is GPU colocation?

GPU colocation is renting room, power and cooling in someone else's data center for GPU servers you own. You buy the hardware, ship it in, and run your own software on it. The provider runs the building.

To answer the wider question first: what is a colocation data center? It is a facility that houses many customers' equipment and supplies the physical layer around it. People shorten it to a colo data center, or just colo. The split of jobs usually looks like this:

• You supply: the servers, storage and network gear, the operating system and everything above it, and your own staff or contractors for anything beyond basic tasks.
• The provider supplies: the building and security, power delivered to your racks, backup power, cooling, connections to fiber carriers, and on-site staff who can do simple physical work for you (remote hands).

Colocation is sold in very different sizes. Some customers rent part of a rack or a single cabinet. Others lease a whole hall or a whole building, often called wholesale colocation. A colocation GPU server deployment for an AI company is usually at the larger end: many racks, sized as one cluster, on a multi-year contract.

What is high density colocation, and why do GPUs need it?

High density colocation is colocation built for racks that draw far more power, and give off far more heat, than ordinary servers. GPU servers for AI training and inference are the main reason it exists, which is why you also see it called AI colocation.

A hall built for web servers and storage was designed around modest power per rack. Fill those racks with modern GPU servers and three things break at once:

• Power delivery. Each rack needs much more electricity than the busbars, breakers and feeds were sized for. Many older halls can only take a GPU rack by leaving the racks around it empty.
• Cooling. Air cooling reaches its limit. Dense GPU racks often need liquid cooling: cold plates on the chips (direct-to-chip), rear-door heat exchangers, or immersion. Each needs pipes, pumps and heat rejection designed in from the start.
• Floor and network. Dense racks are heavy, and GPU clusters need fast links between servers, so cable paths and floor loading matter more than in an ordinary hall.

There is no industry-standard line where "high density" starts. Providers use different thresholds, and the numbers keep moving as each hardware generation draws more. Ask for the per-rack figure the provider will put in the contract. The marketing figure doesn't count.

GPU colocation vs cloud GPUs vs building your own

There are three common ways to get GPUs running. They differ in who owns the hardware, how fast you start, and how long you are committed.

Cloud GPUs (GPU-as-a-service). A cloud provider owns the GPUs and the building. You rent instances by the hour, or reserve them for a term. You can be running today, and you can stop when the work stops. You pay a margin on the hardware for that flexibility, and you get the hardware, network and location the provider offers.

Colocation. You own the GPUs; the provider supplies the building, power and cooling. You pay for hardware upfront (or finance it) and pay the provider per unit of power per month. You control the hardware choice, the network design and the software. You need lead time for the space to be ready and for your hardware to arrive.

Building your own. You own the land, the building, the power supply and the hardware. You get the most control and carry all the risk: site selection, permits, a power connection, construction, and staff to run it. Read how long it takes to build a data center before you plan around this option.

You can mix them: cloud GPUs for bursts and experiments, colocation for the steady base load you know you will run for years.

What should you check before signing a GPU colocation lease?

Ask each provider the same questions in writing and compare the answers line by line.

1. Power per rack, in the contract. What the provider commits to deliver to each rack, and what happens if you need more in year three when you refresh hardware.
2. Cooling type. Air, liquid or both. If liquid: which method, who supplies and maintains the cooling equipment inside the rack, and what water or coolant spec your servers need.
3. Power source and how firm it is. Where the electricity comes from (the grid, on-site generation, or both), whether supply can be interrupted or curtailed, and what backup is in place. Firm power means steady, contracted and backed up. Ask the same question about the fuel or supply contract behind it.
4. Room to expand. Whether there is space and power on the same site for your next phase, and whether you get first right to it. Splitting one cluster across sites is expensive.
5. Fiber. Which carriers reach the site, how many separate routes, and how cross-connects are priced.
6. Remote hands. Who is on site, what hours, what tasks are included, and how fast they respond to a failed server or cable.
7. Lease length vs hardware life. A lease much shorter than the useful life of your servers forces a move mid-life. A lease much longer leaves you paying for space after the hardware is retired. Aim to line them up, with options to extend.
8. Delivery date and what happens if it slips. For space that is not built yet, ask what the provider's commitment is and what remedy you have if it is late.

Also read how power is billed. Some providers bundle energy into one rate; others charge for capacity and bill the energy you use separately. Know which you are comparing.

Why is power now the hard part of AI colocation?

Space and cooling can be built. Power is the slow part, because large new loads have to wait for a grid connection. That wait is the main reason high-density capacity is scarce in many markets.

Demand is rising fast. A US Department of Energy summary of a Lawrence Berkeley National Laboratory report found that data centers used about 4.4% of total US electricity in 2023, or 176 TWh. The report expects that to reach roughly 6.7 to 12% by 2028, or 325 to 580 TWh (DOE, December 2024). For a sense of what that means at one site, see how much power a data center uses.

New grid connections for loads of this size join a line of other projects waiting for studies and upgrades. The data center interconnection queue explains how that line works and why it can take years.

One response is to build where the energy already is and generate power on site, so the data center does not wait on the grid. That is called behind-the-meter power. For a colocation buyer, that adds questions. Beyond "when does the utility connect you?", ask what fuel runs the generators, under what contract, and what backs it up.

When is GPU colocation the wrong choice?

Colocation is a long-term commitment to your own hardware. It is a poor fit in these cases.

• You need GPUs running this month. Rent them in a cloud, or take space in an existing colocation hall that is already built and powered.
• You need them for weeks or months, not years. Cloud GPUs let you stop paying when the project ends.
• Your demand is uncertain. If you can't forecast your load a few years out, owning hardware and signing a long lease adds risk you may not need.
• You have no one to run the hardware. Colocation hands you the building, not the servers. Someone on your side still installs, monitors and fixes the cluster, or you pay a managed service to do it.
• You need a small amount. A handful of servers is better served by retail colocation or the cloud than by a dedicated build.

If none of those apply and you know your load, owning the hardware and leasing power-dense space for its full life is often the steadier option.