How to Choose Cooling for Mixed-Density Data Halls
Unlock the secrets to effective cooling strategies for mixed-density data halls and enhance AI performance.
The pitch is seductive in its simplicity: AI means high density, high density means liquid cooling, therefore every data center needs liquid cooling. Problem solved, article written, budget approved.
Except that's not how data centers actually work.
Most facilities running AI workloads today aren't monolithic GPU clusters humming at 100kW per rack. They're mixed environments — legacy compute sharing floor space with inference nodes, storage arrays sitting next to training clusters, and general-purpose servers handling everything from ERP to dev environments. The real challenge isn't choosing a cooling architecture for an AI-native greenfield build; it's choosing one that works across a hall where rack densities might swing from 5kW to 50kW within thirty feet of each other.
That's the problem worth solving.
Understanding the Thermal Landscape
Before you commit capital to a retrofit or break ground on a new build, you need an honest assessment of where each cooling method physically breaks down — not where vendors say it starts struggling, but where it actually stops working.
Standard air cooling on a raised floor handles roughly 5kW to 15kW per rack under normal operating conditions. It's low complexity, well understood, and your operations team can troubleshoot it at 2 a.m. without calling a specialized contractor. But push past that range, and you're fighting physics. Hot aisle containment and in-row cooling can stretch that ceiling toward 20kW to 25kW, but only if your airflow paths are engineered correctly from the start — and in most retrofits, they aren't.
Rear-door heat exchangers (RDHx) represent a meaningful middle layer that doesn't get enough credit in the cooling conversation. A passive RDHx can absorb residual heat from racks pushing 20kW to 30kW without requiring facility-level plumbing changes that turn into six-month projects. For data centers with existing chilled water infrastructure, this is often the fastest path to handling moderate density increases. It's not glamorous, but unglamorous solutions that actually ship on schedule have a way of looking brilliant in retrospect.
Direct liquid cooling — cold plates, immersion, rear-door active systems — is where the real density ceiling gets pushed. Cold plate systems can handle 50kW to 100kW per rack. Full immersion can go further. But liquid cooling introduces operational complexity that doesn't appear in the vendor's slide deck: specialized maintenance procedures, fluid compatibility requirements, longer mean time to repair, and a skills gap that will cost you real money if you haven't planned for it.
Thermal profiling isn't optional. It's the work that has to happen before any architecture decision. Map your actual rack densities today, model where they're going over a 36-month horizon, and overlay that against your facility's power and cooling infrastructure constraints. Gut instinct about where your workloads are heading is not a thermal profile.
The Mixed-Density Problem Is Harder Than It Looks
A purely high-density facility is actually a simpler engineering problem than a mixed one. You pick a cooling strategy optimized for one density range, engineer accordingly, and move on. Mixed-density halls are harder because no single architecture is optimal across the full range of equipment that's actually on the floor.
Consider what a mid-size enterprise data center looks like in 2025: rows of legacy servers at 8kW to 12kW per rack handling traditional workloads, a growing cluster of GPU nodes for inference at 30kW to 40kW per rack, and storage infrastructure that runs cool but demands precise airflow management. That's three different thermal profiles in one hall, and the cooling system has to serve all three simultaneously.
The instinct is to engineer for the worst case — design liquid cooling infrastructure for the 40kW racks and let air handle the rest. The problem is that liquid cooling infrastructure is expensive to install even where you don't need it, and over-investing in cooling for racks that never exceed 15kW is capital that could have gone elsewhere.
Zones are the answer most experienced operators land on, but zoning a live data center without disrupting operations is genuinely difficult work that tends to get underestimated in the planning phase.
The practical approach is modular containment: air-cooled zones for legacy infrastructure, a dedicated high-density zone with liquid cooling for AI compute, and a transition strategy for equipment that moves between categories as workloads evolve. This requires coordination between facilities, IT, and operations teams that often don't share the same planning cycles — which is usually the real constraint.
What Actually Drives the Decision
Three factors end up mattering more than any technical specification when organizations make real cooling decisions.
Operational capability is the one that bites hardest. Liquid cooling systems — particularly immersion — require maintenance procedures that most data center operations teams haven't performed before. The fluid needs to be monitored, filtered, and eventually replaced. Leaks in an immersion tank are a different class of problem than a failed CRAC unit. If your team isn't trained for it, your MTTR numbers will look very different in practice than they did in the business case.
Existing facility infrastructure determines what's actually achievable in a retrofit context. Chilled water capacity, electrical headroom, floor load ratings, and raised floor depth all constrain your options before the first vendor conversation happens. Greenfield builds have design freedom; most organizations are working with facilities built to 2010 assumptions trying to handle 2025 workloads.
Power density trajectories matter because the system you install today needs to handle the racks you'll be running in three years. NVIDIA's next-generation compute platforms and AMD's accelerator roadmaps both point toward continued density increases. A cooling architecture that's barely adequate for today's AI hardware may be genuinely inadequate for the next generation — and ripping out and replacing cooling infrastructure mid-cycle is expensive in both capital and downtime.
Where the Industry Is Actually Heading
The cooling industry has moved fast in the last two years, driven by hyperscaler requirements that eventually trickle into enterprise and colocation design standards. A few trajectories are worth tracking.
Rear-door heat exchangers are getting a second look from operators who want to bridge the gap between air and full liquid cooling without committing to the operational complexity of immersion or cold-plate systems. For facilities with mixed-density halls and limited retrofit budgets, RDHx can extend the life of an air-cooled infrastructure by several years while higher-density zones get liquid cooling investment.
Warm water cooling — running facility water at higher temperatures than traditional chilled water systems — is becoming more viable as hardware manufacturers increase their thermal tolerances. Running coolant at 35°C instead of 15°C dramatically reduces the energy required for cooling and opens up free cooling opportunities in moderate climates for more of the year. For operations teams focused on PUE reduction, warm water cooling often delivers more measurable improvement per dollar than more exotic approaches.
AI inference workloads, as distinct from training workloads, introduce a variable density challenge that static cooling infrastructure handles poorly. Inference loads can spike dramatically based on query volume, which means rack-level heat output fluctuates in ways that training clusters — running at near-constant utilization — don't. Dynamic cooling response systems that can adjust capacity at the rack level in near-real-time are an emerging category worth watching.
What Good Decision-Making Looks Like
The operators making smart cooling decisions right now share a few common practices. They're doing actual thermal profiling instead of estimating. They're engaging facilities and IT teams in the same planning process instead of handing down requirements from one group to the other. They're building cooling architecture that accommodates multiple density tiers rather than optimizing for one.
And they're resisting the pressure to over-engineer. Not every rack that will ever touch an AI workload needs to sit behind liquid cooling infrastructure. Matching cooling investment to actual density requirements — current and projected — is the discipline that separates facilities running efficiently from facilities that spent aggressively on capability they don't use.
The question to answer before any cooling decision is straightforward: What does your actual thermal map look like, and where is it realistically going? Everything else follows from that.
Call to Action
Ready to optimize your data center cooling strategy? Explore our marketplace for innovative solutions: InfraSale Marketplace.