AI has rewritten the colocation checklist. Power, density, cooling and proven readiness now decide whether a deployment goes live, and here is how enterprises should evaluate a provider before signing.
Choosing a colocation provider for AI is not the same as choosing one for traditional servers. The old checklist misses the things that now decide whether a deployment works.
Colocation, renting space and power in someone else’s data center, remains one of the fastest ways to get an AI cluster running without building a facility of your own. But AI has changed what a good colocation partner looks like. A provider that was perfect for enterprise applications a few years ago may be unable to power a single modern AI rack.
The reason is that AI workloads concentrate far more power and heat in far less space than the workloads these facilities were designed for. Evaluating a provider on the old terms, square footage and price per square foot, will lead you to sites that look ready and are not. This guide walks through what actually matters, and ends with the questions to confirm before you sign.
How Much Power Can the Facility Actually Deliver?
Power is the first thing to confirm and the most common place where a site falls short. The question to ask is how much contracted power is available per rack today, not on a roadmap, and whether that power is energized now or waiting on a utility connection. A facility can have an entire empty hall and still deliver only a fraction of the power that hall could physically hold at AI density. In constrained markets, new grid connections can take years, so available soon is not the same as available.
How Dense Can the Racks Go?
Density is how much power and equipment a single rack can hold. Traditional racks ran at 5 to 15 kilowatts. Current AI systems run at 80 to 140 kilowatts per rack, and the next generation goes higher. A provider needs to state the maximum density it can support per rack and per row, because a facility built for 15 kilowatt racks cannot reach AI density by adding a few circuits. That change runs from the utility feed all the way to the rack, and in practice it is a rebuild.
Is the Cooling Already Installed?
Above roughly 40 kilowatts per rack, air cooling can no longer carry the heat away, and liquid cooling becomes a requirement rather than an option. The important question is whether liquid cooling is installed today or merely planned. Coolant distribution, leak detection and compatible racks cannot be added quickly to a facility that lacks them. Cooling is a decision made when a hall is built, so a provider that has not already committed to it is offering a construction project, not a ready site.
Can the Network Keep Up With the Cluster?
AI clusters move enormous volumes of data between servers during training and between the facility and users during inference. A provider should offer high-bandwidth, low-latency connectivity inside the facility and multiple carrier options out of it. It is worth asking about the network fabric between racks, the available carriers, and how the site connects to the regions where your users actually are. Networking planned late becomes a bottleneck that fast hardware cannot overcome.
Is the Timeline Real or Aspirational?
An availability date only means something if the power behind it is contracted and the cooling is installed. A promised move-in date that depends on a utility connection still in queue is not a timeline, it is a hope. The strongest providers can put you into high-density space in weeks because the hard parts are already done. The test is simple: ask the provider to tie its date to secured power and installed cooling.
Does the Provider Reach the Regions You Serve?
Where a facility sits affects latency for your users, the rules that apply to your data, and how quickly you can expand. If your users or operations are spread across regions, a provider with a single location will eventually become a limit. Reach also matters for compliance, since some data must be stored and processed within a specific country. A provider with facilities in the regions you serve saves you from stitching together several vendors later.
Can the Site Grow With You?
The cluster you need today is rarely the cluster you need in two years. It is worth asking whether the provider has room to grow with you, both in additional racks and in support for denser hardware as it arrives. A site with no headroom means your next expansion is another search, another contract, and another migration.
What Questions Should Enterprises Confirm Before Signing?
Due diligence for AI colocation comes down to a short list of questions, and a weak answer to any one of them points to where the delay will come from.
How much contracted power is available per rack today, stated in kilowatts rather than square feet, and is it energized now or waiting on a utility connection?
What is the maximum rack density the facility supports, both per rack and per row?
Is liquid cooling installed and operating today, or does it still require a build-out?
Does the facility offer high-bandwidth networking inside and multiple carriers out, connecting to the regions your users are in?
Is the move-in date backed by secured power and installed cooling, or does it depend on a connection still in queue?
Is there room to add racks and support denser hardware as the deployment grows?
What Actually Separates AI-Ready Colocation
The right AI colocation partner is the one that can prove readiness, not just describe it. Power, density, cooling and timeline are the four points where ready sites separate from legacy ones, and every question above traces back to them.
Vertical Data helps enterprises evaluate, place and deploy AI infrastructure in colocation facilities that are ready now, so these questions are answered before a contract is signed rather than discovered after.
