Financial Modelling for Data Centres: Seven Assumptions That Decide the Investment Case

Where engineering reality enters the spreadsheet

At first sight, a data-centre model can look wonderfully complete. Rows extend across twenty years. Debt service is calculated to the cent. Escalation rates, exit multiples and tax effects sit in their proper cells. The workbook balances.

Then one asks how the GPUs will age, how PUE changes at partial load, what happens if energisation slips by twelve months or whether the ground beneath the heaviest halls has actually been characterised. The apparent precision begins to thin.

This is the central weakness in many data-centre models: the spreadsheet is financially disciplined, but the assumptions entering it are technically naive. The most important variables are not accounting conventions. They are physical and operational conditions. Seven of them, in particular, determine whether the model describes an investable asset or merely a tidy expectation.

1. Compute does not depreciate like the building around it

The shell, structure and much of the electrical infrastructure may remain useful for decades. The accelerators inside can lose their economic position within three to five years. They have not necessarily failed. They are simply overtaken by a new generation capable of producing more compute for every kilowatt-hour consumed.

That distinction matters because an operating GPU and a competitive GPU are not the same asset. Training workloads tend to migrate towards the newest hardware, where interconnect, memory and performance per watt justify the price. Older equipment may move down a cascade: from flagship training to fine-tuning, from fine-tuning to inference, and eventually to specialised or lower-value workloads.

A model that places all IT equipment on one uniform depreciation line misses this migration. Compute should be modelled in generations, with an explicit revenue and utilisation profile for each stage. Power price should be linked to the cascade because the same older GPU fleet may remain profitable in a low-cost facility and become uneconomic in a high-cost one. Terminal value without this logic is less a forecast than an assumption concealed by formatting.

2. PUE is not a constant

Power Usage Effectiveness is often entered once and copied to the right. The physical system does not behave that way. PUE changes with IT load, outdoor temperature, humidity, cooling architecture, maintenance condition and the age of equipment. A facility operating below design load may carry fixed cooling and electrical losses over fewer productive kilowatts. A hot period may push cooling consumption upward just when grid conditions are most strained.

German regulation gives this engineering variable a legal boundary. The Energy Efficiency Act establishes efficiency requirements for data centres, making PUE not merely a performance metric but a potential retrofit driver.

A credible model should therefore use PUE sensitivities or a distribution rather than a single immutable number. It should show how part-load performance, climate variation and a mandated improvement programme change operating cost. If the model cannot explain what a cooling retrofit does to capital expenditure and returns, it has not yet captured the risk that PUE represents.

3. The utilisation ramp carries the equity risk

Commissioning is a date. Stabilised occupancy is a condition reached later. Between them lies the period in which fixed costs have arrived but contracted revenue may not have done so in full.

Strong demand for AI capacity can make this interval appear harmless. Pre-leasing is real, but so are delayed grid connections, construction slippage, incomplete customer fit-outs and late delivery of compute hardware. A hall may be technically available while the workload intended for it is not.

This is where the equity story lives. The model should stress six-, twelve- and eighteen-month delays in the utilisation curve and track the consequences for liquidity, interest during construction, covenant headroom and internal rate of return. A base case that survives only when every dependency arrives on schedule is not a base case. It is a coordination miracle.

4. Power price is inseparable from procurement structure

Electricity commonly represents the largest operating cost. Yet a single flat euro-per-megawatt-hour assumption says almost nothing about how that electricity will actually be bought.

Merchant exposure, a baseload power purchase agreement and a pay-as-produced renewable PPA with firming can produce similar average prices in a simple model while creating very different hourly risk. Shape, imbalance, curtailment, volume mismatch, grid charges and escalation clauses determine whether the apparent hedge performs when the data centre is drawing power continuously.

Renewable procurement also interacts with regulatory requirements and the asset's commercial positioning. The power case should therefore be built as a structure, not a price: contracted volumes, delivery profile, residual exposure, firming cost, indexation, tenor and counterparty strength. This is one of the most designable parts of the transaction. Treating it as a single escalated number gives away that advantage.

5. Ground conditions belong in capital expenditure

A data centre is a dense collection of settlement-sensitive loads. Buildings, tanks, generators, transformers, cable routes and cooling infrastructure all depend on the ground behaving within narrow limits. A soft layer, an old working or a cavity discovered after foundation design can turn a standard solution into piling, grouting or ground improvement.

The financial problem is not only the cost of the remedial work. It is the time at which the condition is discovered. Late information affects design, procurement sequence, lender confidence and the construction programme simultaneously.

Modern subsurface investigation can map anomalous zones across a site before the transaction or design becomes difficult to change. The findings should enter the model as probabilities and scenarios: the likely foundation case, a defined remediation case and the cost and delay attached to each. Ground risk is not a footnote to technical due diligence. It is an uncertain capital line waiting to be measured.

6. Regulatory readiness has a calculable price

Data-centre services sit within Europe's critical digital infrastructure framework. NIS2 and its national implementation bring registration, risk-management, supply-chain, incident-reporting and management-accountability obligations. The German Energy Efficiency Act adds efficiency, renewable-energy and waste-heat requirements.

Compliance gaps are not abstract legal observations. They require people, systems and construction. A weak security organisation may need monitoring, governance and incident-response capability. An inefficient cooling system may require physical modification. A credible waste-heat concept may depend on a new connection, an external offtaker and a commercial agreement that does not yet exist.

Due diligence should translate these gaps into cost, timing and dependency. The model can then distinguish operating expenditure from capital works, identify what is required before closing or operation, and show the effect of non-compliance or delay. A legal annex may describe an obligation accurately. Only a quantified remediation plan shows what it does to value.

7. Waste heat can become an asset - but only where a system exists to use it

Every megawatt of electricity consumed by compute eventually appears as heat. In the server hall it is a cooling load. Beyond the site boundary it may be a resource, but only if temperature, distance, demand profile and infrastructure align.

A nearby heat network or industrial user can turn waste-heat recovery into a revenue or cost-offsetting line. Seasonal mismatch, low delivery temperature, connection cost and uncertain offtake can turn the same concept into an obligation without an economic outlet. Geology may also matter if the subsurface is considered as a seasonal thermal buffer.

The model must therefore begin with physical feasibility. How much usable heat is available, at what temperature, during which hours and at what boundary? Who needs it, how far away, and what equipment must sit between the two systems? Only after those questions are answered does a revenue line become defensible.

The pattern behind the seven assumptions

Each of these variables begins outside finance. Compute depreciation begins with chip performance and workload economics. PUE begins with cooling and load. Utilisation begins with delivery dependencies. Energy cost begins with procurement structure. Ground risk begins below the site. Regulatory readiness begins with actual systems and processes. Waste-heat revenue begins with thermodynamics and infrastructure.

The model becomes useful when these disciplines meet inside it. Technical, geological and regulatory findings should appear as quantified sensitivities, not as cautious language in separate reports. That is how an investor can see which risks affect price, which can be engineered out, which should be transferred and which remain fundamental to the asset.

A spreadsheet cannot make an uncertain project certain. It can, however, show honestly where the uncertainty resides. That is the difference between a model that calculates and one that informs a decision.