Financial Modeling for Data Centers: The 7 Critical Assumptions
Most data-center financial models we review are financially clean and technically naive. The spreadsheet mechanics are impeccable; the assumptions feeding them come from brokers' decks. Here are the seven assumptions that actually decide whether your model predicts anything.
1. Compute depreciation is not building depreciation
A shell-and-core facility depreciates over decades. The GPUs inside it can be economically obsolete in three to five years, not because they stop working, but because the cost per unit of compute of the next generation makes them uncompetitive for training workloads. If your model treats IT equipment as a single line with a uniform depreciation schedule, your terminal value is fiction. Model compute in generations, with explicit assumptions about redeployment to inference workloads as each generation ages.
2. PUE is a sensitivity, not a constant
Power Usage Effectiveness drives the largest operating cost line, and it is anything but constant: it varies with load, climate, cooling technology and age. German regulation adds a hard edge, the Energy Efficiency Act (EnEfG) sets PUE ceilings for existing and new facilities. A model without PUE sensitivities cannot tell you what a mandated cooling retrofit does to returns. Ours treat PUE as a distribution with a regulatory ceiling, not a cell with a point estimate.
3. The utilization ramp hides the equity story
Between commissioning and stabilized occupancy lies the period in which equity is actually at risk. AI demand has made operators optimistic: pre-leased hyperscale capacity is real, but so are delayed grid connections and slipped chip deliveries. Stress the ramp (six, twelve, eighteen months of delay) and watch what happens to the IRR before you believe the base case.
4. Power price and PPA structure
Electricity is 40-60% of OpEx. How it is procured (merchant exposure, baseload PPA, pay-as-produced renewables with firming) changes both the cost line and the risk profile. The renewable share also feeds back into EnEfG compliance. A model that takes "€/MWh, flat, escalated 2%" as its power assumption has skipped the most structurable part of the deal.
5. Ground conditions are a CapEx line, not a footnote
Data centers are heavy, settlement-sensitive structures. Undetected soft layers or cavities become foundation redesigns, and foundation redesigns become double-digit CapEx overruns discovered after signing. This risk is measurable before the transaction. Modern subsurface surveys image ground risk at a fraction of the cost of a drilling grid, and yet it appears in almost no financial model we have seen. It should enter as a quantified CapEx sensitivity with a probability attached.
6. Regulatory readiness has a price
NIS2 classifies data-center services as critical digital infrastructure: registration, incident reporting, ISMS obligations, management accountability. EnEfG adds waste-heat and renewable-energy requirements. Neither is optional, and closing the gap costs real money, security organization, monitoring, possibly heat-offtake infrastructure. Regulatory due diligence should produce a number, and that number belongs in the model.
7. Waste heat: from cost to revenue line
The same law that constrains you creates an asset: waste-heat offtake to district heating or industrial users can turn a compliance obligation into a revenue stream, if the site's geology and surroundings support it. Whether that's realistic is a subsurface and infrastructure question first, a financial one second. When it works, it changes the asset's ESG profile and its exit multiple.
The pattern
Every one of these assumptions is an engineering, geology or regulatory question before it is a modeling question. That's why we build financial models alongside technical, regulatory and geological due diligence, the findings of each workstream enter the model as quantified sensitivities, not as disclaimers in an appendix.