IES Data Center Retrofits Challenge the AI Greenfield Building Race

IES data center retrofits have gained a timely new argument: existing facilities can add AI capacity without waiting for an entirely new campus. A September 8 whitepaper promoted by Data Center Dynamics says targeted upgrades can unlock power, cooling, and floor space for denser computing. The conflict is speed. Operators need capacity now, while purpose-built AI facilities can take years to plan, connect, construct, and commission.

The paper comes from building-performance software company IES, so its conclusions should be read as a vendor-backed framework rather than independent research. Still, the underlying problem is widely documented. AI servers concentrate more electricity and heat inside each rack, while grid connections and critical equipment remain difficult to secure.

That creates a contest between retrofitting existing sites and developing greenfield facilities from the ground up. Greenfield construction offers fewer design compromises. Retrofitting starts with constraints, but it also starts with a building, network access, operating staff, and sometimes an established grid connection.

The result is not a universal case for converting every older data center. It is a narrower and more consequential proposition. The fastest usable AI capacity might come from carefully selected parts of the installed data center estate, not only from new megacampuses.

IES Puts Existing Capacity Back Into the Race

The immediate change is that IES has presented retrofitting as a capacity strategy, not simply a maintenance project.

The retrofit framework asks colocation operators to assess existing facilities for AI readiness. Its scope includes electrical distribution, cooling architecture, structural loading, connectivity, and resilience. These are interconnected constraints, not separate items on a renovation checklist.

The framework also promotes phased upgrades and dynamic simulation. Dynamic simulation uses a time-based digital model to test how a building responds as workloads, temperatures, and equipment states change. IES argues that operators can compare interventions under realistic conditions before committing capital.

That distinction matters because a nameplate capacity figure rarely tells the whole story. A facility can have unused floor space but insufficient power. It can have available electrical capacity but lack a practical way to remove heat from dense racks. It can satisfy both conditions yet face structural, hydraulic, network, or redundancy limits.

A retrofit assessment therefore begins by locating stranded capacity. This is infrastructure that exists but cannot support the desired workload without a targeted change. The intervention might involve switchgear, busways, cooling distribution units, rear-door heat exchangers, containment, pumps, controls, or reinforcement beneath selected equipment.

The strongest version of the IES argument concerns sequencing. An operator does not need to convert an entire building into an AI factory. It can identify a suitable zone, model the consequences, and establish a limited high-density deployment alongside conventional workloads.

That approach preserves the value of equipment that still performs adequately. It also avoids treating every existing air-cooled hall as obsolete. Conventional servers, storage systems, and lower-density inference equipment can remain in place while a smaller liquid-cooled environment serves hotter hardware.

This is why the story extends beyond ordinary facilities management. The retrofit becomes a deployment decision. Operators that find viable capacity inside existing sites can offer AI infrastructure while competitors remain tied to construction and utility schedules.

IES has not published evidence showing that every modeled intervention will succeed. Nor does the DCD page provide a universal deployment timetable or return profile. Its useful contribution is the decision structure: measure the whole facility, test interacting systems, and upgrade only where the combined constraints permit it.

That reframes the infrastructure race. The scarce resource is not simply land or server inventory. It is deployable megawatts that can support the complete power, cooling, network, and resilience chain.

Why AI Infrastructure Cannot Wait for Greenfield Campuses

AI demand is rising faster than the physical systems that supply and condition electricity for new data centers.

The International Energy Agency reported that data center electricity demand rose 17 percent during 2025. It also found that capital spending by five large technology companies exceeded $400 billion that year. The agency expected that spending to increase by another 75 percent in 2026.

Those investments do not eliminate physical bottlenecks. The same energy outlook identified tighter supplies of transformers, gas turbines, advanced chips, and other infrastructure components. Planning processes, permits, and grid connections are also delaying projects.

The IEA expects global data center electricity consumption to double by 2030. It expects electricity use from AI-focused facilities to triple over the same period. Demand from individual AI tasks is becoming more efficient, but growing adoption and more intensive workloads outweigh those savings.

This combination puts colocation providers, enterprise operators, and infrastructure investors under immediate pressure. Customers want accelerator capacity sooner than many utilities can deliver new connections. Equipment buyers also compete for switchgear, transformers, cooling hardware, and skilled commissioning teams.

Greenfield development remains essential. A purpose-built site can align electrical architecture, structural design, water systems, networking, and maintenance access around high-density hardware. It can also reserve space for future cooling and power technologies.

Its disadvantage is time exposure. A new building depends on land approval, environmental review, construction, equipment availability, utility coordination, and testing. A delay anywhere in that chain can postpone revenue and leave expensive computing hardware without a suitable home.

Existing sites begin at a different point. They already have some combination of energized capacity, fiber routes, security controls, operations personnel, and customer access. Those assets do not guarantee AI readiness, but they narrow the gap between a capacity decision and a working deployment.

Location can make this advantage more important. AI training clusters need large power blocks, while many inference applications also benefit from proximity to users, enterprise data, or established network exchanges. An existing metropolitan facility might support such workloads sooner than a distant campus awaiting construction.

This does not mean retrofits create electricity. They cannot turn a constrained utility connection into an unlimited supply. However, an operator can sometimes recover capacity through equipment replacement, airflow correction, thermal isolation, or improved controls. It can then direct the released capacity toward a limited AI deployment.

The pressure is therefore asymmetric. Operators with measurable headroom can act while competitors pursue new supply. Operators without that headroom risk spending time on studies that ultimately confirm the need for a new site.

IES data center retrofits enter the conversation at precisely this point. Their value depends less on architectural ambition than on whether they convert existing resources into operational AI capacity before new construction arrives.

IES Data Center Retrofits Depend on System-Level Modeling

The retrofit case works only when operators model the building as one system and test failure conditions before installation.

AI racks affect more than room temperature. They can alter electrical loading, water flow, pump demand, heat rejection, maintenance procedures, and backup-power requirements. Increasing capacity at one layer can expose a bottleneck somewhere else.

A cooling distribution unit, or CDU, transfers heat between the server-side cooling loop and a facility-side system. Liquid-to-liquid CDUs connect to available facility water. Liquid-to-air units reject heat into the room when a suitable water system is unavailable.

Schneider Electric illustrates both options in a reference design for a 3,818-kilowatt Tier III facility. The design places high-density AI clusters beside conventional equipment. It includes air-cooled, liquid-to-air, and liquid-to-liquid retrofit scenarios.

That example supports a phased path, but it also shows why the choice cannot be reduced to buying a cooling product. A liquid-to-air CDU transfers the thermal burden back into the data hall. The room-level cooling system must still absorb that heat without compromising neighboring equipment.

A liquid-to-liquid design can move heat more directly into the facility water loop. However, operators must verify water temperatures, flow rates, pipe dimensions, redundancy, pumping capacity, water quality, and heat-rejection performance. A suitable connection on a diagram does not guarantee sufficient performance during peak conditions.

Electrical upgrades create similar interactions. High-density racks can require changes to upstream distribution, rack power delivery, protection settings, backup systems, and monitoring. Concentrating load also changes the consequences of a localized failure.

Structural loading is another practical boundary. Accelerator racks can be heavier than the equipment an older floor was designed to hold. Pumps, manifolds, heat exchangers, and larger power components may consume adjacent space. Cable routes and maintenance clearances further reduce the area available for computing.

Dynamic simulation can help operators test these interactions over time. A static engineering calculation captures a design condition. A dynamic model can examine changing outdoor temperatures, workload profiles, equipment staging, component failures, and control responses.

That capability matters when operators mix old and new infrastructure. A converted zone might perform correctly at average load but create an unacceptable condition when another chiller is offline. It might also satisfy thermal requirements while weakening the facility’s promised redundancy.

Digital models are not substitutes for field measurements. Their output depends on accurate geometry, equipment data, control logic, and operating assumptions. An incomplete model can provide a precise-looking answer to the wrong question.

The process should therefore begin with instrumentation and validation. Operators need credible data for actual power use, air and water temperatures, pressure, flow, humidity, and equipment performance. They must then compare modeled behavior with observed conditions before trusting future scenarios.

The central mechanism is iterative. Measure the existing facility, calibrate the model, identify the binding constraint, and simulate targeted changes. After installation, compare real performance with the modeled result and update the model again.

This is more demanding than a simple equipment replacement. It also provides a clearer basis for deciding when not to retrofit. A defensible model can show that an intervention would compromise resilience, exceed structural limits, or move the bottleneck without creating useful capacity.

Liquid Cooling Expands the Retrofit Envelope, With Limits

Liquid cooling makes denser retrofits feasible, but it cannot solve inadequate power, weak structures, or poor operational discipline.

Nvidia’s GB200 NVL72 shows why thermal architecture has become central to AI infrastructure. The rack-scale system connects 72 Blackwell GPUs and 36 Grace CPUs in a liquid-cooled design. Nvidia describes the system as one large NVLink computing domain for tightly coupled AI workloads.

The company claims the system delivers 25 times more performance at the same power than H100 air-cooled infrastructure under its stated comparison. That is a vendor benchmark, not an independent measure of every workload. Still, the rack architecture confirms that leading accelerator platforms increasingly arrive with liquid cooling built into their physical design.

Liquids transport heat more effectively than air, enabling operators to remove heat closer to the processors. Direct-to-chip cooling circulates coolant through cold plates attached to heat-producing components. Rear-door heat exchangers cool hot exhaust air as it leaves a rack.

These methods support different retrofit conditions. Rear-door systems can raise usable density without converting every server to direct liquid cooling. Direct-to-chip systems address heat at its source but require a reliable liquid loop, distribution hardware, monitoring, and service procedures.

Hybrid cooling combines liquid systems with traditional air cooling. That approach fits mixed facilities because liquid may capture most processor heat while air still cools memory, storage, networking equipment, or lower-density racks. It also supports staged adoption.

An international study released through the IEA’s technology collaboration program estimated potential energy savings of 8 percent at the server level and 30 to 40 percent at the facility level. It placed overall potential savings between 10 and 21 percent.

However, the same liquid-cooling study identified low current adoption, limited standardization, high initial costs, and concerns about long-term reliability. It also warned that Power Usage Effectiveness can understate some efficiency gains from liquid cooling.

Those qualifications matter. PUE compares total facility energy with energy delivered to IT equipment. It remains useful, but it does not capture every system consequence or describe the amount of useful computing completed.

A lower PUE also does not automatically make a retrofit economically or environmentally preferable. Operators must account for embodied materials, equipment replacement, water use, refrigerants, utilization, and the carbon intensity of electricity. A highly efficient but underused cluster can still waste resources.

Water strategy presents another tradeoff. Some liquid-cooling configurations can reduce dependence on conventional mechanical cooling. Others still require facility water loops and external heat rejection. Results vary with climate, temperature targets, equipment choices, and operating conditions.

Leaks receive attention because liquid is introduced near valuable electronics. Yet the operational challenge is broader than leak detection. Teams need compatible materials, water-quality controls, isolation procedures, spare parts, trained technicians, and clear responsibility across IT and facilities groups.

A retrofit can also create a two-speed operating environment. Conventional racks follow familiar procedures, while liquid-cooled equipment requires different commissioning and maintenance practices. Documentation, alarms, change control, and emergency response must accommodate both.

The skeptical conclusion is straightforward. Liquid cooling expands the set of buildings that can host AI, but it does not make every building suitable. Operators still need sufficient electrical capacity, structural support, heat rejection, and failure tolerance.

IES data center retrofits are strongest when cooling is one component of a verified system plan. They become risky when operators treat liquid cooling as a universal adapter between modern accelerators and legacy facilities.

Retrofitting and Greenfield Construction Serve Different Workloads

The real competition is not old buildings against new buildings; it is near-term deployment against long-term optimization.

A greenfield AI campus can be designed around large, homogeneous clusters. Engineers can plan utility feeds, substations, backup generation, network fabrics, cooling plants, and structural loads as an integrated system. That makes new construction the clearer choice for the highest-density deployments.

Retrofitting favors more bounded requirements. An enterprise might need a private inference cluster near regulated data. A colocation provider might convert part of a hall for customers that need accelerator capacity but not a hyperscale training campus. A research organization might add a liquid-cooled pod to an existing high-performance computing environment.

These uses do not require every site to match the scale of a dedicated AI factory. They require a suitable envelope with predictable performance. That makes segmentation more useful than asking whether an entire facility is AI ready.

A site might support lower-density inference equipment with air cooling. Another zone could accommodate rear-door heat exchangers. A third might support direct-to-chip cooling through a CDU. The rest of the building could remain dedicated to conventional workloads.

This mixed-density model lets operators match infrastructure to actual demand. It also limits the size of the initial commitment. Capacity can expand after customers demonstrate sustained utilization and the operating team gains experience.

Vertiv has argued that most early liquid-cooling deployments will occur in existing facilities. Its deployment guide emphasizes coordination among IT, facilities, and power teams because the building must be adapted around combined rack requirements.

That position comes from another infrastructure vendor and warrants the same caution applied to IES. Vendors benefit when operators purchase modeling, cooling, power, and service products. Their technical frameworks remain useful, but adoption forecasts should not be treated as neutral market measurements.

The competitive question turns on utilization. A retrofit completed quickly has little value if customers do not use it. A greenfield project delivered later can still outperform if it offers greater scale, lower operating complexity, or better access to long-term power.

Hardware cycles add another risk. A facility designed around one generation of rack density or coolant requirements may need further changes sooner than expected. Operators must distinguish durable infrastructure upgrades from adaptations tied too closely to a specific system.

Open interfaces can reduce that exposure. Standardized connections, maintainable piping, modular CDUs, flexible power distribution, and reserved service space give future equipment more room to change. Proprietary assumptions can turn a rapid retrofit into a new form of lock-in.

Resilience also separates credible projects from rushed ones. Colocation customers buy availability as well as power. An upgrade that adds AI capacity but weakens concurrent maintainability can damage the core service proposition.

The right comparison therefore depends on workload size, delivery date, location, power availability, and operating model. Greenfield construction remains the destination for many large clusters. Retrofitting provides a bridge for deployments that fit within verified limits.

This division of labor weakens the simplistic claim that one route will win. The likely winners will use both. They will retrofit where existing infrastructure provides an advantage and build where density or scale makes compromise unacceptable.

Three Signals Will Show Whether the Retrofit Advantage Is Real

The retrofit thesis will be tested by measured deployments, repeatable operating results, and evidence that projects preserve resilience.

The first signal is the conversion of modeled capacity into commissioned AI clusters. Operators should disclose the amount of usable IT capacity created, the types of cooling deployed, and the time from assessment to operation. Announcements without commissioning evidence will not validate the speed argument.

This signal strengthens the IES case if multiple facilities bring bounded AI zones online faster than comparable new construction. It weakens the case if structural, utility, or equipment constraints repeatedly stop projects after lengthy studies.

The second signal is operational performance across changing conditions. Useful reporting should cover utilization, cooling energy, water use, component failures, maintenance events, and performance during hot weather. Average PUE alone cannot establish that a mixed-density retrofit works reliably.

Results should also distinguish modeled estimates from measurements. A digital model earns credibility when observed temperatures, power flows, and control responses stay within predicted ranges. Persistent divergence would expose weaknesses in input data or system assumptions.

The third signal is customer adoption. Colocation providers need sustained demand for retrofit capacity, not just technical readiness. Contracted workloads, repeat expansions, and high utilization would indicate that customers value faster deployment inside established facilities.

Weak utilization would point toward a different constraint. Customers might prefer cloud access, larger greenfield clusters, or infrastructure aligned with a specific hardware generation. They might also hesitate because of availability guarantees or uncertainty around future cooling standards.

These signals should appear over several reporting cycles, not in a single launch announcement. Retrofitting is an operating strategy, so its success depends on repeatability after commissioning. The first installation proves only that one team completed one project.

For enterprise buyers, the immediate question is whether a provider can document the entire capacity chain. That includes utility supply, backup power, rack delivery, cooling, heat rejection, networking, structural support, and maintenance procedures. A claim of available floor space is not enough.

Infrastructure teams should also ask what happens during failure. Can the AI zone continue operating after a pump, CDU, power module, or cooling unit goes offline? Can technicians isolate a problem without interrupting adjacent conventional workloads? Does the retrofit preserve the contracted resilience level?

Developers and AI product teams have a stake in these answers. Infrastructure delays affect access to accelerators, deployment regions, model-training schedules, and inference costs. A successful retrofit program can add capacity closer to existing users and data sources.

Knowledge workers will experience the result indirectly. More distributed inference capacity can support lower-latency enterprise services and private deployments. Yet faster infrastructure growth also increases pressure on local grids, planning systems, and energy supplies.

IES data center retrofits deserve attention because they turn dormant or mismatched infrastructure into a possible source of near-term capacity. The proposal is credible only where detailed measurement confirms that the whole facility can support the change.

The next move belongs to operators. They should publish measured outcomes rather than broad claims of AI readiness. Buyers should ask for those results, compare them with greenfield alternatives, and treat speed as valuable only when it arrives with verified resilience.