The term “AI-ready” has become one of the most frequently used labels in colocation marketing. Almost every facility now positions itself as a colocation option for AI workloads and claims to support high-density rack hosting, but industry data shows a different picture. Uptime Institute’s 2025 Global Data Center Survey found that average rack density continues to rise, but gradually. Most deployments still fall within the 10–30 kW range, while racks operating above that level remain relatively uncommon. A single modern GPU rack used for training can consume 40–130 kW. The gap between marketing claims and engineering reality is often where expensive mistakes begin when selecting a provider.
This guide is intended for teams preparing to deploy GPUs or other high-density workloads in a colocation facility. It is not an explanation of AI infrastructure components. The power, cooling, and hardware specifications that make up an AI data center are covered separately in Understanding AI Data Centers: Key Specifications and Hardware. Instead, this guide focuses on another part of the process: how to verify, before signing a contract, that a specific facility can actually deliver what its brochure promises, and which questions can distinguish a genuinely AI-ready colocation provider from one that is still fundamentally designed around air cooling.
Why Verification Matters More Than the Label
Demand is growing faster than supply. The IEA’s Energy and AI report from April 2025 projected that electricity consumption from data centers optimized for AI would increase more than fourfold by 2030. The IEA has since revised its projections, but the overall direction remains consistent: demand is growing much faster than available capacity, pushing operators across multiple markets to adapt existing facilities.
This distinction matters because high-density deployments usually involve long-term commitments. GPU deployments commonly run on three- to five-year contracts, and the risks are asymmetric. If a facility cannot deliver the power or heat rejection capacity it promised, the problem cannot be fixed with software. You have only two options: accept degraded hardware performance or throttling, or move the deployment. Migrating an operating GPU cluster can take months and cost millions.
The rule for everything that follows is simple: treat every “AI-ready” claim as unproven until the provider can show the engineering evidence behind it. Operators that genuinely have the capability should be comfortable answering these questions. An evasive answer is also an answer.
Power: Verify the kW Actually Available, Not the Design Capacity
Power is the first and often the hardest constraint, and it is also where marketing terminology is most frequently used creatively. A facility may advertise “design capacity,” “planned capacity,” or “campus capacity,” but none of those figures guarantees that your cage will actually be able to draw the contracted amount of power at go-live.
Ask for these figures in writing rather than relying on explanations given during a site visit. A common scenario is that a brochure states a campus capacity of 20 MW, but once you ask how many kilowatts can actually be delivered to your specific rack position, including the busway configuration and whether the relevant substation or electrical zone is already energized, the answer changes to “to be confirmed during commissioning.” Campus capacity and cage capacity are not the same thing, and this is often where the first “AI-ready” claim starts to fall apart. The same principle applies to adjacent rack availability, reservation schedules, and redundancy at high density. The detailed questions are included in the checklist at the end of this guide.
Determining the right amount of capacity, including how many kilowatts to contract now and how much to reserve for future growth, is a discipline of its own. It is covered in more detail in our data center capacity planning guide for first-time buyers.
Cooling: “Liquid-Ready” Does Not Mean Liquid Cooling Is Already Running
Above roughly 30–40 kW per rack, air cooling starts to become insufficient, and direct liquid cooling (DLC) becomes a critical capability. This is also where the gap between marketing language and operational reality becomes especially visible, because “liquid-cooling ready” can mean almost anything, from “DLC is already running in production today” to “there is space where a CDU could theoretically be installed.”
The simplest way to distinguish between the two is to ask how long DLC has actually been operating in the facility and at what scale, rather than whether it is “planned” or “on the roadmap.” An operator that has already run liquid cooling in production has gone through commissioning, water-quality management, and failure-response procedures. An operator that has never operated it before will effectively be learning with your hardware.
Uptime Institute’s guidance on resiliency considerations for direct liquid cooling can serve as a useful reference for the types of questions a provider should be able to answer confidently, including leak detection and where responsibility sits when an issue occurs in the facility water system versus the technology cooling loop. The detailed questions are included in the checklist below.
Physical Details That Marketing Pages Rarely Mention
High-density deployment failures are not always electrical. Many projects are delayed by logistics and structural limitations instead.
Floor loading is one of the most commonly overlooked issues. A fully populated high-end GPU rack using liquid cooling, such as an NVIDIA GB200 NVL72 system, can weigh around 1,500 kg once coolant is included. Compare that with a conventional raised floor that may only be designed for 500–750 kg per rack position.
Make sure the load rating applies across the entire delivery path, not only at the final rack location. The loading dock, freight lift, corridors, and every part of the route must be able to accommodate a fully assembled cabinet. Rack dimensions also need to be checked. Deep GPU chassis and rear manifolds often require deeper cabinets and wider hot aisles than older rows were designed to support.
Ten minutes spent reviewing structural drawings during a site visit can prevent the worst version of this problem: discovering the limitation on installation day.
Network: The Cluster Solves Only Half the Problem
AI workload performance depends heavily on how quickly data can enter the environment and how efficiently results can leave it. The east-west fabric inside the cluster is part of your own design. Everything outside the cage boundary becomes the facility’s responsibility and should be assessed with the same level of scrutiny as power.
Cross-connect provisioning is frequently underestimated even though it can become a significant source of friction. A multi-site training or inference deployment may require dozens of cross-connects. If the provider cannot commit to provisioning times and pricing in writing, those limitations can quietly become architectural constraints.
The more carriers physically available in the facility, the more competitive your transit pricing can be and the easier it becomes to build redundant paths. Direct access to internet exchanges and cloud on-ramps completes the picture. For inference workloads serving local users, direct peering can shorten the path to end-user networks. For hybrid training pipelines, it can reduce the need to move large datasets over the public internet.
Operations: Who Will Handle Your GPUs at 3 AM?
High-density hardware increases the demands placed on day-to-day operations. GPU servers are heavy, expensive, and poorly suited to improvised handling. Liquid-cooled systems also introduce procedures that many facility technicians may never have worked with before.
Ask specifically how the provider handles remote hands services for GPU-class hardware. Have their technicians been trained to work with liquid-cooled systems and blind-mate manifolds? What is the response SLA at night? Can the provider store and manage spare hardware on-site?
A four-hour response commitment matters much more when a rack is operating at 100 kW than when the deployment only requires 5 kW.
Shortlist: Ten Questions to Bring to Every Site Visit
- How many kW per rack can you deliver to my specific location today, and what evidence can you provide?
- Can you provide adjacent racks at full density for my future expansion, and under what terms?
- Is utility power for my area already energized, or does it still depend on an upgrade?
- Is direct liquid cooling already operating in this facility, and at what scale?
- What are the temperature, flow, and water-quality specifications for the facility water available at my location?
- Where exactly is the boundary between the provider’s responsibility and my responsibility for the cooling system?
- What leak-detection coverage is in place, and what is the response procedure for coolant loss?
- What is the floor-load rating across the entire delivery path to my cage?
- What cross-connect provisioning time can you guarantee, and what is the monthly cost?
- Are your remote hands technicians trained to work with liquid-cooled GPU systems?
Conclusion
AI-ready colocation is a verifiable engineering condition, not simply a label. It means having the kilowatts actually available at your rack, liquid cooling already operating in production rather than existing only on a roadmap, floors and delivery routes capable of supporting the hardware load, cross-connects backed by clear SLAs, and an operations team trained to handle the hardware you are actually deploying.
A provider that genuinely meets these requirements should be able to demonstrate evidence for each one during a site visit and through contractual documentation. Treat the checklist in this guide as a minimum standard. A facility that can meet it may be able to support your high-density deployment for years. A facility that cannot may ultimately cost far more than the difference in monthly colocation pricing.
If you are evaluating facilities in Jakarta for a GPU deployment or other high-density workloads, contact the Digital Edge Indonesia team to discuss delivered rack density, liquid-cooling readiness, and interconnection options at EDGE1, EDGE2, and the CGK Campus, supported by engineering evidence for each requirement.
