The fastest way to waste an expensive GPU delivery is to let the servers arrive before the site can power and cool them.
A 40HC container can be transported quickly. That does not make it an AI data center. The real product is the operating system around the servers: rack layout, electrical distribution, cooling, heat rejection, fire protection, monitoring, service access, and the site approval pathway.
Miss one of those interfaces and the project may own a full container of compute that cannot reach full load.
This guide explains how a **modular AI data center** should be evaluated, when a 40HC module makes commercial sense, and which technical information buyers must confirm before requesting a reliable quotation.
A Modular AI Data Center Is More Than a Container Shell
A modular AI data center is a factory-integrated infrastructure unit built around a defined compute load. It can combine:
– GPU server racks and structural support.
– Main power input, protection, metering, PDU, and grounding.
– Direct-to-chip liquid cooling or precision air cooling.
– CDU, manifold, pump, chiller, dry cooler, or condenser interfaces.
– Fire detection and project-engineered suppression.
– Leak, temperature, smoke, electrical, and cooling-system monitoring.
– Front, rear, and side maintenance access.
– Weather protection, drainage, freeze protection, and outdoor interfaces.
The factory integration matters because these systems affect one another. A server change can alter rack space, floor loading, branch-circuit current, heat load, coolant flow, pressure drop, pipe size, and heat-rejection capacity at the same time.
Buy the server first and “fit it into the container later” is not a deployment strategy.
The correct sequence starts with the server list and site conditions. The enclosure follows the load.
Why 40HC Is a Practical Format for AI Infrastructure
The 40HC format offers a useful balance between usable internal volume and standard global logistics. It can be factory assembled, tested, transported, lifted, and connected as a repeatable module.
That creates four commercial advantages.
1. Factory Work Replaces Site Work
Racks, piping, electrical panels, controls, sensors, and safety systems can be installed under controlled manufacturing conditions. The buyer reduces the amount of field fabrication performed by multiple contractors at the same time.
2. Deployment Can Be Phased
A project can begin with one module and add capacity as utility power, network connectivity, and GPU demand grow. This helps prevent a large conventional building from sitting underutilized during the early phase.
3. The Design Can Be Repeated
Once a module has been validated against a server family and site architecture, the engineering basis can be reused for later phases. Repetition does not remove site review, but it reduces unnecessary redesign.
4. The Module Can Be Tested Before Shipment
Factory acceptance testing can verify the build against approved drawings before the equipment reaches the site. Electrical checks, cooling-loop tests, alarms, controls, leak detection, and documentation can be reviewed while the manufacturer still has full access to the module.
Go modular when schedule certainty and phased expansion are worth more than maximizing every square foot of a permanent building.
Pro Tip:
Confirm the shipping route, crane access, foundation, utility entry points, and outdoor heat-rejection location before freezing the internal layout. A module that fits the servers but not the site is still the wrong design.
Do Not Calculate Capacity by Server Count Alone
“How many GPU servers fit in one 40HC container?” sounds like a simple question. It is not.
Practical module capacity is the lowest limit created by:
– Rack U-space and server depth.
– Server weight and structural loading.
– Continuous electrical capacity.
– Branch-circuit and connector limits.
– Air-side and liquid-side heat load.
– Coolant flow and pressure drop.
– Cable and piping routes.
– Replacement clearances and emergency access.
– Heat-rejection capacity at the site’s design ambient temperature.
The useful engineering rule is:
Usable AI capacity = the lowest limit set by space, power, cooling, hydraulics, structure, and serviceability.
An ACT liquid-cooled reference design uses five rack and manifold groups for 35 high-density AI servers. The documented reference operating point includes approximately 31.5 m3/h total flow and 15 LPM per server. Depending on the selected server version, the reference IT load is roughly 213.5 to 225 kW, with an approximately 300 kW total input design basis for that project configuration.
Those numbers are not a universal capacity promise. Change the server, and the operating point changes.
The air-cooled reference architecture uses five 46U racks, providing 230U of total rack space. A documented configuration uses a 264 kW nominal cooling reference and a 238 kW modeled IT load. Usable rack capacity still depends on server depth, airflow, cabling, power density, and maintenance clearance.
Capacity must be calculated cabinet by cabinet, not guessed from the length of the container.
Liquid Cooling or Precision Air Cooling?
The cooling decision should follow the GPU heat load, not a marketing preference.
| Decision Factor | Liquid-Cooled AI Module | Precision Air-Cooled AIDC |
| Best fit | High-density direct-to-chip GPU servers | Lower or mixed rack densities with air-cooled servers |
| Primary design inputs | Liquid/air heat split, flow, pressure drop, supply temperature | Server airflow, inlet temperature, containment, static pressure |
| Internal architecture | Cold plates, manifolds, CDU-based secondary loop | Cold aisle, hot aisle, precision indoor cooling units |
| Outdoor heat rejection | Dry cooler or chiller selected around the operating point | Outdoor condenser or project cooling plant |
| Main engineering risk | Poor flow balance, contamination, trapped air, leak response | Recirculation, bypass airflow, filter resistance, hot spots |
| Maintenance focus | Quick disconnects, valves, pumps, water quality, leak detection | Filters, fans, coils, aisle sealing, condensate and airflow |
Choose Liquid Cooling Around the Heat Split
A liquid-cooled server may still reject part of its heat to the room air. The engineering team must confirm how much heat enters the cold-plate loop and how much remains on the air side.
That split determines both the CDU duty and the residual air-conditioning requirement. A CDU sized only from the server nameplate can be wrong if the liquid capture ratio is not defined.
The liquid path normally follows this sequence:
1. GPU cold plates absorb server heat.
2. Server manifolds distribute and collect coolant.
3. The CDU controls flow, pressure, and heat exchange.
4. The primary loop transports heat outside the module.
5. A dry cooler or chiller rejects heat to the environment.
Every component must be selected at the same design operating point.
Choose Air Cooling Around the Airflow Path
Air cooling remains practical when rack heat density and server design allow stable inlet temperatures without excessive fan and compressor energy.
The airflow path must prevent hot exhaust from returning to server intakes. Rack layout, cold-and-hot-aisle containment, precision cooling capacity, filter pressure drop, outdoor condenser performance, and partial-load control all affect the result.
Do not compare cooling systems using nominal capacity alone. Compare them at the target ambient temperature, required supply condition, part load, redundancy state, and expected maintenance condition.
Pro Tip:
Ask the server supplier for heat split, required coolant temperature, flow range, allowable pressure drop, water-quality limits, quick-disconnect type, and material compatibility. “Liquid-cooled server” is not enough information to design the loop.
The Outdoor Heat-Rejection System Is Part of the AI Data Center
The container does not make heat disappear. It only moves heat to another boundary.
For liquid cooling, the CDU transfers heat from the server loop to a facility-side loop. The dry cooler or chiller must then reject that heat under the site’s real climate conditions.
Dry coolers can reduce water dependence and simplify closed-loop heat rejection when the required supply temperature and local dry-bulb conditions allow it. Chillers provide lower coolant temperatures and broader control, but they add compressor energy, controls, maintenance, and another failure layer.
The selection should consider:
– Peak IT heat load and liquid capture ratio.
– Maximum and minimum outdoor temperature.
– Required supply and return temperatures.
– Glycol concentration and freeze protection.
– Altitude and air-density correction.
– Coil fouling and site dust.
– Fan redundancy and noise limits.
– Pump head and total system pressure drop.
– N, N+1, or 2N availability target.
Buying the CDU and outdoor cooler from unrelated operating points is a common integration failure. The heat exchanger, pump curve, pipe network, coolant mixture, and outdoor equipment must be checked together.
Power Architecture Must Follow the Final GPU Load
AI servers create dense, continuous electrical loads. The module power design must be based on the final server power supply, input voltage, connector arrangement, redundancy strategy, and expected operating profile.
North American projects may use a 480V three-phase module input, with downstream distribution or transformation selected around the server load. The final architecture should verify:
– Utility and transformer boundary.
– Main breaker and short-circuit rating.
– Continuous-load sizing and required derating.
– Phase balance and harmonic performance.
– Branch protection and PDU outlet mapping.
– Metering and power-quality monitoring.
– Surge protection and grounding.
– Auxiliary loads for pumps, fans, controls, fire systems, and lighting.
– Spare capacity for startup, redundancy, and future expansion.
The cooling system belongs in the power balance. A module with 225 kW of IT equipment does not have a 225 kW site demand after pumps, fans, controls, and heat-rejection equipment are included.
For ROI calculations, use usable IT capacity and total facility demand, not the server nameplate alone.
North American Approval Cannot Be Added at the End
Electrical components with UL Listed, UL Recognized, cULus, or CSA status can support a North American project. They do not automatically certify the complete containerized data center.
The final pathway depends on the approved bill of materials, system scope, installation method, utility requirements, local code adoption, field inspection, and the Authority Having Jurisdiction.
Fire protection requires the same discipline. A project-engineered clean-agent system may reference NFPA 72 and NFPA 2001 and can include smoke detection, heat detection, staged alarms, release delay, manual release, emergency stop, and system interlocks. The protected volume, enclosure integrity, agent quantity, local approvals, and commissioning procedure remain project-specific.
Avoid vague claims such as “NFPA certified container” or “universally UL certified.” Ask what is certified, which standard applies, who owns field coordination, and what evidence will be delivered.
Build the approval pathway into the design before the electrical panel and fire system are ordered.
What Factory Integration Should Include
A credible modular AI data center supplier should coordinate the interfaces before shipment, not simply mount third-party equipment inside a steel shell.
The factory scope should include:
– Approved general arrangement and rack elevation.
– Electrical single-line, wiring, protection, and grounding drawings.
– Cooling schematic, flow basis, pressure-drop review, and equipment operating point.
– Control narrative, I/O list, alarm matrix, and emergency logic.
– Fire detection and suppression design within the contracted scope.
– Cable, pipe, sensor, drainage, and service-access routing.
– Factory acceptance test procedure and recorded results.
– Packing, lifting, shipping, installation, and commissioning documents.
FAT should verify configuration, electrical safety, controls, alarms, leak detection, pump operation, flow distribution, temperature sensors, fan operation, and documented corrective-action closure. Site acceptance testing must repeat the relevant checks after transport, placement, grounding, utility connection, field piping, and network integration.
No evidence, no shipping release.
Pro Tip:
During FAT, measure the hydraulically farthest liquid branch or the thermally most difficult air-cooled rack position. A strong total-flow or average-temperature number can hide the exact location that will limit full-load operation.
A Better ROI Framework for AI Compute Containers
The lowest purchase price does not identify the lowest-cost AI deployment.
Evaluate the project using five commercial measures.
1. Time to Usable Compute
Measure the time from purchase order to commissioned GPU capacity, not the date the empty container arrives. Include engineering approval, factory build, FAT, transport, foundations, utilities, field connection, and SAT.
2. Cost per Usable IT Kilowatt
Use this framework:
Cost per usable IT kW = module + cooling plant + electrical infrastructure + site works + logistics + commissioning, divided by validated usable IT capacity.
Do not divide by theoretical rack capacity.
3. Facility Energy Overhead
PUE changes with climate, load, cooling architecture, supply temperature, control strategy, and equipment performance. Treat PUE as a modeled project result, not a permanent product label.
Annual non-IT facility energy can be estimated as:
IT energy x (PUE – 1).
Run this model across several ambient temperatures and part-load conditions.
4. Stranded Capacity Risk
Oversized infrastructure locks cash into unused electrical and cooling capacity. Undersized infrastructure prevents expensive GPUs from operating at full power. Phased modular expansion can reduce both risks when the site utilities and future module layout are planned early.
5. Maintenance and Downtime Exposure
Count filters, pumps, fans, compressors, valves, controls, and water-treatment tasks. Review how failed equipment can be isolated and replaced. Redundancy has value only when the control sequence, service access, and spare-parts plan allow it to work.
The best module is not the one with the most equipment. It is the one that turns the highest percentage of installed infrastructure into stable, maintainable compute.
What Buyers Should Send Before Requesting a Quote
A reliable quotation starts with engineering inputs. Send the supplier:
1. Server manufacturer, model, quantity, dimensions, weight, and U-space.
2. Typical, rated, and peak power per server.
3. Air-side and liquid-side heat split.
4. Coolant type, required supply temperature, flow, pressure drop, and connection type.
5. Project location, altitude, and hourly design climate data.
6. Available voltage, frequency, transformer, and short-circuit information.
7. Required cooling architecture and outdoor heat-rejection preference.
8. Availability target: N, N+1, 2N, or phased redundancy.
9. Fire-protection, monitoring, DCIM/BMS, and local approval requirements.
10. Indoor or outdoor installation, delivery route, foundation, crane, drainage, and schedule.
If these inputs are missing, the supplier can provide only a budget estimate. A fixed server count and fixed PUE promise made before this review should be treated carefully.
Final Verdict: Buy an Operating Point, Not an Empty Box
A 40HC AI compute container can shorten deployment, reduce field coordination, and support phased GPU expansion. Those benefits appear only when the module is engineered as a complete system.
Start with the server load. Define the cooling operating point. Match the power architecture. Coordinate fire protection and field approval. Verify the assembled module before shipment.
Then scale.
For high-density AI projects, the winning question is not “How many servers can you put in the container?”
Ask this instead:
How much validated, maintainable AI compute can this module operate at my site conditions?
That is the capacity that produces revenue.
Frequently Asked Questions
How many GPU servers fit in a 40HC AI data center container?
The quantity depends on server dimensions, U-space, weight, power, heat split, cooling flow, pressure drop, cable routing, and maintenance clearance. A reference design may use 35 high-density AI servers, but the final number must be recalculated for the selected server.
Is liquid cooling always better for an AI data center?
No. Liquid cooling is usually favored as rack heat density rises or when the server platform requires direct-to-chip cooling. Precision air cooling can remain practical for suitable server densities and climates. The decision should compare total system energy, heat-rejection requirements, maintenance, and availability.
Does the CDU provide all cooling for the container?
The CDU controls the server-side liquid loop and transfers heat to the facility-side loop. A dry cooler, chiller, or other heat-rejection system is still required. Residual air-side heat may also require room-level cooling.
Is a container UL certified because it uses UL-listed components?
No. Component certification does not automatically certify the complete container. The final compliance and inspection pathway depends on the assembled system, project scope, local code, and AHJ requirements.
What information is required to size a modular AI data center?
At minimum, provide the server list, power profile, heat split, liquid-cooling requirements, site climate, voltage, redundancy target, fire and monitoring requirements, deployment location, and schedule.
