The cascade nobody budgets for
When a colocation operator commits to serving AI workloads, the infrastructure gap is easy to see: current average rack power density in enterprise data centers is approximately 12kW per rack (per AFCOM 2025), while AI deployments routinely require 30–50kW or more. The delta is filled by upgrading cooling, power distribution, and UPS systems.
The mistake is treating those three upgrades as a list rather than a system.
In a legacy data center, each infrastructure layer was designed with implicit assumptions about the other layers. Cooling capacity was sized for a power density that pre-dates GPU clusters. PDU and busway sizing was specified for loads that don’t have AI step-load profiles. UPS systems were selected for battery chemistries and ambient temperatures that have since changed. When one layer is changed, those implicit assumptions break — and the breakage cascades.
Synthesis: The cascade is not a project management problem. It is a physics problem. The systems are coupled. Changing one system changes the operating conditions of the other two in ways that only become visible during commissioning — which is after the money has been committed.
The failure sequence
The cascade follows a consistent sequence when operators attempt component-by-component upgrades.
Step 1: Cooling upgrade. Liquid cooling infrastructure is installed to serve high-density rows. Manifolds are routed through existing overhead space. During layout, conflicts emerge with cable tray and busway that were positioned for air-cooled PDU placement — the two physical systems compete for the same overhead corridors. Resolution requires electrical infrastructure redesign that was not in scope.
Step 2: UPS upgrade. Replacement UPS units are specified for the new power density and the new AI load profile. During commissioning, step-load testing reveals that the existing power distribution circuit impedance creates voltage deviations that fall outside the new UPS’s transfer criteria. The UPS specification was correct; the distribution impedance assumption in that specification was not. Power distribution redesign is now on the critical path.
Step 3: Power distribution upgrade. New busway or FMP distribution is installed for the required density. During commissioning, the interaction between new distribution impedance and the new UPS step-load response must be verified — but the cooling system has now changed the thermal environment in the server room, which changes the battery ambient temperature, which affects battery state-of-health in ways that were not modeled when the replacement UPS was specified.
Each step generates rework in at least one of the other steps. The total cost of three sequential upgrades with rework loops exceeds the cost of a single concurrent design. The excess cost arrives as change orders, not as line items in the original budget.
[KNOWLEDGE_GAP: Published retrofit cost-per-kW data broken down by concurrent vs. sequential execution, and change order frequency data for staged vs. concurrent data center infrastructure upgrades, are not available in current knowledge files. The pattern above is consistent with practitioner accounts but has not been rigorously quantified in published research.]
Why the density threshold matters
Not every legacy data center faces this cascade. The severity of the coupling between systems increases with density.
At average rack densities below approximately 15kW, air cooling remains viable, UPS step-load dynamics are within the tolerance of most installed systems, and power distribution impedance is not a precision variable. Staged upgrades are feasible.
Above 20kW average density — and certainly at 30kW+ — all three dependencies become hard constraints simultaneously. Liquid cooling is no longer optional. UPS step-load response is a specification item, not a tolerance. Power distribution impedance affects voltage quality in ways that matter for GPU cluster operation.
Synthesis: Equinix has reported that AI-driven deals require approximately 33% higher power density than general colocation demand. If current average density is approximately 12kW, AI-specific deployments are landing at 16kW and above — already inside the zone where the cascade risk is real, and trending toward densities that are firmly inside it.
The revenue cost of the density cap
Operators who recognize the cascade risk but cannot execute a concurrent upgrade often respond by capping the density they will accept from AI customers. This is rational in the short run. In a market where AI is reportedly driving 60% of large colocation deals (per Equinix disclosure), it may not be rational over an 18-month horizon.
The revenue loss from density refusal compounds: a customer who is turned away at 30kW buys capacity elsewhere, typically at a competitor who built for AI density. That customer relationship is not recoverable. The competitor’s operational experience with high-density AI workloads accelerates.
[HYPOTHESIS: Operators who accept a density cap rather than execute a retrofit or greenfield build will lose 40–60% of addressable AI pipeline within 18 months. This estimate is directional, based on the stated AI share of current deal volume and reasonable assumptions about AI infrastructure demand growth. It has not been quantified in any sourced publication.]
The two viable strategies
Zone-by-zone full re-core. Identify a physical zone — a pod, a cage block, or a defined row section — that can be isolated from adjacent zones while infrastructure work proceeds. Design cooling, power distribution, and UPS upgrades for that zone as a single integrated system. Execute all three concurrently. Commission the zone as a unit before serving AI customers into it. Add capacity by adding re-cored zones.
Requirements: adequate physical zone isolation, ability to maintain adjacent zone operations during work, and sufficient floor area to stage equipment. Not all legacy facilities can provide these conditions.
Greenfield build. Design for AI density from the start. Higher upfront capital. No retrofit complexity, no change order risk, faster time to customer revenue. The economics favor greenfield when: the legacy facility’s remaining useful life is short, land and power are available in the target market, and the re-core cost per kW exceeds greenfield cost per kW by more than the NPV of the legacy facility’s existing revenue.
What is not viable: upgrading any one of the three systems while deferring the others, at densities above the legacy ceiling.
The evaluation framework
Before committing capital to a retrofit strategy, three questions determine which path is viable:
Zone isolation: Can any portion of the facility be physically isolated enough to execute a concurrent re-core without disrupting adjacent operations? If yes, re-core is viable. If no, legacy facility retrofit at AI density is not viable.
Re-core cost vs. greenfield cost: What is the all-in cost per kW for concurrent cooling + power + UPS upgrade in an isolated zone, including change order contingency? How does that compare to greenfield cost per kW in the same market? The comparison needs to include time-to-revenue, not just capital cost.
Density timeline: What density does the AI customer pipeline require, and when? If the timeline to re-core completion exceeds the time before density demand arrives, the window for re-core may have closed.
Synthesis: Most operators who are asking “how do we retrofit for AI?” are implicitly asking “how do we stage the upgrade to manage cost and disruption?” The answer is that staging is not available at AI density. The question to ask instead is “which zones of this facility can be re-cored concurrently, and what is the cost and timeline?” That question has a real answer.