Cannibalization model · references

Sources behind the model

Every external source cited across the model pages, with its headline and what it supports. A primary tag marks a first-hand source (statistics office, regulator, market operator, the researcher's own words); reported marks journalism or analysis reporting on one; paywall marks a source behind a subscription. Grouped by the part of the argument each one carries.

01The grid connection gate · Sweden

  1. Jordbruksaktuellt · 5 Nov 2025 · reported
    The Power2Earth refusal. Lantmännen, Fertiberia and Nordion shelved the Luleå fertiliser plant after Svenska kraftnät declined the 4–5 TWh it needed. The clean binary case for gate 2.
  2. Tidningen Näringslivet · 2026 · reported
    First-come-first-served and its abandonment, the 2024 maturity requirements, the ~20,000 MW cleared, and the data-centre versus industry queue split (5–6 GW of data centres, ~80% in SE3, against ~7 GW of new industrial demand for 2026–2030).
  3. Svenska kraftnät (via TT) · 2 Feb 2026 · primary
    The grant figures: 43 applications for over 9,000 MW of withdrawal, about half data centres, 3,695 MW of consumption granted across 22 positive decisions. The withdrawal queue stood near 32 GW at year-end.
  4. Energimarknadsinspektionen · 11 Jun 2026 · primary
    The regulator's finding that connection times routinely exceed the two-year statutory limit for large projects. The basis for calling gate 2 a delay rather than a denial.
  5. Svenska kraftnät · 4 May 2026 · primary
    Capacity zones (anvisningssystem) and business matching, the tools now replacing pure queue order.
  6. Ellevio · 2025 · primary
    Counter-evidence: Stockholm's shortage easing after 2.42bn SEK of network investment. Part of why the crowding-out claim is left as inference in the region where data-centre applications concentrate.

02The price channel and the demand fork

  1. Pär Holmberg, IFN · Dagens Industri · 10 Feb 2026 · paywall
    The price channel stated by a named authority: because a data centre is capital-intensive, electricity is a small share of its costs, so it can outbid industry and the build-out continues even as it lifts the price. Locates the pressure in southern Sweden, where the data centres want to sit and no surplus absorbs them, and proposes self-firming PPAs as the remedy.
  2. Jonas Kristiansen Nøland · Affärsvärlden · Jul 2026 · paywall
    The willingness-to-pay side of the fork, stated for the Nordic case. A hyperscaler's revenue per kWh of electricity, which Nøland puts at roughly 80–100 SEK/kWh and illustrates with Anthropic renting 300 MW of xAI compute for about 1.25 billion dollars a month, some 53 SEK/kWh, runs about an order of magnitude above power-intensive industry's few SEK/kWh, so the data centre can outbid Yara, Hydro or Elkem for the same power. The same channel Holmberg describes for Sweden, and the reason the hyperscalers are contracting nuclear: Meta's January 2026 agreements with Vistra, Oklo and TerraPower run to 6.6 GW. Nøland writes from a nuclear-advocacy position, so the transferable part is the revenue-per-kWh gap rather than the policy conclusion, and it is an argument about competition for future power, not a documented displacement of a named load.
  3. IEEFA · 2024 · reported
    The price channel in a market that prices capacity explicitly: the PJM auction rising roughly tenfold, with data centres attributed 63% of the 2025/26 increase.
  4. Utility Dive · 2024 · reported
    Corroborates the PJM clearing prices and the constrained-zone figures (BGE, Dominion).
  5. IEA · 2025 · primary
    The data-centre demand figures on the fork chart: ~70 TWh in 2024, ~115 TWh projected for 2030. Also the source for data centres accounting for about 10% of EU electricity demand growth to 2030, and for European connection waits of two to ten years, averaging seven to ten in the FLAP-D hubs.
  6. IEA · Apr 2025 · primary
    The growth-rate comparison: data centre consumption rising around 15% a year from 2024 to 2030, "more than four times faster than the growth of total electricity consumption from all other sectors", while still accounting for under 10% of global demand growth. The distinction between growing fastest and constituting most of the growth rests on this.
  7. Svenskt Näringsliv · Apr 2026 · reported
    The Swedish acceleration and the spread in its forecasts. SE3 applications for large data centres rising from about 1,310 MW in 2024 to 6,692 MW in 2025, citing Svenska kraftnät via Sveriges Radio. Current consumption near 4 TWh, with 2030 projections from RISE at ~5 TWh and Energimyndigheten at 4.4–5.0 TWh against 14–15 TWh derived from the Nordic operators' 7 → 25 TWh path. An industry-body report, so read as advocacy, but the underlying figures are attributed.
  8. IEA · primary
    The ~600 TWh EU technical potential for electrified industrial heat, the long arm of the demand fork.
  9. Brookings · primary
    The jobs-per-site figures (20–30 permanent staff per 100 MW) and the 4–5% county private-employment effect that the model's earlier framing understated.

03Electrification, the breakeven, and the Finnish experiment

  1. Ilkka Hannula, Abated · reported
    The €54/MWh electric-boiler breakeven against gas steam and the hours-below-breakeven framing that the price panel measures.
  2. Ilkka Hannula, Abated · reported
    The Finnish electric-boiler surge that demonstrates B2, the flexibility exit, in the field.
  3. Energiateollisuus · 2026 · primary
    The primary Finnish source: electric boilers doubling to 8% of produced district heat in 2025, reproducing the 2,580 GWh figure independently.
  4. Finnish Energy · Sep 2025 · primary
    Background on how the electric-boiler fleet interacts with the Finnish system.

04Flexibility and thermal storage

  1. Julian Geis, Michael Lindner and Tom Brown, TU Berlin · Advances in Applied Energy 23 (2026) 100284 · 9 Jun 2026 · primary
    Peer-reviewed and open access under CC BY. A sector-coupled PyPSA-DE model of Germany to 2045 under four flexibility scenarios. The source for the price-distribution result the model was missing: scarce flexibility produces both a very high count of zero-price hours and an average price 14–17 €/MWh above the base case, while abundant flexibility leaves few zero-price hours (7%) and the lowest average price. Also the daily-flexibility split for 2045 (batteries 38%, flexible EV charging 30%, resistive heaters 7%, electrolysis 6%) and the finding that electrified heat provides flexibility only where it has thermal storage.
  2. Niklas Fernqvist, Sarah Broberg, Johan Torén and Inger-Lise Svensson · Energy Policy 172 (2023) 113332 · primary
    The Swedish side of the B2 argument, peer-reviewed. The static-tax finding, quoted here as "a set price per kWh (0.36 SEK/kWh in June 2022), independent of the electricity price", which is why a deeper glut does not reach the operator who could absorb it. Also the capacity figures: 72 of the 305 largest district-heating systems hold 1,207 MW of heat pumps and 1,151 MW of electric boilers, 1,540 MW of combined electrical draw, supplying 7.6% of heat production 2017–2021; power-to-heat potential of 0.2–8.6 TWh; and national heat storage of ~150 TJ in 2016 with ~96 TJ available for load variation. Read via the open-access Chalmers copy.
  3. Power Circle · Oct 2025 · primary
    The measurement of how far gate 1 is shut. About 130 electric boilers in Swedish district heating at roughly 10 MW each, some 1,300 MW installed, peaking in 1990 at 6.3 TWh of electricity and running at 0.1–0.4 TWh today, with about 30 MW prequalified to the ancillary-service markets against roughly 300 MW of heat pumps. Also the cost decomposition: 1,264 SEK/MWh for an extra boiler megawatt-hour including spot price, energy tax and network fee, against 694 SEK/MWh with the tax removed.
  4. Swedavia · primary
    The counter-example to the price gate. In operation since summer 2009, cutting purchased energy by 19 GWh a year, 4 GWh of electricity and about 15 GWh of district heat, which Swedavia equates to 2,000 houses. Cold water is drawn in summer for terminal cooling and the warmed return banked underground for winter snow-melting on the stands and ventilation pre-heat. The English page calls it the largest of its kind in the world.
  5. EU Covenant of Mayors · Sept 2023 · primary
    The best-quantified figure for data-centre heat recovery in Stockholm: "Med 20 leverantörer återvann Öppen Fjärrvärme över 100 GWh under 2022", meeting "1,5 % av Stockholm Exergis kundbehov", at about 190,000 EUR per year per MW delivered, with data centres contributing by far the most of any supplier category.

05The carbon price and the ETS revision

  1. ICAP · Jul 2026 · primary
    The 17 July 2026 package that weakens the carbon term feeding the breakeven gate.

06The deindustrialization spiral · verification pass

  1. Clean Energy Wire · reported
    Germany's energy-intensive output down 17.8% versus 2021 and still depressed after four years, which reframes R1 as a completed transition rather than an ongoing shock.
  2. Irish Times · 7 Jul 2026 · reported
    Ireland's 23% share, 7,663 GWh in 2025, up from about 21% in 2024.
  3. Energy Connects · Dec 2025 · reported
    The moratorium lifted and replaced by a self-firming requirement, the same instrument Holmberg proposes for Sweden.

07Storage duration and battery revenues

  1. CAISO DMM · 29 May 2025 · primary
    The CAISO battery-revenue decline, the arbitrage side of B2 seen from the operator's income.
  2. Modo Energy · reported
    The CAISO $/kW-year series, 103 → 80 → 51 → ~38.
  3. Modo Energy · May–Jun 2026 · reported
    The Australian NEM series, shown at its own resolution because the published figures do not aggregate cleanly to a calendar year.
  4. Modo Energy · reported
    The seasonal NEM revenue window behind the duration point.

08Data-centre policy context

  1. Dagens Infrastruktur · 17 Jun 2026 · reported
    The ~9,000 MW of 2025 applications with data centres about half, concentrated in Mälardalen, Stockholm, Uppsala and Gävleborg, and the paused Microsoft Sandviken build.

09System cost · the nuclear–wind comparison

  1. Jonas Wahlfrid · 2026 · reported
    A line-by-line review of the Chalmers / Mistra Electrification study (Göransson, Johnsson et al., May 2025) whose headline figure, new nuclear costing society 660–1490 MEUR/yr more than a wind-dominated system, circulates in the Swedish debate. Recomputing the study's own result with three changes, nuclear capex set to what TVO actually paid for Olkiluoto 3, the nuclear-free case's hydrogen-store growth priced, and the wind collection grid priced, turns the premium into a 444–1880 MEUR/yr saving. It carries the cannibalization mechanism in the study's own numbers: nuclear pushed to load-follow and then billed for the resulting unprofitability, and wind's production-weighted price near zero with 1000–1500 zero-price hours by 2035. The recomputation's Norway import-balance figures are the author's own estimates, stated as such. Reviews the study at research.chalmers.se.

10Offshore wind · auctions and revenue risk

  1. Daniel Wetzel · Welt · 25 Jul 2026 · paywall
    The German mirror of the merchant-against-CfD question. German offshore is auctioned with uncapped negative bidding and no revenue stabilisation, so the developer carries full merchant price risk. After capital costs rose, the August 2025 auction drew no bids on either North Sea site (2.5 GW combined), and TotalEnergies and BP are seeking to return already-awarded areas, writing off advance payments in the hundreds of millions. The industry now asks the state to take the loss risk through indexed contracts for difference, the same instrument Sweden uses for new nuclear, with the state additionally bearing inflation risk. The statutory target is a tripling to 30 GW by 2030. The zero-bid result is corroborated by Clean Energy Wire.

11Method and foundations

  1. David J. C. MacKay · UIT Cambridge · 2008 · primary
    The quantitative-energy touchstone, free in full at the link. MacKay's method, adding up supply and demand in one consistent unit (kWh per person per day) and letting no number pass without a source, is the standard the figures on these pages try to meet. It predates the cannibalization problem, so it carries none of the model's argument, but it is where the habit of putting every quantity in the same unit and sourcing every unit comes from. Also readable as an HTML edition at Life Itself.
  2. Donella H. Meadows · The Sustainability Institute · 1999 · primary
    The method behind the diagrams. Meadows ranks twelve places to intervene in a system, from constants and parameters at the weak end up to the goal of the system and the paradigm it comes from at the strong end. The model's two gates are leverage points in her sense, and low ones: a tax rate and a queue rule are parameters and rules near the bottom of her list, which is one reason the model treats them as adjustable rather than fixed. Meadows co-wrote The Limits to Growth with Jørgen Randers, the same lineage the system pages draw on.

12Jobs, value added and tax revenue per MWh

  1. iRecruit · 2026 · reported
    Staffing per megawatt by size class: 2.0–3.0 at 1–5 MW, 1.75–2.5 at 5–20 MW, 1.0–1.5 at 20 MW and above, and 0.2–0.3 for ultra-hyperscale sites over 100 MW. Google's 500 MW Kansas City campus with about 200 permanent staff is the worked example, which the source itself puts at 0.4 jobs per MW — above the ultra-hyperscale band rather than inside it. This is the input to the jobs-per-MWh arithmetic: one MW at an 85% load factor delivers 7.45 GWh a year, so the band runs from roughly 5000 to 37 000 MWh of electricity per job-year.
  2. Latitude Media · 2026 · reported
    Independent corroboration of the staffing ratio from a different direction: 25 to 40 permanent operators per 100 MW at the most automated campuses, which is 0.25–0.4 jobs per MW and brackets the Kansas City figure. Also gives capital per permanent job at a named project, Jay in Maine, at about $5.5 million.
  3. Food & Water Watch · Jan 2026 · reported
    The aggregate check on the per-megawatt arithmetic: as few as 23 000 permanent data-centre jobs in the United States in 2024, 0.01% of employment, against the 4.4% of national electricity data centres used in 2023. Also the definitional dispute with the industry's 603 000 figure, which comes from a PricewaterhouseCoopers report for the Data Center Coalition and counts the whole of NAICS 518210 — web hosting, cyber security and data labelling included — using posts rather than full-time equivalents and adding an indirect multiplier. The part that does not require trusting either total: Virginia holds about a third of national capacity and 4% of national 518210 jobs. Food & Water Watch campaigns for a moratorium on data-centre construction and the PwC report was commissioned by an industry body; neither is disinterested, which is why the geography rather than either headline is what the argument rests on.
  4. Brookings · 2026 · primary
    The counter-evidence, and it is good. About 770 American facilities linked to county employment data for 2003–2024, with synthetic controls applied to the 93 counties that received their first large data centre between 2008 and 2024 against roughly 3000 controls: total private employment 4–5% higher within five to six years, construction up 11%, information-sector work up 22%, wages up 3–4%. The authors' own cautions carry as much weight as the result — naive before-and-after comparisons overstate the effect about threefold, because data-centre counties were already growing faster than their neighbours, and the gains concentrate in hyperscale clusters, with colocation sites showing no significant information-sector growth and counties with four or more facilities showing 23%. This sits in tension with the per-megawatt arithmetic and the tension is not resolved: operating employment per unit of electricity is a narrower quantity than the total economic effect of building the facility.
  5. SVT Västernorrland · reported
    The aluminium comparator: Kubikenborg Aluminium at Sundsvall employs about 470 people and produces about 135 000 tonnes of primary aluminium a year.
  6. Kollega · reported
    The electricity side of the same comparator. Kubal's own statement that one öre per kWh costs it 16 million kronor a year implies 1.6 TWh, which its separate claim to use about 1% of Swedish electricity corroborates. That gives roughly 3400 MWh per job-year against a data centre's 5000 to 37 000. A third check — 135 000 tonnes at the 13–15 MWh per tonne typical of a modern smelter — gives 1.8–2.0 TWh, so the derived figure may be low and the ratio generous to the smelter.
  7. Skogsindustrierna · reported
    18 TWh of electricity a year, just over 14% of Swedish consumption. Note the boundary before using it: the figure covers svensk skogsindustri as a whole, sawmills included, not massa- och pappersindustrin alone.
  8. Pappers · reported
    About 19 000 employed in massa- och pappersindustrin, of whom about 11 500 are workers. Kept next to the entry above as a warning rather than a calculation: 18 TWh covers the whole forest industry while these 19 000 cover pulp and paper alone, so dividing one by the other mixes two boundaries and is not done here.
  9. Öresundsinstituttet · reported
    Facebook's Luleå halls gave close to 400 jobs once in operation, 250 of them to people already living in the region, and each new establishment can want as much as a terawatt-hour a year. The two numbers together give about 2500 MWh per job-year, which is better than the smelter — but 400 jobs against Luleå's roughly 120 MW is 3.3 jobs per MW, several times any published staffing ratio, so the count evidently reaches beyond permanent facility staff. Recorded here with that caveat rather than used in the arithmetic.
  10. Industriarbetsgivarna / Industriekonomerna · 2023 · primary
    The multiplier that stops the jobs-per-MWh comparison being unfair to computation alone: every direct industrial job in Sweden carries about 1.1 further jobs among domestic suppliers, an employment multiplier of 2.1, with industry employing around 800 000 direct and indirect. Indirect employment is not a data-centre-specific effect, so it belongs on both sides of the comparison.
  11. Skatteverket · primary
    The energy-tax reduction for datorhallar, in force since 2017, was abolished with effect from 1 July 2023, decided by the Riksdag in December 2022 on the government's proposal in prop. 2022/23:1, the stated aim being to give data centres an incentive to use electricity more efficiently. The Norrlandsavdraget for electricity consumed in certain northern municipalities is separate and still applies. This closed a subsidy; it did not change which level of government is paid. Municipalities are financed by kommunalskatt on residents' wages, while corporation tax and energy tax accrue to the state, so a load can carry large value added and still leave its municipality little — which is the fiscal half of the jobs-per-MWh argument.