11 September 2026 · MDCX

Factory-tested vs field-built modular AI data centers

AI densification is changing what “capacity” means on a critical path. When racks move into SXM- and OAM-class density, the old habit of designing a custom hall and discovering heat rejection, power distribution, and controls on the pad stops being a schedule strategy. It becomes the schedule risk.

The useful contrast for operators is not whether a vendor says “modular.” It is where systems are proven: inside a factory envelope before ship, or for the first time on your site. That is the difference between ordering a product and running another construction wave.

MDCX builds that product as standard Liquid Cooling and Immersion Cooling blocks — an envelope of space, cooling, and power. You bring the silicon. First and expansion blocks are quoted at 90 days ex-works from down payment. Freight and customs stay buyer-arranged. Delivery is four segments quoted alone, never one fake door-to-door clock.

A factory-tested envelope is commissioned as plant before it ships — heat rejection, power, and controls on the same unit that lands on the pad.

The critical-path question is where systems are proven

Facilities teams and EPC leads already know air-cooled rooms will not hold next-wave density. The mid-commercial question is sharper: do you keep integration risk on the pad, or retire it in a factory before the block leaves the line?

A factory-tested envelope is prefabricated so heat rejection, power, and controls are commissioned as plant behavior before ship. The commissioning gate is real equipment running as the unit that will sit on your pad — not a paper design that still has to be assembled in the weather.

A field-built (or stick-built) room assembles those same systems on site. Mechanical finish comes first; BMS or DCIM often bolts on later; heat-rejection surprises show up late. Expansion is another construction wave, with another set of trade interfaces and another discovery cycle.

“Modular” as marketing language is not enough. The test that matters is provenance: where heat rejection is sized, where power topology is proven, and where controls ship as part of the block. That is the operator vocabulary procurement can put into an RFP without inventing a fake end-to-end promise.

Industry context often cites conventional stick-built halls in the 18–36 month range. That figure is industry context for schedule pressure — not an MDCX delivery claim, and not a competitor dunk. MDCX’s honest clock starts at 90 days EXW plus four separately quoted segments.

A factory-tested envelope is commissioned as plant before it ships — heat rejection, power, and controls on the same unit that lands on the pad.

Field-built: systems assembled on the pad. Illustrative construction context — not a named site.

What ships in the box (envelope, not GPUs)

What arrives is an envelope: space, cooling, and power distribution sized for AI density. MDCX does not sell GPUs. The customer brings the servers. Blocks are GPU-agnostic so silicon roadmaps can move without forcing a new facility architecture every generation.

Liquid Cooling is the empty-pad lead story for SXM / OAM-class density. Immersion Cooling is the secondary fork when the hall is PCIe-heavy — useful, but not the default empty-pad narrative. See the Liquid Cooling hub and the Immersion Cooling hub for line-level detail.

Two Liquid Cooling Aliases cover the common pad conditions in the same 45 ft High Cube footprint:

Pick the Alias from site power reality, not from marketing preference. Empty pad → L1240C45. Power already on the pad → L1800C45. Methodology and product framing live on Why MDCX.

L1800C45 envelope with the side open — space, cooling, and power in one factory-tested block.

Factory-tested vs field-built: operator comparison

Integration. Integration is completed before freight, so pad work is installation of a tested unit rather than first-time assembly of plant.

Heat rejection. Liquid density fails quietly in a spreadsheet and loudly on a pad. When heat rejection is sized with the block and proven as plant behavior in the factory, late-room discoveries stop owning the critical path.

Controls. Every MDCX block includes CIOS on the site plane today. Field-built rooms often finish mechanical work first and add BMS or DCIM afterward — a second integration cycle after the room already looks “done.”

Expansion. The next megawatt of envelope is another product order on the same 90 days EXW factory clock — not another multi-trade construction wave. That is how densification becomes repeatable capacity rather than a one-off project.

Expansion is another product order on the same factory clock — not another construction wave.

Prefab AI DC delivery without a fake end-to-end clock

SERP copy often quotes a single “dirt to compute-ready” number. That habit hides who owns freight, customs, and grid interconnection. MDCX stays honest at ex-works.

• 90 days EXW from down payment for first and expansion blocks. When the production line is full, the first block may take up to 120 days; expansion blocks remain on the 90-day EXW quote.
• Freight and customs are buyer-arranged. They are not inside the factory clock.
• Four delivery segments are quoted alone — never summed: survey 3–4 weeks, manufacture 12–15 weeks, shipping ~7 weeks typical, install 3–4 weeks.
• Grid interconnection remains a separate critical path outside MDCX delivery.

This framing is deliberate. Prefab AI DC delivery only helps procurement when the clock matches commercial reality: factory gate first, logistics second, site install third, utility interconnection on its own rail. See Why for the methodology behind Speed as a product attribute rather than a slogan.

CIOS: the site plane that ships with the block

Hardware without an OS is just metal. CIOS (Compute Infrastructure OS) is open-core Apache-2.0 and included with every MDCX block.

Shipped today (site plane): path-addressed telemetry; Modbus TCP and SNMP v2c; alarms that become tickets on the same path; usage records for operational truth (CIOS does not invoice); and at least seven days of offline autonomy on one site box.

Roadmap only — not shipped: multi-site fleet plane; GPU-level telemetry; AI Ops; Omniverse-as-shipped; DCM; 800V HVDC. Those items stay clearly fenced so operators do not confuse a site plane that ships with a future multi-site story.

CIOS is the controls half of the factory-tested claim: the envelope arrives with a site plane already in the box, not as a bolt-on after mechanical finish. Details and source: mdcx.org/cios.

CapEx starting point and next step

Published starting CapEx is from $3M/MW — indicative, not a quote. Model the rest in the TCO tool. Do not treat that figure as a fixed bid, and do not mix IT load with total facility load when you compare scenarios.

For empty pads, start with L1240C45. When the pad already carries power distribution, compare L1800C45. For site operations scope, read CIOS. For schedule and product logic, read Why.

Secondary: TCO · L1240C45 · L1800C45 · CIOS

FAQ

What is the difference between a factory-tested modular AI data center envelope and a field-built room?

A factory-tested envelope is prefabricated so heat rejection, power distribution, and controls are proven as plant behavior before the block ships. A field-built room assembles those systems on the pad, so integration risk and late commissioning discoveries stay on the critical path. For AI density, the useful question is not whether something is labeled “modular,” but where the commissioning gate happens.

How fast can an MDCX modular Liquid Cooling block ship — and what does 90 days EXW include?

First and expansion MDCX blocks are quoted at 90 days ex-works from down payment. That factory clock does not include freight or customs, which are buyer-arranged. When the production line is full, the first block may take up to 120 days; expansion remains 90 days EXW. Delivery is quoted as four segments alone — survey, manufacture, shipping, and install — and those segments are never summed into one fake end-to-end timeline. Grid interconnection remains a separate critical path.

What ships in an MDCX block, and which Alias should an empty pad start with?

MDCX ships an envelope: space, cooling, and power. The customer brings the silicon. Liquid Cooling leads for SXM / OAM-class density; Immersion Cooling is the secondary path for PCIe-heavy halls. Empty pads typically start with L1240C45 (UPS inside the box). When the pad already has power distribution, L1800C45 puts more IT in the same 45 ft footprint. Every block includes CIOS on the site plane today; multi-site fleet plane, GPU-level telemetry, and AI Ops remain Roadmap. Starting CapEx is published from $3M/MW and should be modeled in the TCO tool — it is indicative, not a quote.