• by: Bisola Orimalade
• 0 Comment(s)

What the Shift to AI Workloads Means for Colocation Density and Cooling

The colocation model has always been straightforward. You rent space, power, and cooling capacity. The facility handles the infrastructure. You handle the kit.

That model still holds, but the physics have changed. AI workloads generate heat at a scale that traditional colocation was not designed to manage. Operators unprepared for this gap are learning the hard way through throttled performance, unplanned capital spend, or contracts that fail to deliver as expected.

This article sets out what is changing, what it means in practical terms, and where infrastructure decisions tend to go wrong.

The Density Problem in Plain Terms

Traditional enterprise racks run at somewhere between 5 kW and 15 kW. That is the range most colocation facilities were built around.

AI infrastructure is a different category. GPU clusters for training or inference can run at 30 kW to 100 kW per rack. The latest generation of NVIDIA-based GPU servers requires 132 kW, and the next generation is expected to push beyond 240 kW per rack.

The physics of a data centre are unforgiving: almost every watt of power consumed becomes heat that must be removed. Double the power density, and you double the thermal load. A facility designed to cool 10 kW racks cannot cool 80 kW racks just by running the air conditioning harder.

This is not a future problem. Operators deploying AI workloads in colocation facilities today are running into ceiling constraints that were invisible when those facilities were specified.

Why Air Cooling Has a Hard Ceiling

Air cooling works by moving large volumes of air across hot equipment. The problem is that air has a limited capacity to carry heat. Beyond about 15 kW to 20 kW per rack, traditional raised-floor cooling with hot and cold aisle containment struggles. Hot spots develop, airflow becomes turbulent, and fan energy consumption rises significantly.

Containment strategies, in-row cooling units, and optimised airflow management can extend the range to moderate AI densities. But they do not solve the fundamental constraint. At 50 kW, 80 kW, or 100 kW per rack, you need to remove heat closer to the source.

This is where liquid cooling enters the picture, specifically rear door heat exchangers, direct-to-chip cooling, and immersion cooling, depending on the density and workload profile.

Rear Door Heat Exchangers

For facilities needing to increase density without major infrastructure changes, rear door heat exchangers are the most practical near-term option. They attach to the back of server racks and intercept exhaust air directly, using chilled water to remove heat before it reaches the room. They support higher rack densities within existing white space without a full retrofit of the cooling plant.

For many colocation operators moving into moderate AI density, this is the first step.

Direct-to-Chip and Immersion Cooling

Higher densities require cooling closer to the silicon. Direct-to-chip cooling runs chilled liquid directly to a cold plate on the processor. Immersion cooling submerges servers in dielectric fluid. Both approaches can support the extreme thermal loads that GPU clusters generate, but they require purpose-built infrastructure and different operational procedures.

These are not drop-in solutions. The decision between them depends on workload type, existing facility design, and the operator's long-term density roadmap.

What Colocation Operators Are Getting Wrong

Most of the difficulty is not technical. It is contractual and commercial.

Assuming power and cooling are aligned

Many colocation contracts specify power in kilowatts per cabinet, but not how that power is cooled. An operator can procure a 30 kW cabinet and find that the facility's cooling infrastructure was not built to handle sustained loads at that density. The power is there. The thermal capacity is not. Treating all AI workloads the same, inference and training workloads behave differently. Training runs sustain near-peak GPU utilisation for extended periods. Inference workloads are more variable. The cooling requirement for a sustained training cluster differs from a mixed inference deployment. Getting this wrong at planning leads to over-specified spend or under-specified cooling.

Not asking about the cooling roadmap

Colocation facilities have different strategies for handling density growth. Some are investing in liquid cooling infrastructure now. Others manage with enhanced air cooling. Before committing to a facility, operators should ask specifically what the provider's cooling roadmap looks like at 50 kW, 80 kW, and 100 kW per rack. A vague answer is a signal worth taking seriously.

Underestimating lead times

Cooling infrastructure changes in a colocation facility do not happen quickly. If you plan to deploy AI workloads in 12 months, the facility needs to be specified now. Delays between design, procurement, and commissioning are common reasons AI infrastructure projects run over schedule.

The Commercial Dimension

Power and cooling decisions in colocation are not just operational. They have a direct impact on cost.

Cooling accounts for 30 to 40 percent of a data centre's total electricity consumption. As rack densities rise, that proportion increases unless more efficient cooling methods are used. Rear door heat exchangers and liquid cooling solutions reduce the energy needed to manage a given thermal load, which affects PUE and operating costs over the deployment's life.

For AI operators with large, sustained workloads, the energy cost of cooling is a significant part of the total cost of ownership. Getting the cooling architecture right at the outset is not a technical nicety but a financial decision.

There is also the question of colocation pricing. Facilities that can genuinely support high-density AI workloads are commanding a premium. CBRE data from H1 2025 shows that colocation pricing for 250 kW to 500 kW deployments reached 184 US dollars per kilowatt per month, with requirements above 10 MW seeing price increases of up to 19 percent in the same period. Operators who understand the infrastructure specification in detail are better positioned to evaluate whether that premium is justified.

How QIS Supports Infrastructure Decisions

QIS works with data centre operators, AI and HPC operators, and technology investors as an infrastructure partner, not a one-size approach to product sales.

Our work covers cooling specification, product sourcing, and infrastructure planning across a range of colocation and on-premises environments. We work with trusted manufacturing and supply partners to source the right equipment for the density and workload profile in question, whether that is rear door heat exchangers for an existing facility or guidance on liquid cooling infrastructure for a new deployment.

We do not tie recommendations to a single product line. The starting point is always the infrastructure problem, not the product.

Interested in Working Together?

If you are planning an AI or HPC deployment, assessing a colocation facility for density readiness, or working through a cooling upgrade decision, QIS can provide independent guidance and access to the right infrastructure products.

Contact the QIS team: queensburyinfrastructuresolutions.com/contact

Frequently Asked Questions

Can a standard colocation facility support AI workloads without major changes?

It depends on the density required. Many standard facilities can support moderate AI inference workloads with enhanced air cooling or rear door heat exchangers. Training clusters at 80 kW to 100 kW or more per rack typically require purpose-built cooling infrastructure. The key question to ask any colocation provider is what their maximum sustained power density per cabinet is, and how that power is cooled at peak load.

What is a rear door heat exchanger, and when does it make sense?

A rear door heat exchanger is a cooling unit that fits directly onto the back of a server rack. It uses chilled water to remove heat from the exhaust air before it enters the room. It is a practical solution for operators who need to increase rack density in an existing facility without a full infrastructure retrofit. It works well for moderate AI density, typically up to around 30 kW to 50 kW per rack, and can be deployed without taking a facility offline.

How do I know whether a colocation provider is genuinely AI-ready?

Ask for specifics rather than marketing claims. Request the maximum sustained kW per cabinet the facility can support, what cooling method is used at that density, and what the provider's roadmap looks like for higher densities over the next 24 to 36 months. Ask whether liquid cooling infrastructure is available or planned. A provider that is genuinely AI-ready will be able to answer these questions with technical detail, not general assurances.

Is liquid cooling always necessary for AI workloads in colocation?

Not always. The right cooling approach depends on the density and workload profile. Lower-density AI inference deployments can often be managed with enhanced air cooling and containment. It is at higher densities, particularly sustained GPU training workloads above 30 kW to 50 kW per rack, that liquid cooling becomes necessary rather than optional. The decision should be made based on the actual thermal load of the deployment, not on assumptions about what AI workloads require in general.