“Can Europe's electricity grid keep up with the growth of AI data centres by 2030?”
Nine answers to one question, scored on the principles a strategy team relies on: what you can open, what was tested, what changes when the world does, and what can be held to account.
The eight chat answers were collected on 16 Oct 2026 with one prompt and are kept as delivered. Every rule is published next to its result and applied the same way to all nine answers. Every figure is read from files on record.
What a chat answer was never built to do
A chat tool writes a strong answer and stops. A strategy team needs the answer checked, argued, kept current and held to account. Of the 7 things that takes, the chat products' own documentation describes none in full.
CapabilityDSGHT.aiChatGPTincl. Deep ResearchClaudeincl. ResearchGeminiincl. Deep ResearchMistral Vibeformerly Le Chat
Every figure checked against the evidence it cites before you see it
Checked by code, not by the model that wrote it: a sentence whose figure is not in its evidence is cut.
Yes
ChatGPTNo
ClaudeNo
GeminiPartlyA search check the user can run after the answer; nothing is dropped
Mistral VibeNo
Goes to the institutions that own the numbers
Regulators, grid operators and statistical offices are queried for the question's key quantities.
Yes
ChatGPTPartlyDeep research can be limited to sites the user picks
ClaudeNo
GeminiNo
Mistral VibeNo
Disagreements kept as two named sides and put to an adversarial panel
Separate model judges try to refute each dispute before it is shown.
Yes
ChatGPTNo
ClaudeNo
GeminiNo
Mistral VibeNo
One analysis, revised over time, with a dated change record
Not a new answer each time you ask: the same space, updated, with what changed and why.
Yes
ChatGPTPartlyMonitoring tasks re-check and notify on change; each run is a new answer
ClaudePartlyDocs keep a version history; Claude edits when asked
GeminiNo
Mistral VibeNo
Scenario probabilities with a dated history of every move
Each move carries the evidence that caused it.
Yes
ChatGPTNo
ClaudeNo
GeminiNo
Mistral VibeNo
Forecasts registered before the outcome, scored after
Dated checks with a threshold, written down before anyone knows the answer.
Yes
ChatGPTNo
ClaudeNo
GeminiNo
Mistral VibeNo
Re-runnable from a fixed evidence base with a stated cutoff
The same question returns the same auditable record.
Yes
ChatGPTPartlyCan be limited to chosen files or sites; no cutoff date
ClaudePartlyProjects keep a fixed file set; no cutoff date
GeminiPartlyCan run on selected files only; no cutoff date
Mistral VibePartlyLibraries keep a fixed file set; no cutoff date
Cites its sources inline
All five do this.
Yes
ChatGPTYes
ClaudeYes
GeminiYes
Mistral VibeYes
As of 23 Oct 2026. Read from each vendor's own product documentation and help centre. "No" means none of that documentation describes the capability.
Scored on what a strategy team needs, not on prose
One question, nine answers, scored on 7 published principles and a blind judged block. Each rule is written next to its row and applied the same way to every answer. DSGHT.ai is ahead of all eight on 7 of 7 measured principles.
Highlight one answer
PrincipleDSGHT.ai
ChatGPTChatChatGPTDeep ResearchClaudeChatClaudeResearchGeminiChatGeminiDeep ResearchMistralLe ChatMistralDeep Research
Evidence breadth
How many sources behind the answer can you open?
DSGHT.ai: 1,962sources
Best of the eight: 55 (Claude Research)
ChatGPT: ChatGPT17
ChatGPT Deep Research: ChatGPT Deep Research0Names its sources, no links
Claude: Claude38
Claude Research: Claude Research55
Gemini: Gemini13
Gemini Deep Research: Gemini Deep Research45
Mistral Le Chat: Mistral Le Chat21
Mistral Deep Research: Mistral Deep Research44
Rule. Sources behind the answer that a reader can open. DSGHT.ai: distinct sources in the space’s append-only source ledger. Each chat answer: distinct links cited in the answer as delivered.
DSGHT.ai. 4,760 claims on file, 3,272 verified against their page. The briefing that was judged cites 34 links directly.
How many cited sources are the institutions themselves?
DSGHT.ai: 15primary sources cited
Best of the eight: 11 (Claude Research)
ChatGPT: ChatGPT6
ChatGPT Deep Research: ChatGPT Deep Research0No links to classify
Claude: Claude3
Claude Research: Claude Research11
Gemini: Gemini6
Gemini Deep Research: Gemini Deep Research6
Mistral Le Chat: Mistral Le Chat3
Mistral Deep Research: Mistral Deep Research4
Rule. Links cited in the answer whose publisher is a regulator, a grid or market operator, a statistics office, a central bank or a peer-reviewed journal, by one declared publisher table applied to every answer.
DSGHT.ai. Counted on the 34 links in the judged briefing, the same unit as every chat answer.
DSGHT.ai: 95%of figures sourced38 of 40 sentences with a figure
Best of the eight: 94.7% (Gemini Deep Research)
ChatGPT: ChatGPT42.1%8 of 19
ChatGPT Deep Research: ChatGPT Deep Research0%0 of 16
Claude: Claude26.5%9 of 34
Claude Research: Claude Research56.4%31 of 55
Gemini: Gemini22.7%5 of 22
Gemini Deep Research: Gemini Deep Research94.7%18 of 19
Mistral Le Chat: Mistral Le Chat45.5%10 of 22
Mistral Deep Research: Mistral Deep Research91.7%33 of 36
Rule. Share of the sentences stating a figure that carry a citation in the same sentence or table row. Counted part: everything before the Scenarios section, which is the executive summary, the key drivers and the tensions, where the shared prompt asked for evidence with a source. Scenarios, signposts and recommendations are the answer’s own forecast (probabilities, axes, thresholds, check-by dates, proposed moves) and are not counted, nor is the source list. A figure is a number written in digits, other than a year, a date, a list, section or row number, a label such as “Force 1” or “EU-27”, or a scenario’s probability. Titles and questions count as sentences. A citation is an [S#] or [n] tag that resolves to the answer’s source list, an inline link or URL, the tool’s own footnote number or source chip, or a “(Publisher, year)” reference whose publisher is in the source list. One script applies this to all nine answers.
DSGHT.ai. 38 of 40 sentences with a figure carry a source, in the briefing as shipped on 2026-10-23: the judged text, word for word.
Without a source in the judged briefing.
**Why it matters:** If the 3-4 year timeline holds, it favours the grid axis' low pole (A, C).
If 7-10 years persists, it points to its high pole (B, D).
Does each cited figure open at the sentence that supports it?
DSGHT.ai: 100%of cited figures open at their passage38 of 38 cited figures3,272 claims in the space open at their passage
Best of the eight: 0% (ChatGPT, Claude, Claude Research, Gemini, Gemini Deep Research, Mistral Le Chat, Mistral Deep Research)
0%for the other seven: links to a page, not a passage; ChatGPT: names its sources, no passage; ChatGPT Deep Research: no cited figures; Gemini: names its sources, no passage.
Rule. Share of the answer’s cited figures whose citation opens at the exact supporting passage. A cited figure is a sentence that states a figure (a number in digits, other than a year, a date, a list, section or row number, a label, or a scenario’s probability) and carries a citation in the same sentence or table row, counted in everything before the Scenarios section; one script applies this to all nine answers. DSGHT.ai: a figure counts when its [S#] tag names a page on which the space holds a claim whose passage was matched character for character on the stored page, with every number of the claim inside it, and the figure is in that claim or its passage. Chat answers: a citation counts only when its link carries the passage itself (a text-fragment link). Their links point at whole pages, so none does.
DSGHT.ai. 38 of 38 cited figures in the judged briefing. Across the space, 3,272 of 4,760 claims open at their passage, on 227 source pages.
Were the disagreements put to a test before you saw them?
DSGHT.ai: 20tensions adversarially tested
Best of the eight: 0 (all eight)
0for all eight answers: lists tensions in the text, untested.
Rule. Disagreements between named sources that were put to an adversarial panel after they were found: three refuters with different lenses, a tension killed when two of them refute it.
DSGHT.ai. 71 tensions in the space; the latest round tested 20 and killed 1.
Does the answer change when the world does, and say why?
DSGHT.ai: 17dated versionssince 12 Jul 2026
Best of the eight: 1 (all eight)
1for all eight answers: one snapshot.
Rule. Dated versions of the scenario probabilities kept on file, each later version stating why each probability moved. A single answer is one version.
DSGHT.ai. 17 versions since the first run; 14 state why each probability moved.
DSGHT.ai: Preferredby an independent blind judge on living, tension rigour and decision usefulness, against the answers it scored.
None came out ahead of DSGHT.ai. They read well, but stop at the day they were written: no dated updates, no tested tensions, nothing registered to be scored later.
Chat answers collected on 16 Oct 2026 with one prompt. The DSGHT.ai briefing that was judged shows only evidence from documents published on or before 16 Oct 2026; space counts are read from the space as it stands. The method and the full record are below.
Read blind, DSGHT.ai comes out ahead
With every name removed, an independent judge from another model family read each answer next to ours, twice, in both orders. No other answer came out ahead of DSGHT.ai. The reasons it gave are the ones a board would give, shown here with the space's own examples. The chat answers read well, but they stop at the day they were written.
Instrument check: the judge preferred the intact answer over a copy with a planted defect in 28 of 28 cases, and on those control pairs and on our answer set against an identical copy of itself, its verdict changed when the order was swapped in 2 of 45 checks. Our limit for both is 1 in 5. Passed. On the real comparisons the order changed the verdict in 3 of 24; each of those counts as a tie, not a win.
Two institutions, one figure. Data-centre electricity demand in 2030, kept apart rather than averaged.17 dated versions, then a dated check. Each move states its reason; the next signpost is registered to be scored.
Average time from inception to operation of European electricity interconnectors, in ACER's next electricity infrastructure monitoring report: at most 7 years
One answer vs. a living map
Asked on the same day, four chat tools opened with four differently worded verdicts, from “yes” in aggregate to a flat “No.” Each is a well-written document, fixed at the moment it was generated. DSGHT.ai keeps the question open: the evidence, where it disagrees, the futures it implies and the dates on which they can be checked.
One answer
The same prompt, asked of four chat tools.
How the answer opens
“Answer: in aggregate energy terms, yes. On location, timing and the EU’s own ambition, no.”
Claude Sonnet 5.5 (effort Medium), chat, web search. Asked 16 Oct 2026.
1 document3,238 words38 links
It stays as written on 16 Oct 2026. To update it, ask again.
A living map
The DSGHT.ai future space on the same question. It unfolds in four layers, and each one stays open to inspection.
1
Map of evidence
4,760 claims extracted from 1,962 sources. 3,272 of them (69%) are verified against the exact passage on 227 source pages.
Every dot is one distinct source in the space’s ledger: 1,962, gathered over 11 ingest runs and never deleted. Below them, at the same scale, the links each chat answer shows its reader.
DSGHT.ai future space, as of 16 Oct 20261,962
Links in each chat answer, same scale
Claude38 links
ChatGPT17 links
Gemini13 links
Mistral Le Chat21 links
Publisher class. Tap one to find it in every field.
Classes come from the declared publisher table the platform uses (source-quality-v2-2026-10-18), applied the same way to both sides. 1,166 sources fall outside the table and are counted, not hidden; 187 of them are search-result links whose publisher was not resolved. A link count is not a reading count: a chat tool may read more pages than it cites. It is what the reader is given to check.
Breadth is not proof. What makes a source useful is a sentence you can check: 3,272 claims carry a passage verified on 227 of these pages. Open one.
Every claim opens to the sentence behind it
A link points at a document; finding the sentence that backs a number is left to the reader. In the space, 3,272 claims carry the exact passage from their source page, located by character position and checked before they are shown. Tap any of them.
Verified claims from the space (8 of 3,272)
How the chat answers cite the same kind of fact
Claude
“The Cloud and AI Development Act aims to at least triple EU data-centre capacity in 5–7 years and speed up permitting.”[S5]
Source tags such as [S5] point to a numbered list of links at the end of the answer.
ChatGPT
“The International Energy Agency (IEA) projects global data-centre consumption of around 945 TWh in 2030 in its 2025 base case.”
Citations appeared as source chips in the chat window. Its links were recovered from a follow-up in the same chat, so individual sentences are not linked.
Gemini
“Global data centre electricity consumption is projected to double from 460 TWh in 2024 to over 1,000 TWh by 2030 in the base case”(IEA, 2024)
Sentences cite an author and year; the links sit in a list at the end of the answer.
Mistral Le Chat
“IEA projects global data-centre electricity to roughly double from ~415 TWh (2024) to ~945 TWh by 2030 (~3% of global demand), with AI-optimised server consumption growing ~30%/year.”[IEA, Energy and AI]
Sentences link to the report page they draw on.
Where the sources disagree, you see both sides
Chat answers can name tensions too. Here, each side is a claim verified against its own page, the two are kept apart instead of averaged, and a deciding signpost says which side the world is taking, by a date. Of the 20 tensions put to the adversarial panel in the latest round, it killed 1.
EUR-Lex (EU law)
98,5 TWh
The energy consumption of EU data centres is expected to rise to 98,5 TWh by 2030, a 28% increase.
1.7×between the two figures
Ember
168 TWh
European data centre power demand is expected to grow to 168 TWh in 2030 and 236 TWh in 2035, an increase of almost 150% in ten years.
1.7× between the two figures
Why it matters
A board planning grid or power capacity for European data centres should test both a ~100 TWh and a ~170 TWh 2030 scenario rather than a midpoint.
Deciding signpost, check by 2027-12
Watch: Measured European data centre electricity consumption for 2026 (EU reporting scheme / IEA Energy and AI update)
Threshold: Below 85 TWh favours the Commission trajectory; above 105 TWh favours the Ember trajectory
Source: European Commission / IEA
Adversarial panel
3 refuters with different lenses (logic, evidence, materiality) tried to kill it on 19 Oct 2026. 0 of 3 succeeded.
Stands
Every move dated, every move explained
The four scenario probabilities, every time the space was updated since 12 Jul 2026. The moves are small and are drawn small, on an axis that starts at zero. Each one carries the reason the update gave for it.
A Electrified Compute Continent
B Queue Europe
C Empty Halls, Full Grid
D Stranded Ambition
Signpost triggered
19 Oct 2026, 18:38 UTCUpdate 17 of 17
12 Jul19 Oct
What changed in this update, and why · 0 new claims read
19%1%
Electrified Compute Continent
30%
Queue Europe
21%
Empty Halls, Full Grid
30%1%
Stranded Ambition
A Electrified Compute Continent19% (−1 pts)
“Adjusted by -1pp based on ambient claim shifts (no direct signpost trigger)”
B Queue Europe30% (±0 pts)
“Held at prior — no material evidence for this scenario”
C Empty Halls, Full Grid21% (±0 pts)
“Triggered signpost(s): AI compute energy efficiency improvement (FLOPS per watt for leading AI training/inference chips, year-on-year)”
Signpost: AI compute energy efficiency improvement (FLOPS per watt for leading AI training/inference chips, year-on-year)
D Stranded Ambition30% (+1 pts)
“Triggered signpost(s): A major hyperscaler publicly cancels or indefinitely defers a previously announced European data-centre campus citing grid and regulatory uncertainty”
Signpost: A major hyperscaler publicly cancels or indefinitely defers a previously announced European data-centre campus citing grid and regulatory uncertainty
Source: the space’s probability history, 17 recorded updates. Reasons are quoted as the update wrote them.
Dated checks, written down before the answer is known
Every scenario names what would confirm it, with an indicator, a threshold, a source and a date. Each of the 52 is registered as a forecast in an append-only file: if a later run moves the threshold or the date, that becomes a new forecast and the old one still resolves against what it said.
202620272028202920302031261245155611121
By end of Q4 2026 · 2
EU-wide metered data-centre annual electricity consumption figure reported under Energy Efficiency Directive database
> 85 TWh
Exceeds 85 TWh by Q4 2026. Source: European Commission (DG ENER / EED database). Scenario A, Gridlock & Geographic Flight. Resolves after 16 Mar 2027.
European industrial electricity price premium relative to US/Middle East
> 100%
EU industrial electricity prices remain >100% higher than US average through Q4 2026. Source: Eurostat / International Energy Agency. Scenario D, The Energy Rationing & Stagnation Trap. Resolves after 16 Mar 2027.
By end of Q1 2027 · 6
Annual European data-centre renewable PPA volume (TWh contracted per year)
> 25 TWh
Exceeds 25 TWh/year contracted by data-centre operators in a single calendar year. Source: BloombergNEF Corporate Clean Energy Procurement tracker. Scenario A, Green Acceleration Pact. Resolves after 14 Jun 2027.
Annual hyperscaler European data-centre CapEx as a percentage of total global hyperscaler CapEx (Europe's share of announced investment)
< 12%
Europe's share of hyperscaler global CapEx falls below 12% in any calendar year — signalling a confirmed redirection of investment away from Europe. Source: IDC European Data Centre Tracker; individual hyperscaler 10-K and earnings call disclosures (SEC EDGAR for US-listed firms). Scenario D, Fragmented Sovereignty Trap. Resolves after 14 Jun 2027.
EU Grids Package dedicated grid connection guidance adoption status
≥ 15
Adopted into national law by at least 15 Member States by Q1 2027. Source: EUR-Lex / European Commission. Scenario B, The Power-Scale Supercluster. Resolves after 14 Jun 2027.
New grid-connection offers to data centres with energisation before 2030 issued by EirGrid and CRU
≥ 1
at least 1 offer with pre-2030 energisation date. Source: EirGrid and the Commission for Regulation of Utilities (CRU). Scenario B, Queue Europe. Resolves after 14 Jun 2027.
Data-centre share of Dutch national electricity demand for 2025 (Statistics Netherlands)
≤ 7.5%
at most 7.5%. Source: Statistics Netherlands (CBS). Scenario C, Empty Halls, Full Grid. Resolves after 14 Jun 2027.
Data-centre share of Irish national electricity consumption (CSO metered release)
≤ 19%
at most 19%. Source: Central Statistics Office Ireland (CSO). Scenario D, Stranded Ambition. Resolves after 14 Jun 2027.
By end of 2027-03 · 1
European data-centre electricity consumption for 2025, in TWh (Ember estimate)
≥ 105 TWh
at least 105 TWh. Source: Ember. Scenario A, Electrified Compute Continent. Resolves after 15 May 2027.
By end of Q2 2027 · 2
Ireland national data centre metered electricity demand share
> 23%
Exceeds 23% of total national electricity demand by Q2 2027. Source: Central Statistics Office (CSO Ireland). Scenario A, Gridlock & Geographic Flight. Resolves after 13 Sep 2027.
First aggregated EU data centre energy reporting results under EED
< 80 TWh
Total reported EU data centre consumption below 80 TWh by Q2 2027. Source: European Commission (EED Reporting Database). Scenario C, Sovereign Efficiency & Distributed Nodes. Resolves after 13 Sep 2027.
By end of 2027-06 · 4
EU data centre annual electricity consumption
< 75 TWh
below 75 TWh. Source: European Commission. Scenario C, Sovereign Efficiency & Distributed Nodes. Resolves after 14 Aug 2027.
Share of cross-border capacity needs in Europe that are not being addressed by projects (ACER monitoring)
< 50%
below 50%. Source: ACER (Agency for the Cooperation of Energy Regulators) electricity infrastructure monitoring report. Scenario A, Gridlock & Geographic Flight. Resolves after 14 Aug 2027.
EU-27 data-centre electricity consumption for 2025 in the EED European Database on Data Centres, in TWh
≥ 85 TWh
at least 85 TWh. Source: European Commission (Energy Efficiency Directive data-centre reporting). Scenario C, Empty Halls, Full Grid. Resolves after 14 Aug 2027.
Share of cross-border capacity needs in Europe that are not being addressed by projects, per ACER monitoring report
> 50%
at or above 50%. Source: ACER (Agency for the Cooperation of Energy Regulators) electricity infrastructure monitoring report. Scenario B, Queue Europe. Resolves after 14 Aug 2027.
By end of 2027-09 · 5
Average grid-connection time for large loads in Europe's established data-centre hubs
≤ 5
at most 5 years. Source: ACER / national TSO connection statistics. Scenario A, Gridlock & Geographic Flight. Resolves after 14 Nov 2027.
Average time from inception to operation of European electricity interconnectors, in ACER's next electricity infrastructure monitoring report
≤ 7
at most 7 years. Source: ACER electricity infrastructure monitoring report. Scenario A, Gridlock & Geographic Flight. Resolves after 14 Nov 2027.
Installed data-centre IT capacity in the EU, all segments
≥ 13 GW
at least 13 GW. Source: European Commission / Council of the EU data-centre capacity figures. Scenario A, Electrified Compute Continent. Resolves after 14 Nov 2027.
Data-centre grid-connection requests and access rights in France
≥ 18 GW
at least 18 GW. Source: National TSO / regulator connection statistics. Scenario A, Electrified Compute Continent. Resolves after 14 Nov 2027.
Average time from inception to operation of European electricity interconnectors, in ACER's next electricity infrastructure monitoring report
≤ 7
at most 7 years. Source: ACER electricity infrastructure monitoring report. Scenario B, Queue Europe. Resolves after 14 Nov 2027.
By end of 2027 · 15
Average EU grid connection approval time for data centres (months from application to energisation)
< 36
Falls below 36 months on average across DE, NL, IE corridors. Source: ENTSO-E Ten-Year Network Development Plan (TYNDP) monitoring reports. Scenario A, Green Acceleration Pact. Resolves after 29 Apr 2028.
EU cross-border electricity interconnection capacity additions (GW commissioned per year)
≥ 5 GW
≥5 GW new cross-border capacity commissioned in a single year across EU member states. Source: ACER Annual Report on the Results of Monitoring the Internal Electricity Market. Scenario A, Green Acceleration Pact. Resolves after 29 Apr 2028.
Average grid connection queue wait time for >50MW data-centre applications in Germany, Netherlands, and Ireland (months)
> 84
Exceeds 84 months (7 years) average across the three markets simultaneously. Source: ENTSO-E National Reports and Bundesnetzagentur (German Federal Network Agency) connection queue data. Scenario B, Managed Constraint. Resolves after 29 Apr 2028.
Annual new data-centre capacity approved (MW) in Ireland, Netherlands, and Germany combined
< 500 MW
Falls below 500 MW/year approved across all three markets combined, indicating effective moratorium conditions. Source: EirGrid (Ireland), TenneT (Netherlands/Germany) published connection agreement statistics. Scenario B, Managed Constraint. Resolves after 29 Apr 2028.
AI compute energy efficiency improvement (FLOPS per watt for leading AI training/inference chips, year-on-year)
> 40%
Year-on-year improvement in FLOPS/watt for top-tier AI chips exceeds 40% in two consecutive years, indicating an efficiency breakthrough cycle. Source: MLPerf Benchmarks (MLCommons industry consortium) and IEA Global Energy and AI report. Scenario C, Efficiency Escape. Resolves after 29 Apr 2028.
Flagged as triggered by scenario updates since 24 Aug 2026; the forecast still resolves on its date.
Number of EU member states with active moratoriums or binding caps on new data-centre electricity connections above 100 MW
≥ 3
At least 3 EU member states simultaneously enforcing moratoriums or binding caps on new data-centre connections above 100 MW. Source: European Commission DG Energy state-aid and permitting database; ENTSO-E national grid operator annual reports. Scenario D, Fragmented Sovereignty Trap. Resolves after 29 Apr 2028.
EU cross-border electricity price spread (average price differential between highest and lowest electricity price zones in EU, €/MWh)
> 80
Average EU intra-day price spread exceeds €80/MWh sustained for more than 30 days annually — indicating fragmented national markets failing to arbitrage through interconnection. Source: ACER Market Monitoring Report (annual); European Power Exchange (EPEX SPOT) real-time data. Scenario D, Fragmented Sovereignty Trap. Resolves after 29 Apr 2028.
A major hyperscaler publicly cancels or indefinitely defers a previously announced European data-centre campus citing grid and regulatory uncertainty
≥ 1
At least 1 announced campus cancellation or indefinite pause by a tier-1 hyperscaler citing grid/energy constraints. Source: Major tech media / press releases (TechRadar, Datacentres.com, Tom's Hardware). Scenario D, Fragmented Sovereignty Trap. Resolves after 29 Apr 2028.
Flagged as triggered by scenario updates since 24 Aug 2026; the forecast still resolves on its date.
European data-centre electricity consumption for 2025, in TWh (Ember estimate)
≥ 105 TWh
at least 105 TWh. Source: Ember. Scenario A, Electrified Compute Continent. Resolves after 29 Apr 2028.
Installed data-centre IT capacity in the EU, all segments
≥ 13 GW
at least 13 GW. Source: European Commission / Council of the EU data-centre capacity figures. Scenario A, Electrified Compute Continent. Resolves after 29 Apr 2028.
Data-centre grid-connection requests and access rights in France
≥ 18 GW
at least 18 GW. Source: National TSO / regulator connection statistics. Scenario A, Electrified Compute Continent. Resolves after 29 Apr 2028.
EU-27 data-centre electricity consumption for 2025 in the EED European Database on Data Centres, in TWh
≥ 85 TWh
at least 85 TWh. Source: European Commission (Energy Efficiency Directive data-centre reporting). Scenario C, Empty Halls, Full Grid. Resolves after 29 Apr 2028.
IEA estimate of EU data-centre electricity use in 2026 (TWh)
≥ 85 TWh
at least 85 TWh. Source: International Energy Agency (Energy and AI). Scenario B, Queue Europe. Resolves after 29 Apr 2028.
Share of cross-border capacity needs in Europe that are not being addressed by projects, per ACER monitoring report
> 50%
at or above 50%. Source: ACER (Agency for the Cooperation of Energy Regulators) electricity infrastructure monitoring report. Scenario B, Queue Europe. Resolves after 29 Apr 2028.
Average time from inception to operation of European electricity interconnectors, in ACER's next electricity infrastructure monitoring report
≤ 7
at most 7 years. Source: ACER electricity infrastructure monitoring report. Scenario B, Queue Europe. Resolves after 29 Apr 2028.
By end of Q4 2027 · 5
Global data center total power capacity demand
> 150 GW
Exceeds 150 GW by Q4 2027. Source: International Energy Agency (IEA). Scenario B, The Power-Scale Supercluster. Resolves after 15 Mar 2028.
EU Digital Decade Policy Programme edge node deployment count
> 4000
Exceeds 4,000 deployed edge nodes by Q4 2027. Source: European Commission (Digital Decade Report). Scenario C, Sovereign Efficiency & Distributed Nodes. Resolves after 15 Mar 2028.
New data centre grid connection offers energised in Ireland post-moratorium
< 3
< 3 connections. Source: Commission for Regulation of Utilities (CRU Ireland). Scenario D, The Energy Rationing & Stagnation Trap. Resolves after 15 Mar 2028.
IEA Base Case growth in European data-centre electricity use, 2024 to 2030
≥ 70%
at least 70%. Source: International Energy Agency (Energy and AI update). Scenario B, Queue Europe. Resolves after 15 Mar 2028.
at least 1 Amsterdam-region connection offer with pre-2030 energisation date. Source: TenneT and Netherlands Authority for Consumers and Markets (ACM). Scenario C, Empty Halls, Full Grid. Resolves after 15 Mar 2028.
By end of 2027-12 · 6
European data centre annual electricity consumption
> 120 TWh
exceeds 120 TWh. Source: International Energy Agency. Scenario A, Gridlock & Geographic Flight. Resolves after 14 Feb 2028.
Installed colocation data centre IT capacity in Europe
> 10 GW
exceeds 10 GW. Source: Data Centre Dynamics. Scenario B, The Power-Scale Supercluster. Resolves after 14 Feb 2028.
Share of European data centre operators identifying electricity access as their biggest challenge
> 85%
exceeds 85%. Source: European Data Centre Association. Scenario B, The Power-Scale Supercluster. Resolves after 14 Feb 2028.
Average connection capacity usage rate of data centres connected to the French public transmission system
> 30%
exceeds 30%. Source: RTE France. Scenario C, Sovereign Efficiency & Distributed Nodes. Resolves after 14 Feb 2028.
Share of Frankfurt total electricity consumption consumed by data centers
> 35%
exceeds 35%. Source: City of Frankfurt Climate Department. Scenario D, The Energy Rationing & Stagnation Trap. Resolves after 14 Feb 2028.
IEA estimate of EU data-centre electricity use in 2026 (TWh)
≥ 85 TWh
at least 85 TWh. Source: International Energy Agency (Energy and AI). Scenario B, Queue Europe. Resolves after 14 Feb 2028.
By end of Q2 2028 · 1
EU Grids Package adopted with binding grid-connection timelines for large loads
≥ 1
at least 1 adopted regulation with binding large-load connection timelines. Source: European Parliament Legislative Observatory / Council of the EU. Scenario A, Electrified Compute Continent. Resolves after 13 Sep 2028.
By end of 2028 · 1
Average Power Usage Effectiveness (PUE) across EU data centres (lower is more efficient; 1.0 is theoretical perfect)
< 1.25
EU average PUE falls below 1.25 by end-2028, indicating aggressive efficiency adoption beyond minimum compliance. Source: EU Commission Joint Research Centre Data Centre Energy Efficiency reporting (under EU Energy Efficiency Directive Article 12). Scenario C, Efficiency Escape. Resolves after 30 Apr 2029.
By end of Q4 2028 · 1
Cloud and AI Development Act in force with binding data-centre permitting fast-track
≥ 1
at least 1 act in force with binding data-centre permitting fast-track. Source: Official Journal of the European Union (EUR-Lex). Scenario D, Stranded Ambition. Resolves after 16 Mar 2029.
By end of 2030 · 2
European data-centre electricity consumption as percentage of total EU electricity demand
> 9%
Reaches or exceeds 9% of total EU electricity consumption (the EC's 2030 projection). Source: Eurostat Energy Statistics (nrg_cb_e dataset) cross-referenced with EU Commission JRC data-centre monitoring. Scenario B, Managed Constraint. Resolves after 30 Apr 2031.
Average grid-connection time for new data centres in established European hubs (FLAP-D and similar markets), in years
> 4
at or above 4 years. Source: National TSO connection-queue statistics (EirGrid, TenneT, RTE, NESO). Scenario B, Queue Europe. Resolves after 30 Apr 2031.
By end of 2030-12 · 1
Average grid-connection time for new data centres in established European hubs (FLAP-D and similar markets), in years
> 4
at or above 4 years. Source: National TSO connection-queue statistics (EirGrid, TenneT, RTE, NESO). Scenario B, Queue Europe. Resolves after 14 Feb 2031.
27 of 52 forecasts carry a probability stated in advance and will be Brier-scored when they resolve.
Method
Question
“Can Europe's electricity grid keep up with the growth of AI data centres by 2030?”
The eight answers
ChatGPT, ChatGPT Deep Research, Claude, Claude Research, Gemini, Gemini Deep Research, Mistral Le Chat, Mistral Deep Research, asked on 16 Oct 2026 with one prompt and kept as delivered.
Our answer
The board briefing the space produces, using only documents published on or before 16 Oct 2026. Rules and judges were fixed before judging.
Ask your own question of a living map
Bring the strategic question your team keeps re-asking. We will show you the space it becomes.
