Power, cooling and layout in AI data center infrastructure all follow from one decision: the server platform. What to settle first, from rack power and liquid-cooling heat rejection to module layout and factory prefabrication.
AIDC Infrastructure Notes · Part 1
Most AI data center projects open with the servers: which platform, how many nodes, how much compute. Engineering then asks a different set of questions. How much power can the site actually deliver? Where does the heat go? How do the racks get in, and how does anyone service them once they are in?
Those answers set both how much compute a site can hold and when it can start running. In AI data center (AIDC) infrastructure, power, cooling and physical layout are not three separate work packages; they are three views of the same server configuration. Prefabricated compute modules are designed from exactly that starting point.
Two racks can look identical and draw very different power, because what sits inside them is not the same. Once the server count is fixed, the load list still has to pick up networking, cooling plant and the rest of the auxiliary load before anyone can see what the facility really demands.
That load list drives everything downstream: transformers and switchgear, the number of feeders, cable sizing, busway. Redundancy belongs in the same conversation, along with the blunt question of what has to keep running while something else is being serviced. Add a feeder or reroute cable after the racks and distribution boards are already placed, and the change rarely stays local. Splitting medium voltage, low voltage, transformers and batteries into separate functional modules is a common way to keep that manageable, and it is how the ETENZ AI Data Center Power E-House is divided.
The industry is also testing new ways to distribute that power. In August 2026 NVIDIA presented an 800 VDC architecture for AI factories, together with a path for moving over from existing AC facilities in stages. Shifts like this mean infrastructure has to be designed against what today's servers need and against where the next expansion is heading.
For any real project the sequence stays the same: fix the server platform, the deployment size and what the site can supply, then judge which architecture fits and how much space to reserve for it.
Liquid cooling shifts part of the heat-removal job into piping. In a typical liquid-to-liquid cold-plate scheme, coolant picks up heat at the server cold plates, exchanges it with the facility loop through a coolant distribution unit (CDU), and outdoor equipment then rejects it to ambient.
Follow that path and the matching conditions stack up quickly. The inlet coolant temperature the servers accept narrows the choice of heat-rejection source. Pipe runs, valves and heat-exchanger pressure drop set what the pumps have to do. Ambient temperature caps what the outdoor side can actually reject. These have to be checked against each other; lock one of them on its own and another usually overturns it later.
Air cooling deserves the same line-by-line check. Many configurations put cold plates on the main processors while everything else is still cooled by air. Ventilation, air conditioning and the liquid loop inside a compute module should each carry a share of the load that matches the equipment actually installed.
This is why cooling equipment has to reach the general arrangement early. Where the CDU sits, where the pipes run, whether the outdoor units have enough air intake, discharge and service clearance: all of it drives module dimensions and site footprint. The outdoor leg is usually handled by a Closed-Loop Dry Cooler or a heat-exchange unit, and its rating and footprint should be fixed on the same schedule as the module itself.
How many racks a compute module holds is never answered by floor area alone. Rack dimensions, fully loaded weight, the power and cooling method, and the working space in front of and behind the equipment all set that number together.
Which side a server slides out of, how maintenance tools get in, whether pipe couplings can be reached, whether a bent cable ends up blocking the aisle: these belong in the layout review, not in a punch list. Take the dimensions from the equipment vendor's installation documents rather than estimating from the rack outline. Withdrawal travel, connector orientation and minimum cable bend radius usually decide aisle width more than the rack itself does.
A prefabricated module also has to survive the trip from factory to site. Whether equipment ships inside the module or is installed on arrival changes base-frame loading, equipment fixing, the entry route into the module and the order of installation. How many transport sections the module splits into, and how piping and cabling are reconnected on site, get decided alongside the equipment layout rather than after it.
So it pays to put servers, racks, distribution boards, CDU, cable tray and piping onto one drawing while the concept is still open. A clash found there costs a revision. The same clash found after fabrication costs a great deal more.
Prefabricated delivery moves enclosure fabrication, equipment supports, cable tray, piping and part of the system testing into the factory. While that runs, the site gets on with foundations, utilities and connection points. The two tracks run in parallel, and the schedule is set by whichever one finishes later.
Whether those tracks meet cleanly depends on how well interfaces and installation sequence were thought through at the start. The factory needs to know where equipment enters, where each service terminates and which checks can be closed out before shipping. The site needs to prepare foundations, power and cooling connections from the same drawing set.
Phased builds can be organised the same way, with modules grouped by compute, power and cooling demand. Which equipment goes in first, which direction the build grows next, and how the spare interfaces get used should all be visible on the phase-one general arrangement.
ETENZ handles structural design and manufacturing around prefabricated modules, and integrates the electrical, environmental control, thermal management, fire protection and security subsystems a project calls for. On AIDC projects, cooperation can start at enclosure and interface prefabrication, extend to subsystem integration, or run through to a finished module delivered ready to install. OEM and ODM programmes follow the same layers.
Once a shortlist of server models and a deployment plan exist, both sides can start checking rack arrangement, equipment weights, the power method and the cooling conditions, then fold the site and transport limits back into the design. From there, enclosure design, auxiliary equipment selection and manufacturing preparation all run off the same basis.
Part 2 follows the heat from the server to the outdoor air, and looks at how cold plates, CDUs, piping and the heat-rejection source have to fit together.
The AI Data Center Power E-House splits a compute site's power train into four function cabins delivered as one set: medium-voltage distribution, low-voltage distribution, solid-state transformer and battery storage. Each of the three function cabins measures 3495 mm wide by 13716 mm long by 4150 mm high; the battery cabin is built to the 2438 mm standard container width. ETENZ builds all four at the works with the panel-lineup foundations, busway and tray routes, auxiliary distribution and fire-and-security interfaces already in place, so the cabins are set down side by side on site and only the cabin-to-cabin connections remain. The extra width leaves a service aisle both in front of and behind the panel lineup; heavy outgoing circuits run on overhead busway while power and control cables drop through a brush-sealed floor entry, so the two routes never cross. Built to IP54/55 and a C3-C5 corrosion system with a 25-year design life, available from bare-shell supply through OEM/ODM.
A closed-loop dry cooler rejects the heat of the circulating fluid straight to ambient air. The loop stays sealed, so there is no make-up water, no water treatment and no drift. Once ambient falls below the changeover point, mechanical cooling can be switched off and the dry cooler supplies directly — the main source of annual energy saving in data centres, liquid-cooled compute containers and process cooling. ETENZ selects the unit against the project duty and integrates the enclosure-side interface.
Thank you for sharing ETENZ content
A custom prefabricated enclosure needs more than a corrosion category. ETENZ turns exposure, substrates, coating build-up, inspection and repair requirements into a package that can be quoted, built and verified.
ETENZ is advancing BESS enclosures for the Zhongwei grid-forming energy storage project, with 143 twenty-foot-class units and their structural, fire-service and electrical interfaces prepared in the factory.
ETENZ is supplying the BESS enclosure scope for the China Three Gorges high-voltage direct-connected storage project in Haixi, Qinghai, contracted through NR Electric: 72 battery enclosures with their internal structural steelwork. The first unit has passed first-article review and the first batch shipped in late July 2026.
ETENZ is advancing BESS enclosures for the Zhongwei grid-forming energy storage project, with 143 twenty-foot-class units and their structural, fire-service and electrical interfaces prepared in the factory.
ETENZ News Center
A custom prefabricated enclosure needs more than a corrosion category. ETENZ turns exposure, substrates, coating build-up, inspection and repair requirements into a package that can be quoted, built and verified.
ETENZ is advancing BESS enclosures for the Zhongwei grid-forming energy storage project, with 143 twenty-foot-class units and their structural, fire-service and electrical interfaces prepared in the factory.
ETENZ is supplying the BESS enclosure scope for the China Three Gorges high-voltage direct-connected storage project in Haixi, Qinghai, contracted through NR Electric: 72 battery enclosures with their internal structural steelwork. The first unit has passed first-article review and the first batch shipped in late July 2026.
