Machine learning approaches to dynamic TCO modeling under global warming scenarios across 10 global markets
This research extends prior Monte Carlo simulation work on global datacenter location optimization by introducing climate-dynamic TCO modeling and extreme weather impact quantification. Analyzing 10 strategic global markets across 100,000+ simulation iterations, we demonstrate that climate change amplifies existing location advantages—widening the TCO gap between Nordic and traditional hub locations from approximately 74% (static baseline) to over 80% under RCP 8.5 by 2050. We integrate IT hardware costs (4 tiers including GPU and ARM architectures) to show that hardware energy is the dominant channel through which climate change affects datacenter economics, with a 4.6x climate exposure ratio between Boden, Sweden and Atlanta, Georgia.
The facility-only climate gap grows 4.9x from 10-year to 25-year horizons. Atlanta faces +$37.1M in climate premium vs +$16.6M for Boden over 25 years.
For a 100MW AI facility (500 racks), hardware energy costs range from $203M (Boden) to $1,074M (Atlanta) under RCP 8.5—a 5.3x spread driven by PUE degradation.
Boden, Sweden maintains cost leadership under all climate scenarios with only +$47M climate premium over 25 years, compared to +$179M at Atlanta.
ARM-based hardware (Apple silicon) reduces hardware energy by 85%, saving $131–491M per facility. Partially compensates for climate-vulnerable locations.
Georgia averages 149 FEMA disaster declarations/year vs Wyoming's 9—a 16x disparity. Survival model concordance: 0.94 using real data.
Bayesian Neural Network uncertainty bands (2σ) encompass physics predictions across all locations, with 94–100% calibration coverage.
100MW reference facility, 500 Standard AI racks (H100-based), climate-adjusted
| Location | Tier | RCP 2.6 | RCP 4.5 | RCP 8.5 | Climate Gap |
|---|---|---|---|---|---|
| Boden, Sweden | Nordic | $14,300M | $14,314M | $14,347M | +$47M |
| Iceland (Reykjanes) | Nordic | $14,336M | $14,355M | $14,388M | +$52M |
| Kristiansand, Norway | Nordic | $14,328M | $14,347M | $14,381M | +$53M |
| Evanston, Wyoming | US Secondary | $15,200M | $15,224M | $15,262M | +$61M |
| Sines, Portugal | Emerging | $14,831M | $14,873M | $14,935M | +$104M |
| Florence, SC | US Secondary | $15,487M | $15,534M | $15,602M | +$115M |
| Johor, Malaysia | Emerging | $16,037M | $16,105M | $16,206M | +$170M |
| Atlanta, Georgia | Traditional | $16,220M | $16,292M | $16,398M | +$179M |
US retail electricity prices and generation mix by state (2015–2025). 55 price records + 4.5MB generation data for 5 US locations.
Historical daily temperature, precipitation, and wind data. 46,044 records across 9 weather stations near datacenter locations.
69,769 disaster declarations (1953–2026). Georgia: 149/yr vs Wyoming: 9/yr—real extreme event frequency for insurance modeling.
Monte Carlo simulation parameters for 10 global markets. Facility TCO, CapEx, PUE, power rates, and risk metrics from 100,000 iterations.
Physics-based shift functions replace static distributions with year-by-year climate-adjusted parameters. PUE degrades at 0.008–0.012 per °C with humidity compounding.
PUE(t) = PUE_base + ΔT × β × (1 + H_f) × (1 + 0.05ΔT)
Weibull AFT model predicts time-to-outage from climate covariates and real FEMA disaster data. Concordance index: 0.94.
log T = βTx + σε, ε ~ Gumbel
Quantile Gradient Boosted regression with nonlinear event scaling. Premium ~ event_ratio1.5. RMSE: $0.71M.
σ_ins(t) = (E(t) / E_0)1.5
MC Dropout BNN (12K params) independently validates physics shifts. 94–100% of true values within 2σ uncertainty bands.
p(y|x) ≈ (1/N) ∑ f(x; θ_n)
Climate impact multiplied across IT equipment at 100MW scale
Climate Δ (25yr): Boden +$99M vs Atlanta +$456M
85% energy reduction. Saves $131–491M per facility.
Balanced training/inference. Mid-range climate exposure.
Lowest CapEx, limited AI throughput. Edge/inference only.
Full source code, pipelines, configs, and LaTeX paper. MIT License.
6 executed notebooks, processed data, FEMA records, hardware TCO results.
Full research paper with methodology, results, and analysis.
Full manuscript with methodology, results tables, and bibliography.
6 Jupyter notebooks covering EDA, dynamic TCO, survival analysis, BNN, and synthesis.