How expensive is it to run a datacenter mostly on renewables?

A techno-economic model of the least-cost way to run an always-on, off-grid datacenter on (mostly) renewable power — solar, wind, batteries and a backstop — and of when that costs less than burning gas, across Europe and the US, 2025–2040.

▶  Source code & full methodology on GitHub

The question

AI datacenters use a large and growing amount of electricity. This model asks the narrower question each builder controls: if you pair a new datacenter with its own solar, wind and batteries, how clean can it run, and what does that cost compared with just burning gas?

A datacenter needs power every hour, and sun and wind don't deliver every hour. The bill therefore hinges on the backstop for dark, windless spells (a Dunkelflaute): a gas turbine, which emits, or a clean option — hydrogen, pumped storage, hydro, nuclear. The model finds the least-cost mix and the delivered cost per MWh (LCOE — levelized cost of energy), for Europe and the US, every year to 2040.

Key findings

• Europe: moderate renewable shares cost less than carbon-priced gas; the last decile is a large premium. A 70% renewable build crosses below gas ~2027 and 80% ~2033, but on France's measured weather 85% reaches parity only ~2040 and 90% not within the horizon.
  More

  The baseline is gas rising $122→$163/MWh as EU carbon prices climb. France's measured wind is poor (capacity factor 0.135), so riding out multi-day winter Dunkelflaute takes heavy overbuild — roughly 10× solar + 9× wind + 6h battery at the 90% target. Moderate renewables are the low-cost portion; the last decile is a premium on a poor-wind hub.

• Batteries cover overnight gaps; wind covers multi-day winter lulls. A solar+battery system reaches a ceiling at ≈68% renewable — no affordable battery bridges a week-long lull — while adding wind extends the same firm system to ≈92%.
  More

  Wind's lulls don't coincide with overcast spells, which is why adding wind raises the achievable winter renewable share. At the same 65% target (2035, EU), the no-wind build delivers $138/MWh vs $99 with wind. Even with wind, a firm battery-only system tops out near ~94% renewable: the last few percent need long-duration storage or hydrogen.

• The firming choice — what covers the dark, windless weeks — changes the delivered cost more than the generation build does. A fully gas-free build firmed by self-produced hydrogen delivers $191→$142/MWh in Europe, crossing below gas ~2035.
  More

  Gas is the lower-cost emitting option; purchasing green H₂ for the same turbine adds ≈$25–30/MWh (a premium that narrows as EU carbon rises); self-produced H₂ — an electrolyser plus tank storage, charged on surplus sun — is the cheaper zero-carbon route. Where terrain allows pumped storage, it firms the same sun+wind for $31–56/MWh less than green H₂ — often the difference between a marginal site and a competitive one.

• Cheapest clean sites in Europe: water first, then windy coasts and sunny islands. At 2030, the cheapest 24/7 carbon-free sites are: Aurland ≈$46, Harsprånget ≈$46, Kaprun ≈$46, Buksefjord ≈$46, Hellisheiði ≈$63 (per-site real ERA5 weather).
  More

  Firm hydro and geothermal sites skip the firming question entirely. Among sun+wind sites, the cheapest are those whose resource co-locates with pumped-storage terrain (islands and sierras); flat sites fall back on dearer H₂ firming. A 978-cell scan of the whole continent adds: the best build-it-yourself geography is windy coast, not sunny interior — mostly-land coastal cells from the Baltic to Galicia cluster at ~$119–121/MWh, while the part-sea cells (~$95 at onshore prices) really cost ~$158+ once their sea wind is priced at offshore capex. Details in the Geography chapter.

• Small modular nuclear: a glide-path reference, not today's competitor. Modelled as an exogenous reference line, first-of-a-kind ≈$175/MWh (EU) / $150 (US) gliding to ≈$85 over 10–12 years — competitive with Europe's deep-renewable builds if that glide materialises, never with cheap US gas.
  More

  The SMR line is deliberately simple and never part of the optimisation: a linear FOAK→NOAK decline, then flat — no Wright's-Law learning, since there is no deployed fleet to learn from. An $85 NOAK undercuts the EU 90% build (≈$219/MWh in 2040) if NOAK costs materialise — uncertain, given FOAK history — but never approaches the $58 US gas baseline.

• In the US, cheap gas sets a baseline renewables do not cross within the horizon. Even 70–80% renewable targets bottom out ≈$60/MWh vs the $58 flat-gas baseline (crossover >2040); they reach parity with a stressed (+60% fuel) gas baseline by ~2030.
  More

  At $4/MMBtu, renewables compete with a ~$29/MWh fuel bill, not the $58 all-in gas LCOE — so the pure cost-optimum is ≈0% renewable today and only ~⅓ (solar-only, no battery) by 2040. A clean US datacenter reduces exposure to gas and carbon prices; in Europe it is the cheaper plant.

• AI datacenter procurement accelerates the technology learning curve. Every doubling of cumulative deployment cuts battery system cost ~19% and solar ~25% (Wright's Law), and GW-scale datacenter procurement is concentrated in the technologies with the steepest learning curves.
  More

  Battery turnkey prices fell ~31% in 2025 alone. The deployment trajectory behind these projections already leans on the AI clean-power buildout to keep additions growing — so each clean campus pulls the parity years above forward for everyone else.

• Off-grid is itself a premium. Staying on the grid with a renewable-energy contract sits below the off-grid high-renewable optimum in both regions — off-grid buys siting independence, at a cost.

Delivered cost & parity ($/MWh of load)

The renewable target is the minimum share of the datacenter's yearly energy that must come from solar + wind + battery (the rest is gas). Firm (always-on) workload; gas backup sized to 100% of load. "Crossover" = the first year the build's delivered cost drops below the gas baseline.

Cost trajectories

Dig deeper