Skip to content

Instantly share code, notes, and snippets.

@mrzmyr

mrzmyr/grid-surplus-report.md

Created June 6, 2026 09:47
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save mrzmyr/2edea94d67d4f9d5c490ced809215da3 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save mrzmyr/2edea94d67d4f9d5c490ced809215da3 to your computer and use it in GitHub Desktop.
Download ZIP
EU Electricity Surplus (DE/NL/SE/ES) → GPU Cluster + Cheap Inference — research brief
Raw
grid-surplus-report.md

EU Electricity Surplus → GPU Cluster + Cheap Inference Play

Research DE / NL / SE / ES. Question: how much power overproduction, can flexible GPU clusters eat cheap surplus, sell discounted AI inference.

Compiled June 2026. Numbers cited. Wholesale-only unless noted (grid fees/taxes extra).

Cross-country scorecard (2024 → 2025)

DE NL SE (north SE1/SE2) ES
Negative-price hrs 2024→2025 457 → 573 (record) 458 → 585 SE2 720 → 679; SE1 674 784 → 798 (70% truly neg vs 32%)
RE curtailed / wasted ~9 TWh RE (2024), solar +97% ~4.4 TWh "lost" + 0.7 op-curtail mostly economic (no clean TWh) 1.5 → ~9 TWh (~6x in 2yr)
Cheapest avg price €78→89/MWh nat'l €75→83/MWh nat'l SE2 ~16.5 €/MWh (2025), SE1 ~26 €65 nat'l; midday capture €7.5–17
Why surplus wind north + solar midday, N→S bottleneck huge rooftop solar, grid congestion hydro+wind north, SE2↔SE3 bottleneck solar boom, "electricity island" (3GW France link)
Surplus geography Schleswig-Holstein/north (wind) Flevoland/Friesland/Brabant (solar belt) Norrbotten / Node Pole (Luleå/Boden) Toledo/Ciudad Real/Granada/Extremadura
Cheap window durability shrinks ~2028 (SuedLink) grows (solar still climbing) durable to 2030s (NordSyd slow) structural to ~2028+ (Biscay link only 5GW)

Siting rank for cheap power: SE (north) > ES > NL > DE.

  • SE north = cheapest + cleanest + durable + free cooling + proven (Meta, miners, green steel).
  • ES = deepest growing glut but shallow negatives (PPA zero-floors).
  • NL = best flexibility programs (GOPACS, ATR85 non-firm connection).
  • DE = biggest market, window narrowing, grid-fee complexity.

Hard truth: your thesis breaks on 3 facts

Stated thesis = "cheap surplus power → discounted real-time inference." Fails:

  1. Power = only ~5–20% of inference TCO. GPU capex/depreciation dominates (35–50%). Halve power → total cost moves single digits. Cheap power NOT main lever.
  2. "Run only when cheap" fights capex. Negative hrs = ~6–9% of year. Expensive GPUs need ~80–95% utilization. Idle $30k GPU to chase free power loses more in depreciation than saved on energy. (Texas study: 95% uptime → ROI 502 days; low duty-cycle → economics collapse.)
  3. Real-time inference can't wait for wind. Latency-bound, demand-driven, can't tolerate 2-min preempt. Worst-fit workload. Good fit = interruptible batch: training, fine-tuning, batch/offline inference, rendering.

Market verdict: Crusoe (poster child stranded-renewables→AI) added 4.5 GW gas to firm power. Soluna splits: batchable training/crypto → remote intermittent renewables; latency inference → near users on firm power. Nobody sells real-time inference off raw surplus.

Recommended pivot

  • Product: cheap interruptible batch / training + batch inference (relaxed SLA, ~70% cheaper). NOT discounted real-time inference.
  • Power buy: day-ahead spot + as-produced solar PPA (€35–55/MWh) stacked with balancing-market revenue (aFRR ~€20/MW/h capacity, DR payments). Flexibility = 2nd revenue stream.
  • Siting: behind-the-meter co-location at curtailed generators. Sidesteps the real binding constraint: grid-connection queues (7–15 yr DE/NL, moratoriums). Price not the bottleneck — interconnect queue is.
  • Best geos: north Sweden (firm-cheap-green) → Spanish curtailment hotspots (behind-meter PPAs near €0–15/MWh) → NL with ATR85 + GOPACS flex revenue.
  • Green pitch: "soak otherwise-curtailed renewables, audited hourly carbon." Honest + defensible. Avoid "24/7 carbon-free" without hourly matching.

Per-country detail

🇩🇪 Germany

  • Curtailment 2024: ~9.3 TWh RE (offshore wind 4,562 GWh, onshore 3,384, solar 1,389 = +97% YoY). 3.5% of RE gen curtailed.
  • Congestion-mgmt cost 2024: €2.77–2.8 bn (down from €4.2 bn 2022). RE operator compensation €554m.
  • Negative hrs: 69 (2022) → 301 (2023) → 457 (2024) → 573 (2025, record). Extreme −€250.32/MWh on 11 May 2025. Floor −€500.
  • Surplus = north wind (Schleswig-Holstein ~150% of local demand) + solar midday. N→S bottleneck. SuedLink (~2028+) will shrink the window.
  • Flex pathways: §19(2) StromNEV grid-fee reductions (up to 80–90%, but rewards baseload = anti-flex), §14a EnWG. National Data Center Strategy (Mar 2026): 4x AI/HPC by 2030, recognizes DC flexibility in tariffs.
  • Achievable: 573 negative/near-zero hrs accessible anywhere via wholesale; cheap decile <900 h/yr. Grid fees dominate delivered cost.

🇳🇱 Netherlands

  • Negative hrs: 85 (2022) → ~315 (2023) → 458 (2024) → 585 (2025, 7% of hrs). May 2025 alone ~145 neg hrs.
  • Lost renewables: 16 PJ ≈ 4.4 TWh (2025), +33% YoY. Operator-curtailed 708.6 GWh. Solar farms made >28% of annual output during negative hrs.
  • Grid congestion = headline issue. 14,044 businesses (~9 GW) on offtake waiting list + ~8,000 feed-in. Economic drag up to $38bn/yr. Zuid-Holland declared full Dec 2024.
  • Per-capita solar highest in Europe (>1 kW/person, ~24 GW).
  • Flex unlock: GOPACS congestion platform (min 100 kW bid, pays the spread). ATR85/TDTR85 non-firm connection = 85% capacity + ~50% cheaper tariff, curtail ≤15% of time w/ day-ahead notice. Ideal for interruptible GPU load — gets connection where firm 24/7 load can't.
  • Site where feed-in congestion + cheap power coincide (solar belt), NOT where offtake full. Use capaciteitskaart at postcode level.

🇸🇪 Sweden (north)

  • Zone prices 2024: SE1 ~26, SE2 ~25, SE3 ~36, SE4 ~50 €/MWh. South pays ~2x north. SE2 full-year 2025 = €16.5/MWh (~1/3 of SE3).
  • Negative hrs: SE2 720 (2024) → 679 (2025, most in all Europe); SE1 674.
  • Europe's #2 net exporter ~33 TWh (2024); −27% net-export share = #1 in Europe.
  • North wind record 40.8 TWh (2024), growing. SE2↔SE3 bottleneck ~7,300 MW → NordSyd reinforces to >10,000 MW but build runs to 2033 = cheap stays cheap.
  • Strong precedent: Meta Luleå (2013), bitcoin miners (KnC/Hive/Northern Data), green steel Stegra ~20 TWh/yr, HYBRIT. BUT industry also competes for same power.
  • Free cooling: Boden −9 to +17°C, year-round, cuts PUE.
  • Flex markets: Svk FCR/aFRR/mFRR, bids as small as 1 MW (Dec 2024).
  • ⚠️ Connection queues/capacity in north a real gate. Verify per site.

🇪🇸 Spain

  • Curtailment: 1.5 TWh (2023) → ~9 TWh (2025 Jan–Nov), ~6x. July 2025 spike ~11% of RE gen. Solar curtail rate 0.4→3.9→4.7%.
  • Negative hrs: zero in 2023 → 784 (2024) → 798 (2025), now ~6.3% of hrs but 70% truly negative. May 2025 record 269 zero/neg hrs.
  • Negatives shallow (avg −€2.1/MWh) — PPA zero-floors + weak interconnection.
  • Midday solar capture price collapsed €61 (2023) → €7.5–16.8/MWh (2025). Duck-curve spread €40–70/MWh.
  • Surplus geo: solar in Toledo/Ciudad Real/Granada/Extremadura/Castilla-La Mancha (>30% local curtail); wind in Aragón.
  • Cause: solar boom + "electricity island" (3 GW France link, 2.8% ratio vs 15% EU target). Bay of Biscay link +2.2 GW ~2028. Post-Apr-2025 blackout = conservative grid operation forces more curtailment.
  • Hyperscalers validating: AWS €33.7bn Aragón, MS+Blackstone >€30bn.
  • Best move: behind-the-meter at constrained solar plant in curtailment hotspot → PPA near €0–15/MWh. ⚠️ No clear formal REE demand-response remuneration found — case rests on price arbitrage + curtailed-PPA negotiation.

Caveats

  • 2025 full-year curtailment figures not yet officially confirmed (DE/ES absolute totals disputed; rates solid).
  • Headline €/MWh = wholesale energy only. Grid fees/levies/taxes dominate delivered cost, not avoided by cheap energy unless qualify for reductions (DE §19, NL ATR85).
  • öre→EUR ~11.3 SEK/EUR, approximate.
  • Achievable flexible-buyer €/MWh depends entirely on duty-cycle flexibility — pull hourly spot data per candidate grid node before committing capital.

Research via 5 parallel agents (4 country + 1 business/market-mechanics). Sources: Bundesnetzagentur/SMARD, TenneT, Svenska Kraftnät, Red Eléctrica/OMIE, ENTSO-E, Ember, Fraunhofer, Montel, pv-magazine, IEA, plus practitioner data (Crusoe, Soluna, io.net). Full source URLs available on request.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment