Method & how much to trust it
What is and isn't modelled
- Default = firm, always-on: gas backup covers 100% of load during lulls, so the worst case is a known, capped fuel cost. Premium/AI workloads never shed and collapse to this case; interruptible (spot) workloads can shed load when the lost compute is worth less than the gas to serve it (
--workload). - Optimisation: for each region and year, the least-cost solar overbuild × wind overbuild × battery hours meeting the renewable target, evaluated on chronological hourly dispatch; gas is sized to the peak residual deficit, never a decision variable. Every reported optimum is confirmed with exact dispatch.
- Costs: Wright's-Law learning curves on an S-curve deployment trajectory, per-technology WACC, battery augmentation (top up faded cells, not mid-life replacement), CCGT/OCGT selection, EU ETS carbon path. All in real 2025 USD.
- Single real site per region — the largest directional caveat: one off-grid datacenter gets no geographic smoothing; a multi-site portfolio (
--sites) softens multi-day lulls and lowers high-renewable cost. The documented experiment (§4.7) puts the effect at ≈40% off the EU 90%-RE delivered cost for 3–5 partially-correlated sites (≈$161 → $91–98/MWh, 2025, reduced fidelity) — the direction is robust; the magnitude depends on the calibrated inter-site correlation. Every high-renewable number on this site is therefore a single-site worst case. - Not modelled: sub-hourly load variation, on-site fuel logistics, transmission. Embodied carbon and water use are discussed qualitatively on the Zero-carbon page but not priced.
What the optimiser actually builds
Cost breakdown
The delivered cost, split by what is actually paid for. Two points: the battery is a small share at every target — high-renewable economics are dominated by generation overbuild and firming, not storage — and moving from 70% to 85% renewable is bought almost entirely with more generation capital (the overbuild that rides out multi-day lulls), while the firming and carbon slices shrink.
What moves the answer most
One-at-a-time swings of the key assumptions around the base case, measured as the change in the parity gap (the 90%-renewable build's delivered cost minus gas, at 2030; base +37 $/MWh). The ranking: Wind resource ±10% swings the gap by $31/MWh and Gas price ∓25% by $22 — together they account for most of the range — while the smallest levers (RE WACC 4% / 7%, $2; Battery capex ∓20%, $3) change it little. Whether deep-renewable Europe reaches parity with gas depends on wind quality and the gas price, not on battery costs or financing.
Reduced-fidelity, one-at-a-time swings on the synthetic weather generator (the free-knob mode behind sensitivity runs; the headline itself uses measured ERA5) — at this fidelity the firm battery-only system tops out just below the 90% target, so the gap is measured at the ~89% penalty optimum. Treat magnitudes as indicative and the ranking as robust. Reproduce: python datacenter_lcoe.py --tornado --region eu → output/eu_tornado_results.json.
How this compares with other studies
No published study prices exactly this build (islanded, always-on, real single-site weather), but the closest literature brackets it — and agrees on the shape: hourly-matched clean power is cheap until the last few percent, which only long-duration storage or clean-firm capacity closes affordably. Grid-connected studies land below this model (they sell surplus and lean on the grid); islanded studies with sunnier sites or more aggressive cost projections land at or below it too. Nothing in the literature contradicts the directional findings; the levels differ for stated, checkable reasons.
| Study | What it prices | Their finding | This model |
|---|---|---|---|
| Riepin & Brown 2022/24 (TU Berlin, PyPSA) | Grid-connected 24/7 hourly-matched clean supply, Germany / Ireland 2025, surplus sold to the grid (€2020) | Wind+solar+battery only: 100% hourly matching costs €194–229/MWh (+141–242% vs annual matching); adding hydrogen LDES brings it to €99–114; 90–95% matching costs only a small premium — the last ~2% roughly doubles the cost | Same non-linearity, same fix: this model's battery-only 90% RE is $278/MWh while the H₂-firmed 100% build is $191 — deep targets need LDES, not more batteries. Levels sit above theirs because this build is islanded (no grid sales) on poorer-wind France |
| Xu, Manocha, Patankar & Jenkins 2021 (Princeton ZERO Lab) | Grid-connected 24/7 CFE, California / PJM ~2030 ($2020) | 100% CFE ≈ $68–100/MWh all-in; the premium is strongly non-linear (98→100% costs more than 88→98%) and shrinks as the surrounding grid decarbonises | Consistent: this model's islanded US 90% RE is $137/MWh (2025) and the gas-free H₂ system $142→$106 by 2040 — dearer than grid-connected 24/7 CFE, which is exactly the off-grid premium the Overview flags |
| offgridai.us 2024 (Baranko, Campbell, Hausfather, McWalter, Ransohoff) | Fully islanded solar+battery+gas microgrid, US Southwest (~$2024) | 90% solar-served load at $97–109/MWh (optimised $97; solar $0.75/W, batteries $120/kWh) | Closest architecture to this model's US case: this model's 80% RE is $89/MWh and 90% $137 (2025, Texas ERA5) — the same ballpark, with this model's 90% dearer mainly via costlier battery and firm-always-on assumptions |
| Fasihi & Breyer 2020 (LUT) | Off-grid PV+wind+battery+H₂ firm baseload, best global sites (Maghreb-class resource, 7% WACC) | <€119/MWh in 2020 falling to ≈€54 by 2030 on projected technology costs | This model's H₂ system is $191 (2025) → $166 (2030) on French resource; its scan prices the Maghreb edge at ~$111–120 in 2030 — the remaining gap to €54 is their more aggressive 2030 solar/electrolyser cost projections |
Literature numbers are quoted in each study's own currency-year (Riepin & Brown €2020; Princeton $2020; offgridai ≈$2024) — inflate ~15–20% to compare €/$2020 with this model's real-2025 USD. All model numbers in this table are generated from output/*.json.
Key assumptions
| Assumption | Europe | US | Source |
|---|---|---|---|
| Solar PV — LCOE₀ · learning rate | $60/MWh · 25% | $52/MWh · 25% | Lazard v18; Way et al. Joule 2022 |
| Onshore wind — LCOE₀ · LR | $56/MWh · 17% | $61/MWh · 17% | Lazard v18; OWID |
| LFP battery — energy · power | $90/kWh · $120/kW | $90/kWh · $160/kW | BloombergNEF 2024–25; Ember 2025 |
| Gas price | $10/MMBtu | $4/MMBtu | EIA Henry Hub / TTF forward |
| Carbon price trajectory | logistic $70→$200/tCO₂ | linear $0 | EU ETS Fit-for-55 |
| WACC (solar/wind · battery · gas) | 5.5% · 7% · 9% | 5.5% · 7% · 9% | NREL ATB 2024; merchant spread |
| Site capacity factor (solar / wind) | 0.18 / 0.13 | 0.26 / 0.32 | measured ERA5 (France / Texas); costs re-levelled to the site CF |
All in real 2025 USD; costs fall over time via learning curves. Full derivations, data sources and the accuracy summary: model_documentation.md.
Glossary
| LCOE | Levelized Cost of Energy — the all-in cost of one delivered MWh ($/MWh) once capital, fuel, carbon and O&M are spread over the project's life. The model's headline output; lower = cheaper power. |
|---|---|
| Capacity factor (CF) | Average output ÷ nameplate rating. A solar farm with CF 0.23 produces 23% of its peak rating averaged over the year. |
| Renewable target / "90% RE" | Share of the datacenter's served energy that must come from renewables + storage (the rest is gas). |
| Firm / always-on | A datacenter that never shuts down: gas backup is sized to cover 100% of load during lulls, so the worst case is a known, capped cost. The model's default. |
| Overbuild / curtailment | Installing more solar/wind than peak load (e.g. "6× solar") so enough is made on poor days; the excess on good days is curtailed (spilled). |
| Dunkelflaute | German for "dark doldrums" — a multi-day, wide-area spell of low sun and low wind. It sets how much storage/backup a renewable system needs. |
| Firming | Whatever covers the hours renewables + battery cannot: a gas turbine, green hydrogen, pumped storage, hydro, or nuclear. |
| WACC | Weighted Average Cost of Capital — the financing rate; the model uses a different one per technology (solar/wind 5.5%, battery 7%, gas 9%). |
| Learning rate (Wright's Law) | A technology's cost falls a fixed % for every doubling of cumulative production — this drives the cost-over-time trajectories. |
| CCGT / OCGT | Combined-/Open-Cycle Gas Turbine — efficient baseload vs a cheap-to-build peaker; the model picks whichever is cheaper for the gas duty. |
| SMR | Small Modular (nuclear) Reactor — plotted as a firm-clean reference line, never part of the optimisation. |
| PPA / 24/7 CFE | Power Purchase Agreement — a long-term renewable supply contract (the on-grid alternative). 24/7 CFE is the stricter hour-by-hour carbon-free matching standard; both are plotted as reference lines. |
| PHS / LDES | Pumped Hydro Storage / Long-Duration Energy Storage (iron-air, hydrogen) — multi-day storage options that can replace residual gas. |
| Crossover / parity | The year a build's LCOE drops below the gas baseline — when going (mostly) renewable becomes the cheaper choice, not just the greener one. |
Reproduce
pip install -r requirements.txt, then make reproduce && make report regenerates every figure, the exports and this site. The model is pure Python and runs fully offline.
output/*_firm_results.json at commit 8ab84d0 (config 40172b07822a49e6) · source on GitHub · licensed CC BY 4.0. Reproduce: make reproduce && make report.This model was created in early June 2026, before I joined the AI Office in the European Commission. Nothing here represents the opinion of the AI Office or the EU Commission.