AI Data Center Retrofit or New Modular Capacity?

An existing data center can be a good home for a new GPU cluster when its electrical, cooling and physical infrastructure can support the selected equipment through both normal operation and the required maintenance or failure conditions. A separate modular facility becomes worth evaluating when adapting those interfaces would consume too much available capacity, disrupt operating services or impose restrictions on the equipment.

Start with the rack list and a site assessment. Compare both options against the same IT installation, operating requirement and date. That reveals which existing assets can be reused and which constraints would follow the project into a new facility.

Four questions determine whether a retrofit is practical

An empty rack position establishes space. Usable GPU capacity also needs power, heat removal, structural support and an operating plan. Assess these together.

The equipment supplier's requirements should drive the assessment. NVIDIA's planning guidance, for example, describes how cooling constraints can limit rack power density and increase the racks needed for a given server population. A room-wide capacity total leaves that local constraint unresolved. NVIDIA deployment planning

Check the full electrical path

Begin at the supply point and follow the path through transformers, switchgear, uninterruptible power supplies (UPS), distribution and rack connections. Identify the ratings, present load, reserved capacity, protection and required service state at each step. The usable increment may be set by one downstream component even when the incoming supply has capacity.

Existing monitoring is valuable evidence when its period, measurement point and operating conditions are recorded. An average consumption figure should be reconciled with the proposed design load and the operating modes the facility must support. Check the assumptions behind concurrent load, maintenance and growth allowances with the electrical designer.

A retrofit may reuse much of the installed system. It may also concentrate the same overall IT load into fewer, denser racks, changing distribution requirements without a corresponding increase in the site's total load. Vertiv discusses reusing or rearranging existing power and rack infrastructure as part of a liquid-cooling deployment. The specific reuse decision requires an assessment of the actual equipment. Vertiv liquid-cooling deployment guide

Apply the same electrical test to a proposed modular addition. Locate its supply point, account for its facility loads and establish whether it uses shared upstream equipment. A separate enclosure can still depend on the same constrained transformer.

Follow the heat beyond the server

List the load captured by liquid, the residual room load and the equipment that rejects heat outdoors. Include networking, storage and power equipment within their correct boundaries. Confirm the required supply temperatures, flows, pressure limits and coolant compatibility from the selected systems.

The cooling arrangement can change the retrofit considerably. A liquid-to-air coolant distribution unit (CDU) transfers heat from the servers’ liquid circuit into room air. The existing room cooling must then remove that heat. A connection to a facility liquid loop changes the heat path but requires compatible conditions and sufficient outdoor rejection capacity. Vertiv's roadmap for existing data centers covers the associated mechanical space, power, rack and service considerations. Vertiv retrofit roadmap

A useful assessment records both total available duty and the ability to deliver it to the new racks. Check the plant at the project's design conditions and specified service states. Identify pipes, pumps, controls and electrical supplies affected by the change.

Review how the remaining equipment will operate during construction. A new pipe route may cross an occupied area. A connection may require an isolation window. Removing an old unit to create space for a CDU may reduce cooling available to the surviving air-cooled equipment. These effects belong in the option comparison.

For the mechanics of the cooling alternatives, use our liquid-cooling guide and CDU sizing guide. The retrofit decision adds the facility's constraints and construction sequence to that equipment choice.

Worked assessment — the same 320 kW cluster in two different sites

Assume a proposed cluster of eight compute racks at a design electrical load of 30 kW each and four support racks at 20 kW each. The resulting IT load is:

8 × 30 kW + 4 × 20 kW = 320 kW.

These are illustrative planning inputs. They represent no named server configuration or customer project. A real design must use equipment data and a checked load schedule. Cooling, power-conversion losses and other facility loads are additional inputs to electrical sizing.

For the compute equipment in this exercise, assume that 75% of its heat is captured in liquid and 25% enters room air. Assume the support equipment releases all its heat to room air. The liquid duty is 240 × 0.75 = 180 kW. The air duty is (240 × 0.25) + 80 = 140 kW. Together, these account for the 320 kW IT design load. Pump, fan and other relevant heat gains must be added at the correct cooling boundary.

Now compare two hypothetical existing halls.

Hall A has a credible reuse case. Its existing assets may save work, but costs and dates still need scoped quotations and an executable sequence. Hall B requires a choice about dependencies. A separate module could remove work from the occupied hall, provided suitable land, heat rejection, supply and connection routes are available.

If both options share Hall B's constrained distribution section, the module still needs that upstream work. The designer should mark shared dependencies before either option receives a delivery date.

Before progressing the module option for Hall B, repeat the assessment at its proposed location. Verify the full electrical supply path, space for outdoor heat rejection, delivery and maintenance access, and the installation sequence for connecting shared services. Price the works that solve those constraints alongside the module. The decision can then compare a deliverable retrofit with a deliverable addition.

Compare costs on the same boundary

Request both options against the same equipment population, IT design load, service requirements and handover state. Keep shared upstream works visible. A retrofit quotation that includes power distribution and a modular quotation that excludes the upstream connection need reconciliation before comparison.

Then evaluate the operating case over a common period. Include maintenance, energy under the stated workload and climate, replacement obligations and the cost of any transitional operating arrangement. An energy advantage needs a defensible model with comparable boundaries. Annual power usage effectiveness (PUE) compares total facility energy with IT energy. Peak electrical and cooling capacity require their own load schedules.

Our enterprise AI infrastructure TCO article provides a wider business-case structure. The facility assessment should precede the financial comparison so that both options price a technically deliverable installation.

Build a sequence around the live service

A retrofit programme should show how crews enter the operating environment, isolate equipment, protect remaining services, complete works and restore the required operating state. Identify the point at which the proposed cluster can be installed and the later point at which it can run at its agreed load.

For a modular addition, link off-site assembly to civil works, connection availability, transport, site installation and testing. The IT deployment can begin only when its necessary interfaces are ready. Parallel work provides value where dependencies and approvals allow it.

Use dates supported by the responsible party. Record provisional dates as assumptions. If one option requires a utility milestone shared by both, changing the construction method alone may leave the energization date unchanged.

When retaining the existing hall makes sense

A retrofit deserves serious consideration when the equipment fits, the necessary electrical and thermal capacity can be demonstrated, and the work sequence is acceptable to operations. It may preserve useful network, security and operating infrastructure while avoiding a separate site package.

Age alone gives a weak answer. An older hall with documented spare capacity and a suitable interface can be more useful than a newer room whose power and cooling are fully committed. Component upgrades can also improve an existing facility without creating enough headroom for the intended GPU installation. Infinitum's discussion of motor-system retrofit illustrates that narrower opportunity. Use a project-specific load and operating model when estimating savings from such component changes. Infinitum

A separate installation deserves comparison when the present hall imposes a material equipment restriction, its required outages are unacceptable or the new operating environment would be easier to establish on a suitable adjacent site. A third option may be to reduce or phase the initial cluster. Test that against the workload requirement before committing to a larger facility project.

Turn the assessment into a supplier brief

Submit the rack list, electrical diagrams, measured loads with dates and boundaries, cooling-system conditions, site or room drawings, delivery survey and permitted outage windows. Ask each supplier to state the retained assets, changed interfaces, exclusions, required investigations and acceptance conditions.

Where evidence is incomplete, request an assessment that resolves those gaps before a firm facility offer. The result should identify a viable route, its assumptions and the work needed to validate it.

ModulEdge can discuss configuration around your equipment and site. An existing-hall proposal and a new prefabricated module should both start with the same equipment brief. That gives the owner a fair basis for deciding which infrastructure investment supports the intended workload.

Frequently asked questions

Can liquid cooling be added to an existing air-cooled data center?

It can be feasible after assessing the equipment interfaces, heat-rejection path, electrical capacity, space and construction sequence. A liquid-to-air CDU still leaves the room cooling responsible for removing the released heat.

How much does a data center retrofit cost?

A defensible estimate needs the equipment brief and the retained, modified and replaced infrastructure. Request comparable scope schedules for the retrofit and any new-build alternative, including upstream work, temporary systems and testing.

Does a modular addition require more utility power?

Determine its facility input from the proposed load schedule and supply architecture. The project may reuse existing capacity or require more capacity, depending on the current loads, available headroom and required service conditions.

Can PUE be used to size the electrical connection?

Annual PUE describes an energy ratio. Connection sizing needs an electrical design that accounts for peak demand, losses, auxiliary loads and the specified operating states.