The Questions You Need to Ask Before You’re Locked In
Signing a colocation contract is easy. Signing the right one is a different matter entirely.
As enterprises move AI workloads into colocation facilities, the stakes of that decision have risen considerably. A standard data center built for traditional enterprise IT simply cannot support the power density, cooling demands, and network performance that modern AI infrastructure requires. Choosing the wrong facility doesn’t just create operational headaches; it can stall entire AI programs before they produce a single result.
This checklist covers what actually matters before you commit.
Power Density: The First Filter
AI workloads consume power at a scale that most colocation facilities weren’t originally designed to handle. A standard enterprise server rack draws somewhere between 5 and 10 kilowatts. A rack dense with NVIDIA H100s or AMD MI300X accelerators can push 30 to 80 kilowatts or more. That’s not a marginal difference; it’s a fundamentally different infrastructure requirement.
Before signing anything, validate these specifics with the provider: What is the maximum power density per rack, expressed in kilowatts? Can the facility support 30kW, 50kW, or higher per cabinet if your workloads require it? Is high-density power available today or only in future buildout phases? What redundancy model governs power delivery, N+1, 2N, or something else? And critically, how does the facility handle power allocation across tenants to prevent contention?
If a provider is vague on kilowatts per rack or defaults to averages, that’s a signal worth paying attention to.
Cooling: Where Many Facilities Fall Short
Power density and cooling are inseparable. Heat generated by AI accelerators at high utilization is extreme, and traditional air cooling hits its limits quickly at densities above 20 to 25 kilowatts per rack.
The questions that matter here are direct. Does the facility offer liquid cooling, including direct-to-chip or rear-door heat exchanger options? What is the maximum thermal load the cooling infrastructure can handle per rack and per hall? How does the facility maintain temperatures during peak load conditions? What happens during a cooling system failure and is there documented failover with a clear activation timeline?
Liquid cooling is no longer a premium option reserved for hyperscalers. For serious AI workloads, it’s increasingly a baseline requirement.
Interconnects and Network Performance
AI training workloads, particularly distributed training across multiple nodes, are extremely sensitive to network latency and bandwidth. The interconnect infrastructure inside a colocation facility can be the difference between efficient model training and a bottleneck that wastes expensive GPU time.
Validate the following before committing: What fiber and network infrastructure is available within the facility? Does the provider support low-latency cross-connects to cloud on-ramps like AWS Direct Connect, Google Cloud Interconnect, or Azure ExpressRoute? What are the available uplink speeds and are they guaranteed or best-effort? If you’re running multi-node GPU clusters, does the facility support the high-speed fabric interconnects those configurations require?
SLAs: Read What’s Actually Promised
Service level agreements are where colocation contracts reveal their real quality. A headline uptime figure of 99.999% sounds reassuring until you look at what’s actually covered and what remedies exist when things go wrong.
Push for clarity on these points: Does the SLA cover power, cooling, and network separately or as a combined metric? What is the actual financial remedy for downtime and is it proportional to the business impact or just a token credit? How is downtime defined and measured, and who does the measuring? What are the provider’s documented response times for critical incidents?
An SLA that promises high availability but offers a minor service credit as the only remedy for an outage is not a strong SLA.
Compliance and Data Sovereignty
Depending on your industry and the data your AI systems process, the physical location and compliance posture of a colocation facility may be non-negotiable.
Confirm which compliance certifications the facility holds, including SOC 2 Type II, ISO 27001, HIPAA-eligible infrastructure, and PCI-DSS, and verify they cover the specific cage or suite your workloads will occupy, not just the facility broadly. Understand the geographic jurisdiction of the facility and how that interacts with any data residency obligations your organization operates under. Ask whether the provider has experience supporting regulated workloads and can provide documentation to satisfy your own auditors.
Resiliency and Physical Security
Evaluate the facility’s resiliency posture honestly. What is the Tier rating of the facility and what does that mean practically for your uptime requirements? How many independent utility feeds supply the building? What is the on-site backup generation capacity and how long can it sustain full load? What physical security controls govern access to your cage or suite?
The Decision Behind the Checklist
Colocation is a long-term commitment. Facilities that meet today’s requirements but have no path to higher power density or advanced cooling will create a forced migration problem as AI workloads grow. The right question to ask any provider isn’t just whether they can support your current deployment; it’s whether they can support where your AI infrastructure will be in three years.
At Vertical Data, we help enterprises evaluate and structure colocation decisions alongside the financing arrangements that make deploying AI infrastructure at scale actually feasible.
