Engineering Notes on Prefabricated Modular Data Centers and Site Heat Rejection
North American high-compute data center primary-market vacancy hit a record-low 1.4% in H1 2026, even as supply grew to a record 10,903 MW. With 80.4% of the capacity under construction already pre-leased, the pressure to build and deliver faster is immense.
On the utility side, projects requiring new transmission or generation can wait 24 to 48 months (or more) for grid interconnection. By the time the power is finally delivered, the building phase must be completed and ready to operate. That is why factory-built modules (Prefabricated Modular Data Centers, or MDCs) are increasingly being specified.
But modules only fix the construction schedule. They do not fix the physics of rejecting megawatts of heat on a small concrete pad. This post covers both the schedule and the physics, complete with the numbers.
What Prefabrication Actually Changes on the Critical Path
In a traditional "stick-built" job, Mechanical, Electrical, and Plumbing (MEP) work is sequential and happens in the field: set the equipment, run the pipes, pull the wire, and then test. With prefabrication, that work moves into a factory and runs in parallel with site civil work, utility trenching, and permitting.
The Prefabricated Modular Data Center by Numbers:
Standardized, pre-assembled power and cooling modules deploy at least 60% faster than traditional infrastructure.
They offer first-cost savings of 15% or more.
Breakers, transfer switches, VFDs, Dry coolers, and chillers can be tested as a complete system in the factory and certified by UL, ETL, or CSA. This can compress design time from about 24 weeks down to 12.
Once a design is completed, it can be replicated and scaled many times over. This effectively removes design time from future projects, representing a significant reduction in the overall data center construction timeline.
In practice, the biggest benefit is that Factory Acceptance Testing (FAT) replaces months of tedious field troubleshooting. When the module arrives on-site, the only remaining risks are the final site connections: utility, fluid, and controls.
There is one catch: The design must be frozen before fabrication starts. A late change costs significantly more in the factory than it would in the field.
The Load: One 1.44 MW Block at Current Rack Density
To understand the heat problem, let's look at a modern AI rack. A Lenovo NVIDIA GB300 NVL72 is rated at 135 kW nominal Thermal Design Power (TDP) and can draw up to 155 kW at peak. HPE recommends provisioning busway for 192 kW to be safe. Lenovo notes that the rack heat split is roughly 90% to liquid cooling and 10% to ambient air.
For a 1.44-megawatt (MW) critical IT block, that equals about ten racks. Here is how the math works out:
Heat to liquid: 0.90 × 1,440 kW = 1,296 kW
Heat to air: 0.10 × 1,440 kW = 144 kW
Primary loop flow: Assuming a 10 K (18°F) temperature rise on the water, the mass flow rate is: M = 1,296 / (4.18 × 10) ≈ 31 kg/s ((water density ρ ≈ 4.18 kg/m³) or roughly 490 gallons per minute. When using Glycol mixes depends on the Glycol concentration, it will require even higher flow rates.
Facility water temperature: The maximum coolant inlet temperature is 45°C (113°F). At this warm temperature, chiller-less heat rejection using "dry coolers" (radiator-like fluid coolers) works across most U.S. climates. This high temperature tolerance is what makes a compact, prefabricated thermal plant possible in the first place.
Heat Rejection: The Air Side Is Larger Than Most Expect
Let's assume about 1.5 MW of total heat rejection per block (the IT load plus the heat generated by networking, pumps, fans, and system losses). If we aim for a 10 K air-side dT (temperature rise across the dry coolers):
Air mass flow: M = 1,500 / (1.006 × 10) ≈ 149 kg/s
Volume: At an ambient 35°C (where air density ρ ≈ 1.10 kg/m³), that equals about 135 m³/s, or roughly 287,000 CFM (Cubic Feet per Minute) per block.
If you put two blocks' dry coolers on the roof of a module measuring about 1,800 ft² (167 m²), you have to reject about 2.9 MW of heat. That works out to roughly 17 kW of heat per square meter of roof. For context, peak solar irradiance on a hot summer day is only about 1 kW/m².
The discharge from these coolers is a concentrated, buoyant, high-momentum plume of hot air. What that plume does after it leaves the fans determines whether the cooling plant will actually meet its rating.
Why Hot Air Recirculation Hurts Data Center Dry Coolers More Than People Assume
Dry cooler capacity scales directly with the Initial Temperature Difference (ITD)—the temperature of the hot fluid entering the cooler minus the temperature of the ambient air entering the fans. At fixed flow rates, heat transfer is proportional to the ITD (Q ≈ ε · C_min · ITD).
A worked design-day example:
Fluid entering the dry cooler: 50°C (122°F)
Design ambient air temperature: 35°C (95°F)
ITD = 15 K
If the hot exhaust plume recirculates and gets sucked back into the fans, it artificially raises the intake air temperature:
+3 K (5.4°F) at intake: ITD drops to 12 K—a ~20% capacity loss.
+5 K (9.0°F) at intake: ITD drops to 10 K—a ~33% capacity loss.
This is the hidden cost of running warm water to eliminate power-hungry chillers. The same small ITD that saves so much energy leaves very little margin for error if hot air recirculates. On a design day, the fans are already running at 100% speed, meaning there is no reserve power left to recover that lost capacity.
These temperature rises are not hypothetical. In one published chiller-yard Computational Fluid Dynamics (CFD) study, downstream units were ingesting air at up to 133°F (23°F above ambient). Installing a horizontal barrier above the chillers at the height of the exhaust ducts successfully blocked downward recirculation into the condenser inlets.
Without such mitigations, at a 12.8 K intake rise, a dry cooler designed for a 15 K ITD is effectively rendered useless.
Four distinct mechanisms drive hot air recirculation. Each must be checked separately during design:
Wind bending the plume over: When wind speed is high compared to the exhaust plume's velocity and buoyancy, the wind bends the plume downward toward the roof instead of letting it rise. It then gets pulled into the low-pressure recirculation zone on the downwind side of the building.
Gaps between units: HTRI's CFD work on air-cooled heat exchanger banks found that lateral gaps between units and differences in fan ground clearance create direct paths for hot air to recirculate—and wind makes it worse.
Insufficient clearance under the intakes: For units that draw part or all of their air from below (like horizontal-coil dry coolers and air-cooled heat exchangers), too little physical clearance under the cooler causes the intake air velocity to become higher than the exhaust velocity, causing the intakes to violently suck in hot air from the surrounding area.
Neighboring heat sources: Generator exhaust, adjacent buildings, parapet walls, and radiator discharge all add external heat or block clean airflow.
The Data Center CFD Modeling Workflow We Run Before Breaking Ground
Define design conditions: Use the ASHRAE extreme ambient metric, not just the standard 0.4% cooling design value. Pull the local wind rose and model at least three scenarios: the prevailing wind direction, the worst-case direction for the site's building wake, and totally calm conditions. (Calm days often produce the highest intake temperatures because the hot plume simply stalls and blankets the roof).
Build the site model: Include every module, the full dry cooler array with exact fan curves, backup generators, fences, parapet walls, and neighboring structures.
Run external CFD: Record the specific intake temperature of every individual unit, not the average across the array. The capacity of the entire plant is ultimately limited by the hottest unit.
Feed the derating back into sizing: Reselect and size the dry coolers based on the worst-case intake temperature from the model, not the nominal ambient weather data.
Run internal cases at the worst external result: Current best practice simulates the room-level data hall cases under the worst external CFD scenario, ensuring the IT gear survives the resulting equipment derating.
Verify after construction: Install intake-air temperature sensors on every physical unit so you can validate the model against real-world data after startup.
The tools to do this are improving quickly. NVIDIA's Omniverse Blueprint for AI factory design, introduced in March 2025, supports robust digital twins for simulating layouts, power topologies, and thermal behavior before a shovel ever hits the dirt. Cadence reports that digital twins found MaxQ operation with warmer coolant could yield roughly 32% more tokens per watt. The engineering principle is exactly the same as before; what has changed is that modern solvers are now fast enough to run dozens of wind and design-day cases in the time it used to take to run one or two.
Design Responses, Ranked by Cost
When the model shows recirculation, here is how you fix it (from cheapest to most expensive):
Orientation and layout: Align the cooler array with the prevailing wind, keep intakes out of the building's wake, and place generators downwind. Cost: $0 (if done early).
Clearance and spacing: Set fan and intake heights directly from the model data, and close off lateral gaps between units.
Elevate Dry coolers: Dry Coolers are the workhorse of the modern AI factory and elevating them to the roof lever allows them to avoid recirculation of other heat sources while providing theme with more favorable conditions away from the ground level.
Discharge extensions and canopies: Raise the exhaust discharge physically above the recirculation zone.
Wind walls and screens: Strategically block prevailing winds from disrupting the exhaust plume.
Oversize the coil: Add raw surface area to the dry cooler to recover the lost ITD. Cost: Very high. This is the expensive penalty you pay when upfront modeling is skipped.
Field Reference
Our 1.4 MW MDC is built entirely off-site, complete with power modules, cooling skids, pump stations, Dry Coolers and chillers. We install our dry coolers on the roof to avoid hot air intake from other heat sources and to take advantage of better wind patterns.
The facility operates at a Power Usage Effectiveness (PUE) of 1.1, which accounts for all energy consumption and systemic losses. Absolutely no potable water is used for cooling. This aggressive performance target was only possible because the electrical and thermal designs were locked in before any equipment was ordered.
Checklist for Your Next Data Centr
Rack heat split (liquid vs. air) is explicitly confirmed via the vendor spec sheet.
Facility water temperature is chosen, and the dry cooler ITD is calculated directly from it.
External CFD is run for calm, prevailing-wind, and worst-wake cases (reporting intake temp per individual unit).
Dry coolers are selected based on the worst-case modeled intake temperature, not ambient weather data.
Generator exhaust and radiator discharge are fully included in the site CFD model.
The design is 100% frozen prior to factory release.
Intake-air temperature sensors are specified for every unit to allow post-construction validation.
Ultimately, speed to market is no longer just a supply chain or construction challenge—it is an engineering problem as well. Prefabrication removes the field delays, but it places the burden of success entirely on the upfront design phase. If you use the checklist above, respect the physics of air recirculation, and rely on rigorous CFD modeling rather than rule-of-thumb assumptions, you can deploy modular capacity at record speed without sacrificing a single kilowatt of performance.
Comments