| Vose Software

Industry: Telecommunications and IT
Product: ModelRisk
Application: Data center optimization


A 10-Year TCO Curve for a 200-MW Data Center Estate

A modern hyperscale-style data center runs at PUE between 1.2 and 1.6 and consumes electricity at industrial-tariff rates of $0.06–$0.18/kWh depending on region — meaning a 40 MW IT-load campus burns through $25M–$70M in electricity alone per year before any allocation of capital, staff, or maintenance. Multiply by a five-campus estate over a 10-year horizon and the cumulative power bill alone clears $1.5 billion, with a spread driven by three things the deterministic model treats as constants: demand growth, PUE drift, and tariff volatility.

A telecom-and-IT services operator rebuilt its 10-year TCO and capacity-expansion model in ModelRisk to make the spread visible. The deterministic plan said "TCO ≈ $2.5B". The simulation said "mean $3.56B, P10 = $2.88B, P90 = $4.52B — and with the current incremental build plan there is a 63% chance you breach the capacity envelope before year 6."

10-year estate TCO — distribution and tail

The deterministic $2.5B sat below even the 1st percentile of the simulated distribution: it was not a central estimate at all, but an optimistic floor. The whole purpose of the rebuild was to replace that single number with the curve — its mean, its tail, and the capacity-breach probability the point estimate could never expose.

Demand growth: not a single CAGR

The previous plan used a single point-estimate of 8%/yr compound growth in IT load. The rebuilt model treats annual growth as a truncated Normal(mean 8%, σ = 4%, clipped to [0%, 20%]) — bounded because runaway demand growth saturates at physical and budgetary limits, and a plain Normal would draw negative growth. Compounded over 10 years against a 200 MW year-0 estate load (five campuses of ~40 MW IT load each), this produces a year-10 estate-load distribution where:

  • P10 year-10 estate load: ~370 MW
  • P50: ~430 MW
  • P90: ~500 MW

That ~130 MW spread between P10 and P90 is the entire capacity-planning problem in one number.

PUE drift

PUE is not a constant — it drifts as hardware ages, free-cooling availability changes with climate, and workload mix shifts. The model treats per-site PUE as LogNormal(median 1.35, σ_log = 0.06), bounded below at 1.10 (Carnot/thermodynamic floor for the cooling architecture). A LogNormal is the right shape: PUE can spike upward when a chiller fails or ambient conditions degrade, but cannot fall below the physical floor.

Energy per year per site is then:

E_year = IT_load_MW × 8760 h × PUE × (1 - free_cooling_hours_fraction)

producing energy distributions that combine demand uncertainty with PUE drift.

Electricity-price modelling: regime-switching, not GBM

The first iteration of the model used geometric Brownian motion for electricity prices. That is the standard textbook choice and it is wrong for industrial-tariff power: prices in the 2021–2024 European and North American records show regime switches (the gas-price shock of 2022 doubled industrial-tariff power overnight in many countries), not the smooth log-Normal drift GBM produces. The rebuilt model uses a two-regime mean-reverting process with Poisson regime-switch arrivals at λ = 0.12/yr (about one shock per decade per region). In the "calm" regime, price reverts to the long-run mean of $0.10/kWh with annual vol 12%; in the "shock" regime, price reverts to $0.18/kWh with annual vol 22%. Without the regime-switching layer the model under-estimates 10-year energy-cost variance by a factor of nearly 2.

Combining into TCO

Per-year cash flow per site:

Cost_year = Energy_cost + Maintenance + CapEx_amortised + Labor
TCO_10yr  = Σ_t Cost_t / (1 + r)^t  with r = 8%

across five regional sites. The simulation runs 50,000 trials over the 10-year horizon.

The deterministic plan reported TCO of ~$2.5B. The simulation produces a mean of $3.56B with a P90 of $4.52B — the deterministic figure understates the mean by more than $1B and the P90 by roughly $2B. The right tail comes from the joint event "demand at P75 and a sustained price shock in years 4–8" — neither extreme on its own, but together they swing TCO by hundreds of millions.

The capacity-envelope problem

A pure-TCO view hides the operationally critical question: when do we run out of capacity? With the current build plan (modular adds at years 2, 5, and 8), the simulation says:

  • P(capacity breach in any of the 10 years) = 67%.
  • P(capacity breach before year 6) = 63% — driven by the upper half of demand growth catching the thin headroom the modular adds leave between build steps.
  • P(stranded capacity > 30% in year 4) = 0.1% — the current plan runs tight, not slack, so over-provisioning is essentially never the problem.

These are not symmetric problems: a capacity breach forces emergency colocation procurement at a premium and risks customer-facing outages; stranded capacity is a sunk-cost annoyance. The simulation makes plain that the current incremental plan errs heavily toward the breach side — capacity runs hot for most of the decade — which is exactly why the build-strategy comparison below matters.

Which uncertainty hurts most?

Sensitivity on P90 TCO ranks the inputs by tail leverage:

What drives the P90 10-year TCO

Demand-growth volatility dominates — a wider σ on the growth rate moves P90 TCO by ~$140M. The electricity-price regime-shift probability is second; PUE drift third. The standard sensitivity treatment that locks PUE at 1.35 and varies it ±5% gets the ranking wrong because it ignores the joint event of high demand × shock-regime electricity.

Three build strategies

Three CapEx strategies were simulated against the same demand/price distributions:

Three build strategies — 10-year TCO CDFs

  • Front-loaded build (proactive). Build to year-10 P75 demand in year 0. Mean TCO $3.51B, P95 $4.86B — the lowest of the three, because building the capacity once up front avoids both the breach premium and the option carry. Its weakness is stranded capacity in the futures where demand disappoints, not TCO.
  • Incremental adds (current plan). Modular adds at years 2, 5, 8. Mean TCO $3.56B, P95 $4.91B, but it carries the 63%-by-year-6 breach exposure above.
  • Flex-contract hybrid. Smaller in-house build plus pre-purchased colocation options that activate above a demand threshold. Mean $3.86B, P95 $5.50B — the most expensive, because the colocation overflow is priced at a 70% premium to owned-capacity energy and gets used heavily once demand outruns the deliberately small in-house build.

The result overturns the usual intuition: on this model the front-loaded build is cheapest on both mean and tail, and the flexibility of the hybrid is paid for in a materially higher TCO. The decision is therefore not "which is cheapest" but how much the operator will pay in expected TCO to avoid committing CapEx against demand that might not arrive — exactly the trade-off a single deterministic number cannot frame.

What changed

  • Build plan moved toward front-loading. The simulation showed the current incremental plan runs into a 63%-by-year-6 breach probability for essentially the same TCO as a front-loaded build that eliminates it — so capacity was pulled forward rather than left to modular triggers, and the colocation-heavy hybrid was rejected once its 70%-premium overflow cost showed up as the highest TCO of the three.
  • Energy hedge programme initiated. P90 electricity cost informed a 40% volume hedge at a fixed forward price for years 2–5, capping the regime-shock exposure.
  • PUE-improvement business case repriced. A $48M CapEx for advanced liquid-cooling that drops median PUE from 1.35 to 1.22 was approved on the expected-savings plus tail-clip basis — the deterministic IRR alone hadn't cleared the hurdle.

ModelRisk Functionality Used

  • Truncated Normal on annual growth instead of unbounded Normal, preventing the model from drawing negative or runaway compound-growth rates.
  • LogNormal PUE with a hard lower bound at the cooling architecture's thermodynamic floor — a Normal would draw PUE < 1, which is physically impossible.
  • Two-regime mean-reverting electricity-price process with Poisson regime-switch arrivals — corrects the GBM-on-trending-commodity error that suppressed energy-cost tail variance.
  • Capacity-envelope tracking per simulation trial producing the P(breach) and P(stranded) probabilities — not derivable from a deterministic plan.
  • Tornado on P90 TCO ranking demand-growth volatility and regime-shift probability ahead of PUE — directing the next-dollar-of-investment debate.
  • CDF overlay across three build strategies that surfaced the flex-contract hybrid as the dominant choice on the (mean cost, tail risk) plane.

The deterministic plan's $2.5B TCO is one number sitting below the bottom of a curve that runs past $2.9B at the P10 and $4.5B at the P90. Monte Carlo turns the curve into the actual planning artefact — and the artefact, not the single number, is what gets the capacity-strategy and hedge decisions right.