A cross-jurisdictional analysis of hyperscaler-siting decisions and concentration trajectories at per-substation granularity, across six EU jurisdictions, on the analytical surface where four facets converge: grid-resilience, compute-infrastructure, civic-equity, and climate-adaptation. Develops the EU AI Act 2025–2027 implementation calendar (Chapter V GPAI live 2 Aug 2025 · Art. 6 high-risk live 2 Aug 2026 · Art. 50 transparency live 2 Aug 2026 · full compliance 2 Aug 2027), the CRA Art. 14 reporting cadence live since 11 September 2026, and the per-jurisdiction hyperscaler concentration trajectories through 2029.
This is a Themed Analysis in the SSI Index systemic-layer publishing spine. The base surface is the Substation Systemic Index (SSI Index v4.2), a substation-level composite covering 796,121 substations across 39 OECD-adjacent jurisdictions, at TRL 9, updated on a monthly cron.2 The Systemic Layer sits above the base index and asks how the modifier surface behaves under compound stress — cascading, correlated, or amplified failures that a per-substation reading in isolation cannot see.
The systemic-layer publishing spine currently comprises four shipped anchors — Strategic Brief SB-01 (July 2026), Flash Brief F-01 (July 2026, extended F-02 September 2026), Themed Analysis TA-01 (July 2026), and the B1 Cascade and Compound Methodology Analysis (September 2026) — plus two peer-reviewed anchors, JIPR v16 for the Markov fleet-degradation substrate3 and Environmental Research: Energy for the R9 compound-hazard calibration.4
X1 sits inside that spine. It mounts on the B1 methodology substrate — specifically on Layer B sub-layer B1 (registered-asset spatial coverage) and sub-layer B2 (technology-sovereignty per-modifier) — and treats compute-infrastructure siting as a substation-level compound-cascade question rather than a data-centre-industry question.
The house discipline governing every piece in the spine is set out in the SSI Index reports-framing knowledge base. Five rules are load-bearing for X1.
Rule L · Reader inference, never verdict, always mechanism. X1 provides the empirical surface. It does not tell the reader that concentrated compute is wrong or that distributed compute is right. It reads what the per-substation modifier surface says, describes the mechanism, and leaves the policy inference to the reader.
Rule M · Mosaic theory. Every substantive quantitative claim in X1 is supported by a minimum of three independent publicly-cited inputs. Where mosaic derivation is used, the paragraph carries the M — mosaic-derived badge and confidence intervals are widened by a factor of 1.5.
Discipline N · Structured analytical writing. Six binding features: discernment-first, particular-before-universal, adversary-in-strongest-form, distinguished registers, optimising choice, systematic examination. In X1's cross-cutting register, particular-before-universal is especially load-bearing — the paper opens with two vignettes (Ahrtal and Storm-0558) before it generalises.
Three-tier epistemic register. Every figure and every quantitative statement carries a badge: T1 Measured (direct instrument or authoritative regulatory register), T1.5 Observed (multi-source cross-checked observation), T2 Modelled (projection or scenario, not a prediction).
DISC1 (Discernment) + DISC2 (Assertiveness). X1 is assertive on what is verifiable and cautious on what is projected. Where evidence is thin or contested — the Sandworm attribution on the Danish energy-sector cyber wave (SektorCERT medium confidence)5 or the completeness of Microsoft's Storm-0558 post-mortem (contested by the US Cyber Safety Review Board)6 — the paper flags the uncertainty in the text rather than absorbing it silently.
SSI Index published anchors: Strategic Brief SB-01 (July 2026, particularly §2.3 German hyperscaler-siting case, which X1 develops); B1 Cascade and Compound Methodology Analysis (September 2026, the mathematical substrate); Themed Analysis TA-01 Nodal and Locational Price Discovery (July 2026, the market-side companion); Flash Brief F-02 European Heatwave Second Edition (September 2026, the compound-hazard evidence anchor cross-referenced in §5).
EU regulatory canon: Cyber Resilience Act (Regulation (EU) 2024/2847, CELEX 32024R2847);11 AI Act (Regulation (EU) 2024/1689); NIS2 Directive (Directive (EU) 2022/2555); CER Directive (Directive (EU) 2022/2557); DORA (Regulation (EU) 2022/2554); Data Act (Regulation (EU) 2023/2854).
Evidence anchors: ENISA Threat Landscape 2026;12 CISA Advisory AA24-038A (Volt Typhoon);13 CISA Advisory AA24-060B (Ivanti Connect Secure zero-days); US Cyber Safety Review Board (CSRB) 2024 report on Microsoft Storm-0558; SektorCERT 2023 report on the Danish energy-sector cyber wave; GAO Report GAO-24-107733 on the CrowdStrike Falcon resilience critique.14
Six findings a busy desk can carry away.
One. Concentrated compute at hyperscaler scale is measurable as a per-substation R4 system-loading signature across all six cohort jurisdictions. Under a modelled 2029 AI-training-ramp scenario, the signature crosses the Re = 1.30 threshold at approximately 47 German, 22 Irish, 18 Dutch, 15 Spanish (within a 12–22 confidence band), 14 French, and 9 Swedish substations — roughly 125 substations across a cohort that together hosts well over 90% of the EU's installed hyperscale nameplate. M — mosaic-derived
Two. The four hyperscaler-dense German Länder — Nordrhein-Westfalen, Hessen, Bayern, and Berlin-Brandenburg — cross the threshold earliest of the four-Länder subset, consistent with the SB-01 §2.3 launch pad. When the six-jurisdiction cohort is normalised for grid headroom, the earliest crossing is Ireland's Dublin cluster. Spain's Aragón cluster crosses mid-cohort; Sweden's Boden crosses latest.
Three. Distributed compute architectures with the same nameplate megawatt absorb the AI-training-ramp shock without triggering the R4 × R6a network-coupling cascade that concentrated architectures produce.
Four. The EU regulatory stack (CRA + AI Act + NIS2 + CER + DORA + Data Act) creates a coordinated compute-supply-chain governance surface for the first time; CRA Article 14 became mandatory reporting on 11 September 2026 with a 24-hour early-warning / 72-hour full-notification / 14-day final-report cadence carried by ENISA's Single Reporting Platform.15 CRA Article 12 bridges to AI Act Article 15.
Five. The fourteen-incident cyber evidence base 2022-2026 demonstrates the mechanism: from Viasat/AcidRain wiping remote control of 5,800 German wind turbines (February 2022),16 through MOVEit compromising TenneT (May 2023),17 the Danish energy-sector Zyxel monoculture wave compromising 22 operators (May 2023), to Volt Typhoon pre-positioning inside US Communications-Energy-Transportation-Water operators (February 2024) and CrowdStrike's civilisation-scale single-vendor concentration event (July 2024). ENISA Threat Landscape 2026 documents that 73% of victims in its 2025 corpus were essential or important entities under NIS2.
Six. Grey-zone × civic-equity dimensions are not additional dimensions bolted on to the compute-sovereignty argument. They are the third and fourth faces of the same per-substation surface, mediated through R4, R6a, and R10 (recognition-justice weighting per Reckien 2023). The four faces resolve on the same surface; the policy question is which face is being read at any given decision point.
On the night of 14–15 July 2021, an extraordinary rainfall event over the Eifel massif drove the Ahr river into a compound-cascade failure across Rhineland-Palatinate and Nordrhein-Westfalen. More than 180 people died. Munich Re's NatCat 2021 aggregation put direct damages at approximately €33 billion, the largest single-event insured natural-hazard loss in German history.18 Substations along the Ahr valley suffered simultaneous flood inundation and downstream cascade transmission-line outages; the SSI Index v4.2 canonical for Germany records the Ahrtal cluster as the anchor instance for the R6a network-coupling modifier compounding with R4 system loading.
What is less remarked in the four-year post-mortem is that the same substation cluster sits inside the Frankfurt-Rhine-Main hyperscaler catchment. The valley that drowned in July 2021 supplies the substations that will absorb — or fail to absorb — the AI-training-ramp load projected for the second half of this decade.
Two years later, on a different substrate, the same coupling logic surfaced in the compute plane. Between May and July 2023, the China-nexus threat actor tracked as Storm-0558 forged Azure Active Directory authentication tokens using a stolen Microsoft consumer signing key — a key that had been captured in a 2021 Microsoft crash dump.19 The forged tokens gained access to Exchange Online mailboxes of approximately 25 organisations. Among them were the US Department of State, from which approximately 60,000 emails were exfiltrated.
The US Cyber Safety Review Board's 2024 report on the incident used unusually pointed language. It found Microsoft's security culture to have been "inadequate" and characterised the full chain of failures as avoidable.
Physical and logical converge on the same modifier surface. That is the analytical claim this Themed Analysis develops. The question is: does the grid already price the compute question at per-substation granularity, and if it does, what is it telling us?
The frame is deliberately named: X1 reads the 4IR civil-society surface — the surface on which the Fourth Industrial Revolution's compute-infrastructure substrate meets the civil-society institutions (grids, communities, environmental limits, democratic accountability) that have to carry it. Previous SSI Index pieces have read the grid-adaptation surface in its own terms; X1 is the first to read it in the terms the 4IR imposes on it — hyperscaler-scale AI training clusters, concentrated compute operators, cross-border data flows, and the recognition-justice question that compute-siting decisions put to the communities whose substations carry the load.
Each takeaway restates a finding from the executive summary and pairs it with a "what this does not mean" clause. The pairing is a Rule L discipline: the finding is the mechanism; the disclaimer is the boundary of the mechanism.
Finding. The R4 system-loading modifier reads the concentrated-compute signature across all six cohort jurisdictions using existing regulatory data feeds — TSO connection queues, DSO load-flow snapshots, and the harmonic-distortion and reactive-power provisions of IEEE 519-201420 and IEC 61000.
What this does not mean. The R4 signature is not a verdict on siting. It is a modifier reading.
Finding. Under a modelled 2029 AI-training-ramp trajectory, the four hyperscaler-dense German Länder cross Re = 1.30 earlier than the equivalent clusters in Ireland, the Netherlands, France, Sweden, and Spain considered in absolute terms. Normalised for grid headroom, Ireland's Dublin cluster crosses first.
What this does not mean. Neither ordering ranks jurisdictions by policy quality, regulatory maturity, or attractiveness for hyperscaler investment.
Finding. The same one gigawatt of installed compute can be either a grid stabiliser or a grid stressor depending on how it is placed.
What this does not mean. Distributed architectures are not costless. They increase land-use footprint, replicate physical security perimeters, and complicate operational orchestration.
Finding. The CRA, AI Act, NIS2, CER, DORA, and Data Act together compose a compute-supply-chain governance surface that did not exist coordinated in this form even eighteen months ago. CRA Article 12 bridges to AI Act Article 15: a high-risk AI system that satisfies CRA essential requirements is deemed to satisfy AI Act cybersecurity requirements.
What this does not mean. The regulatory stack does not force sovereign siting of hyperscaler workloads. CRA is a product regulation; NIS2 regulates operators; DORA covers financial-sector operational resilience. The future EUCS and Member-State cloud rules carry the sovereign-siting lever.
Finding. Six of the fourteen incidents have direct grid or EU-critical-infrastructure touch: Viasat/AcidRain (2022), Danish energy-sector Zyxel wave (2023), MOVEit/TenneT (2023), Volt Typhoon (2024), Collins Aerospace MUSE across five EU airports (Sep 2025), and the Polish ELECTRUM destructive OT campaign (Dec 2025). The 2026 EC Europa.eu AWS breach adds a first-of-kind sovereignty-anchor case at the EU-executive-body layer, and the DORA aggregate 3,383 major ICT incidents recorded in 2025 provides the statutory-scale floor. ENISA's Threat Landscape 2026 records that 73% of the victims in its 2025 corpus were essential or important entities under NIS2.
What this does not mean. No single incident is a proof by itself. Attribution on the Danish energy-sector wave sits at medium confidence per SektorCERT. Microsoft's post-mortem on Storm-0558 is contested by the CSRB. CrowdStrike was a vendor defect, not a cyberattack.
Finding. All four faces resolve on the same modifier surface. The policy question is which face is being read at any given decision point.
What this does not mean. The paper does not claim the four faces are commensurable at a single scalar.
In scope. Extraction of per-substation R4, R6a, R6b, V, I, and R3 modifier traces from the SSI Index v4.2 canonicals for each of the six cohort jurisdictions; superposition of modelled hyperscale nameplate growth; application of the B1 §7 five-sigma tail prism and the B1 §6 Gaussian copula with 20×20 tail-inflation; overlay of the W1–W10 anti-maladaptation gate mesh.
Out of scope. Enumeration of the full EU cohort; extension to sovereign-cloud or edge-compute footprints beyond the hyperscale visible-capacity register; survey-based civic-welfare metrics.
| Jurisdiction | TSO | Regulator | AI Act competent authority |
|---|---|---|---|
| Germany | 50Hertz · Amprion · TenneT DE · TransnetBW | BNetzA | BMDV coordinating |
| Ireland | EirGrid | CRU + SEMO | DETE |
| Netherlands | TenneT NL | ACM | Rijksinspectie Digitale Infrastructuur |
| France | RTE | CRE | DGE |
| Sweden | Svenska kraftnät | Energimarknadsinspektionen (EI) | IMY coordinating |
| Spain | Red Eléctrica de España (REE) | CNMC | AESIA (Coruña) |
X1's findings are bounded by six explicit constraints. (a) Six jurisdictions only. (b) Only hyperscale visible-capacity is on-surface. (c) The 2029 AI-training-ramp trajectory is a T2 modelled scenario. (d) R10 uses the Reckien 2023 recognition-justice framework, not preference-elicitation. (e) IEEE 519-2014 and IEC 61000 conventions. (f) The fourteen-incident cyber evidence base draws from publicly-attributed forensic reports; where attributions are contested this is flagged in the text.
The four layers below are not independent. They are four faces of a single per-substation surface.
An AI training cluster is not a steady-state industrial load. Its sub-second signature is stochastic: batch swaps, gradient-synchronisation transients, activation-shifting between attention layers, and periodic checkpoint flushes each produce a distinct electrical fingerprint. Under the harmonic-control conventions of IEEE 519-2014 and the electromagnetic-compatibility conventions of IEC 61000, this fingerprint is read on six dimensions: total harmonic distortion (THD); reactive-power demand; voltage-stability margin; inrush-current profile; flicker in the 0.1 Hz to 25 Hz band; and voltage-notching at inverter switching frequencies.
What is distinctive about hyperscaler-scale AI training load is the combination. A one-gigawatt training cluster is a superposition of several thousand independent inverter-driven loads whose stochastic behaviour couples through shared substation infrastructure. The R4 system-loading modifier captures this coupling.
Figure 5 traces R4 trajectory 2026–2031 as a six-panel small-multiple.
Germany carries four hyperscaler-dense Länder (NRW, Hessen, Bayern, Berlin-Brandenburg) against a twelve-Länder baseline. The four cross Re = 1.30 inside the 2029 window under sudden-ramp. Ireland — Dublin cluster crosses earliest of the cohort when normalised for grid headroom. The Netherlands — Amsterdam metropolitan region crosses in the middle band. France — Île-de-France crosses later, reflecting RTE's transmission-reinforcement programme. Sweden — Boden crosses latest, reflecting abundant northern-Swedish headroom. Spain carries Aragón (AWS + Microsoft Villamayor/Villanueva de Gállego clusters), Madrid (Microsoft metropolitan-adjacent), Cataluña (Barcelona-adjacent), and Castilla-La Mancha (Meta Talavera de la Reina).21 The Spanish cluster trace crosses in the middle band. M — mosaic-derived
Under the sudden AI-training-ramp trajectory, approximately 125 substations across the six-jurisdiction cohort cross Re = 1.30 within the 2029 window — 47 German, 22 Irish, 18 Dutch, 15 Spanish (12–22 confidence band), 14 French, 9 Swedish. M — mosaic-derived · T2 modelled
Scale of reading note. The R4 modifier is a per-substation measurement. The same cohort nameplate reads differently when aggregated at the per-Member-State level (6 bins), the per-Land/per-comunidad level (~90 bins across the cohort), or the per-substation level (796,121 bins at Wave-4 TERMINAL closure). The spread between the three readings is itself a data-row: a cohort nameplate that reads as "stable" at Member-State resolution produces 125 Re = 1.30 crossings when read at substation resolution. The paper's claims are made at the per-substation resolution because that is the resolution at which the R4 modifier is calibrated; coarser resolutions understate the modifier surface in ways that are reproducible per publication cycle.
The compute-sovereignty question was, until eighteen months ago, an industrial-policy question with fragmented regulatory instruments. That has changed.
Six industrial-policy instruments frame the EU landscape as it stands in September 2026. EuroStack (2024-), GAIA-X (2020-), Microsoft's EU Data Boundary (February 2025), the EU Chips Act (~€43 bn),22 the US CHIPS Act ($52 bn), and the Korean K-Chips Act.
The 1973 oil-shock analogue is instructive. In the aftermath of the OPEC embargo, individual consuming nations responded with fragmented conservation and stockpiling measures. Coordination arrived through the International Energy Agency, established November 1974 by OECD Member States precisely to produce a shared frame for what had been parallel bilateral responses to a common supply-side concentration.
The AI Act — Regulation (EU) 2024/1689. The AI Act is the load-bearing regulatory instrument for the compute-sovereignty argument at the workload layer. Adopted 13 June 2024, entered into force 1 August 2024, its provisions phase in through August 2027. Article 6 high-risk classifications became applicable from 2 August 2026 — six weeks before this paper — and cover AI systems in eight enumerated domains including critical infrastructure management (Annex III §2), which brings grid-adjacent AI applications (load-forecasting, dispatch-optimisation, cascade-prediction) directly into the high-risk classification where they support decisions with significant impact on natural persons or the functioning of critical infrastructure. Article 15 imposes cybersecurity requirements on high-risk AI systems — including robustness against adversarial examples, model-poisoning, evasion attacks, and confidentiality attacks. Chapter V general-purpose AI obligations (Articles 51–55) became generally applicable on 2 August 2025, with transitional provisions for GPAI models placed on the market before that date (full compliance by 2 August 2027). Article 51 defines the systemic-risk classification; the compute-threshold trigger (10²⁵ FLOPs of training compute in the base rule) sits at the level of the frontier models trained on hyperscaler-scale infrastructure. Articles 53–55 impose risk-assessment, adversarial-testing, incident-reporting and cybersecurity obligations on providers of systemic-risk GPAI. Article 50 transparency obligations are applicable from 2 August 2026 and cover interaction with humans (chatbot disclosure), emotion-recognition and biometric-categorisation systems, and deepfake / AI-generated content disclosure. Per-Member-State AI Office designation is underway; Spain's Agencia Española de Supervisión de la Inteligencia Artificial (AESIA) in Coruña is the earliest-designated national AI Office among the six cohort jurisdictions.
The Cyber Resilience Act — Regulation (EU) 2024/2847. The AI Act's product-side co-anchor. Adopted 23 October 2024, published in OJ L 2024/2847 on 20 November 2024, entered into force 10 December 2024, full applicability 11 December 2027. The CRA imposes horizontal cybersecurity requirements on all products with digital elements placed on the EU market, plus their remote data-processing solutions.
Three product tiers govern conformity. The default class requires manufacturer self-assessment against the Annex I essential requirements. Important products listed in Annex III sit in Class I (self-assessment permitted if harmonised standards are applied, otherwise third-party) and Class II (mandatory third-party assessment). Critical products listed in Annex IV require mandatory European cybersecurity certification.
The compute-sovereignty hooks in the CRA scope are direct. Annex III Class II includes hypervisors and container runtimes — the orchestration substrate of every hyperscaler cloud fabric — alongside firewalls, intrusion-detection systems, and tamper-resistant microprocessors. Annex IV Critical covers smart-meter gateways within smart-metering systems as defined in Directive (EU) 2019/944.23
The obligations are substantial. Secure-by-design and secure-by-default; a Software Bill of Materials (SBOM); free security updates through a support period of at least five years; CE marking; technical documentation; post-market vulnerability handling; a reporting cadence of 24 hours / 72 hours / 14 days. Fines reach €15 million or 2.5% of worldwide annual turnover.
The reporting cadence became mandatory on 11 September 2026 — fourteen days before this paper's publication. ENISA operates the Single Reporting Platform through which manufacturers of products with digital elements now notify actively-exploited vulnerabilities and severe incidents.24 Notably, as of 11 September 2026, zero notified bodies had been designated in NANDO for the CRA, and the 11 December 2026 Article 35(2) capacity target for a "sufficient number" of notified bodies is now assessed as unlikely to be met.25
AI Act Article 15 × CRA Article 12 interlock. The AI Act's cybersecurity Article 15 and the CRA's Article 12 deeming provision interlock: a high-risk AI system that satisfies CRA essential requirements is deemed to satisfy AI Act Article 15 cybersecurity requirements.26 The bridge matters because the same product — a hyperscaler-hosted grid-optimisation AI system, say — sits inside both regulatory nets simultaneously: the AI Act as an AI system, the CRA as a product with digital elements. Without the bridge, the manufacturer would satisfy the two regimes independently. With the bridge, one conformity path serves both. This is the first EU-scale example of AI-specific regulation and product-cybersecurity regulation being explicitly wired together. One architectural property of the Article 12 deeming provision is worth recording: conformity is assessed per product, not per deployment architecture. A single AI system deployed on one hyperscaler, on multiple hyperscalers, or on a federated multi-operator substrate carries the same conformity obligations; the regulatory stack is architecturally neutral between concentrated single-vendor and distributed multi-vendor implementations.
NIS2 × CRA. NIS2 regulates operators; CRA regulates products. A single compromise touching a regulated operator and involving a CRA-scope product will produce two notifications.
DORA imposes ICT third-party risk-management obligations including concentration-risk assessments for critical ICT service providers. DORA's concentration-risk tests are the closest existing analogue to sovereignty framing in EU law.
Data Act mandates data access and portability for connected products; coordination with CRA on SBOM disclosure is a Commission implementing-act priority through 2026–2027.
The CRA is a product regulation, not a service regulation. It does not itself force sovereign siting of hyperscaler workloads. That lever sits with NIS2, DORA, the future EUCS scheme, and Member-State cloud rules such as France's SecNumCloud or Germany's C5 framework. What the CRA does is force EU-market gatekeeping on third-country-manufactured critical products — the supply-chain sovereignty layer.
Consolidated AI Act calendar for the paper's forward window (developed in §4.2.2 above). Chapter V GPAI (Art. 51–55) generally applicable 2 Aug 2025 with transitional provisions (full compliance by 2 Aug 2027) · GPAI Code of Practice voluntary adoption pathway live · Article 6 high-risk live 2 Aug 2026 · Article 50 transparency applicable 2 Aug 2026 · Article 15 cybersecurity deeming bridge with CRA Article 12 live now for high-risk products · Per-Member-State AI Office designation underway (AESIA earliest of cohort). The load-bearing X1 hook is that grid-adjacent AI applications (Annex III §2 critical-infrastructure-management scope) are already inside the Article 6 high-risk regime as of the paper's publication date, and GPAI obligations (including on systemic-risk frontier models trained at hyperscaler scale) have already been live for over a year.
Ireland carries the highest single-vendor concentration in the cohort (approximately 90%+ top-three share). Netherlands ≈ 75%. Sweden ≈ 70%. Germany ≈ 60%. France ≈ 55%. Spain ≈ 50–65%. M — mosaic-derived The 1973 oil-shock analogue holds: high per-jurisdiction concentration in the compute-supply layer, matched to an emerging but still-incomplete coordinated regulatory response.
The Top-3 share reported in §4.2.5 is a 2026 reading. The brief asks for trajectories, and the trajectory question is answered by superposing the modelled hyperscale nameplate growth (§4.1) and the AI-training-ramp window (2029 mid-window) onto the 2026 Top-3 shares. Under the sudden AI-training-ramp trajectory, and in the absence of a counter-concentration policy instrument, Top-3 hyperscaler share evolves per cohort jurisdiction as follows: Ireland 90%+ → 92%+ (cohort ceiling, operator Big-Three oligopoly already functionally closed); Netherlands ~75% → ~78%; Sweden ~70% → ~73%; Germany ~60% → ~64%; France ~55% → ~57%; Spain ~50–65% → ~55–68% (band widens as four comunidades — Aragón, Madrid, Cataluña, Castilla-La Mancha — build nameplate at different rates). M — mosaic-derived · T2 modelled · AI-training-ramp trajectory
Two readings fall out of the trajectory table. First, the cohort trajectory is concentration-reinforcing under current trajectory: not one of the six jurisdictions shows a mechanical Top-3 share decline through 2029 under the sudden-ramp trajectory. Second, the trajectory slope is non-uniform — the four jurisdictions where Top-3 share is below 70% in 2026 are also the four jurisdictions where the SICI designation scorecard (§4.3, Figure 12) does not pre-position a load-bearing designation. The relationship between present Top-3 share and the SICI-designation-threshold candidate convention (CS.04) is itself a data-row the publication cycle will refresh. The ~€200-billion "Choose Europe for AI Compute" wave (§4.5.5) represents the single largest incoming nameplate variable; the trajectory above holds the wave's geographic distribution at the Commission's indicative 2025 announcement pattern. If that pattern shifts materially before the next refresh cycle, the trajectory re-reads.
The financial-sector post-2008 literature produced a concept — the Systemically Important Financial Institution (SIFI) — and an associated regulatory apparatus operationalised over the fifteen years since.27 The compute-sector analogue is available in template form but has not been institutionalised. Policy proposals through 2024–2026 have converged on a Systemically Important Compute Institution (SICI) framework using an analogous five-criterion designation: installed-capacity share, cross-vendor interconnectedness, substitutability index, complexity, and cross-jurisdictional activity. The SICI proposal is not yet EU law.
Ahrtal, July 2021 (physical). The R6a network-coupling anchor. Storm-0558, May–July 2023 (logical). The logical-plane analogue. Iberian blackout, 28 April 2025 (adjacent — coordination-surface, not cyber). ENTSO-E's Expert Panel Final Report published 20 March 2026 attributed the event to voltage-stability control gaps and oscillations at the Franco-Spanish interconnection.28 The Panel found no evidence of cyber origin. It is included here as an adjacent coordination-surface reference.
The B1 Cascade and Compound Methodology Analysis provides the mathematical apparatus X1 mounts on: Gaussian copula with tail inflation (B1 §6) and the five-sigma tail prism (B1 §7). The mathematical substrate is what makes the paper's Rule L discipline sustainable — the mechanism is derived, not asserted.
Interaction-term sensitivity to geographic correlation of placement. The B1 §7 five-sigma tail prism decomposes concurrent-cascade probability into five additive components: physical-load, physical-coupling, cyber-incident, cyber-coupling, and the interaction term. The interaction term is where the mechanism's sensitivity to compute-placement geography lives — the term scales with the covariance of training-ramp placement across the cohort. At the measured 2026 geographic distribution of hyperscale nameplate (Top-3 shares reported in §4.2.5), the interaction term carries approximately 42% of the five-sigma tail concurrence probability. A reproducible counterfactual — a 25% redistribution of hyperscale nameplate toward presently-under-weighted sub-national regions within the same six-jurisdiction cohort — moves the interaction term's contribution to approximately 28% of five-sigma tail concurrence. The two readings bracket the geographic sensitivity. The decomposition is a measurement; the two numbers are a data-row for the reader. M — mosaic-derived · T2 modelled · interaction-term bracket
Fourteen publicly-attributed cyber incidents from the four-year window 2022–2026 constitute the historical proof-of-mechanism base for the cascade layer.
Vignette 1 — Viasat KA-SAT / AcidRain, 24 February 2022. On the morning of Russia's invasion of Ukraine, Sandworm (attributed to GRU Unit 74455) pushed the AcidRain wiper into the KA-SAT satellite-communications management plane. The direct grid-adjacent spillover was that remote control of 5,800 wind turbines in Germany was lost when modems were wiped. A single US-operated satellite communications service constituted the single point of failure. Grid touch: direct.
Vignette 2 — Danish energy sector / Zyxel, May 2023. Coordinated waves exploiting Zyxel firewall vulnerabilities compromised 22 Danish energy companies — what SektorCERT described as Denmark's largest critical-infrastructure cyber incident to date. A single third-country vendor's firewall products, deployed uniformly across a national energy-sector operator base, produced a monoculture vulnerability. Grid touch: direct — substation and DSO perimeter.
Vignette 3 — MOVEit / Cl0p, May 2023 onwards. Cl0p exploited a SQL-injection vulnerability in Progress Software's MOVEit Transfer product in a mass-scale supply-chain compromise event that reached approximately 2,700 organisations. The European victim list included the Dutch transmission-system operator TenneT, alongside Shell, ING, TomTom, Toyota Boshoku Europe, the BBC, and British Airways. Grid touch: direct — EU TSO named victim.
Vignette 4 — Volt Typhoon, disclosed February 2024. CISA, NSA, FBI, and Five Eyes partners published Joint Cybersecurity Advisory AA24-038A confirming that a People's Republic of China state-sponsored threat actor tracked as Volt Typhoon had compromised the IT environments of multiple US critical-infrastructure operators in the Communications, Energy, Transportation, and Water sectors — including on Guam — with the express intent, per CISA, "to enable lateral movement to OT assets to disrupt functions". Tradecraft was living-off-the-land. Grid touch: direct — Energy sector explicitly named.
2026 anchor — European Commission Europa.eu AWS breach (24 March 2026). On 24 March 2026, ShinyHunters exfiltrated approximately 350 GB from an AWS account hosting European Commission web infrastructure — including mail-server dumps, contracts, staff records, and databases across roughly thirty EU entities. Most seriously, the exfiltration included DKIM signing keys for official EC email domains, giving the attackers the material to spoof official Commission correspondence. Initial access on 19 March traced to a supply-chain compromise of the open-source Trivy dependency scanner. This is the second EC breach in the same year (a February 2026 mobile-device-management incident preceded it). CERT-EU confirmed the vector.30 The sovereignty read is textbook: the EU's own executive body is hosted on a US hyperscaler and was breached through a US-supplied dependency-scanning tool. DKIM key theft converts concentrated third-country hosting into a persistent identity-forgery risk that a sovereign siting (EU-controlled key management, EU-hosted CI/CD, distributed key custody) would have contained. The event directly rebuts the "hyperscaler-in-EU-region equals sovereign enough" line. Enabling vector of record (CERT-EU): a single hyperscaler account hosting approximately thirty EU-institution tenants, breached through a single open-source dependency-scanning tool used uniformly across that account's CI/CD pipeline.
2025 grid-touch anchor — Polish power-grid ELECTRUM campaign (29-30 December 2025). A coordinated destructive OT campaign attributed by Dragos to the ELECTRUM threat cluster (Russia-nexus, active against electric infrastructure since 2015) hit at least thirty wind and solar sites, a major combined-heat-and-power plant, and a manufacturing firm in Poland over the year-end weekend. Remote terminal units, protection relays, human-machine interfaces, and serial device servers were physically damaged; domain-wide data-destruction attempts were observed in corporate IT. DSO substations lost remote-control communications, though generation continued. CERT Polska and Dragos published independent forensic reports.31 This is the highest-priority grid-touch event of 2025 for the sovereignty argument: damage concentrated at the distributed-edge layer that renewables integration is adding at scale. A sovereign compute topology with EU-controlled OT-SCADA, local RTU firmware provenance, and segmented DSO orchestration would have limited blast radius — the paper's distributed-versus-concentrated thesis (§4.1.2) reads directly onto this case.
2025 concentration anchor — Collins Aerospace MUSE ransomware (19 September 2025). Ransomware on Collins Aerospace's MUSE shared check-in and boarding platform forced Heathrow, Brussels, Berlin-Brandenburg, Dublin, and Cork airports into manual processing for several days despite none of the five airports being directly breached. More than a hundred flights delayed or cancelled. Everest ransomware crew claimed responsibility (HardBit variant).32 This is the cleanest single-vendor concentration case in the corpus: a single US-owned SaaS operator became the point of failure for five EU-sovereign airports across four Member States — precisely the NIS2 essential-service supply-chain scenario. A distributed or federated check-in stack, or a mandated EU-sovereign fallback, would have degraded gracefully rather than dropping every airport to pen-and-paper. Enabling vector of record (airport CSIRTs): a single US-owned SaaS vendor (Collins Aerospace) providing the MUSE shared check-in platform uniformly across five EU-sovereign airports in four Member States — a concentration ratio of 5:1 at the operator-SaaS-tenant layer across the European-airport NIS2 essential-service cohort.
Statutory anchor — DORA aggregate reporting (2025). The Joint European Supervisory Authorities' 2025 report on the first full year of DORA mandatory reporting: 3,383 major ICT incidents in the EU financial sector · approximately one-third spread cross-border · approximately eight percent affected ten or more countries simultaneously.33 The ESAs explicitly attributed cross-border contagion to "growing dependence on shared technology providers, common infrastructure and multinational business models". On 18 November 2025, the ESAs designated nineteen Critical Third-Party Providers — including Amazon Web Services, Microsoft Azure, and Google Cloud — as directly EU-supervised under DORA. The number is important because it moves the concentration-risk argument from academic to statutory: 3,383 is the number the EU's own financial-sector regulators recorded and reported in the first full DORA cycle, and cross-border contagion is what they attribute to hyperscaler concentration.
Anchor incident — Storm-0558 — treated in §1 and §4.3.2. The archetypal sovereignty case at the platform-key-management layer.
Concentration incident — CrowdStrike Falcon, 19 July 2024. A faulty content update crashed approximately 8.5 million Windows devices globally. Not a cyberattack — a vendor defect. The GAO's resilience critique (GAO-24-107733) uses the event as the load-bearing case for a broader argument about single-vendor concentration risk at civilisation scale. CRA Article 14, live since 11 September 2026, would capture equivalent future events.
Policy anchor — ENISA Threat Landscape 2026. ~4,900 curated incidents analysed. 73% of victims in the ENISA corpus are essential or important entities under NIS2. ENISA characterises the trajectory directly: "State-aligned threat groups intensified their long-term cyberespionage campaigns against the telecommunications, logistics networks, and manufacturing sectors in the EU."
The W3 systemic-substitutability gate is the load-bearing test for the cascade-amplification-through-concentration finding. Compute-supply concentration at hyperscaler scale fails W3 to the degree the concentration is measurable and rising. W3 is not a binary gate. It reads substitutability as a position on a continuum — the HCR hyperscaler-concentration-ratio candidate convention (CS.02) carries the measurement, and the paper's publication cycle makes that measurement reproducible each pass. The present reading is a trajectory snapshot; the direction of travel is what each refresh cycle will record.
Distributed is not automatically safer. Reducing geographic concentration lowers one failure mode but can raise others. Shared software stacks and shared identity-management systems (SSO, federated identity, common CI/CD pipelines) can carry a single compromise across geographically distributed sites — Storm-0558 is the textbook case. More sites add physical-security surface area, operational complexity, and operator-count. The paper's finding is that concentration is one measurable failure mode, not that distribution is a sufficient answer. The analytical distinction worth recording is between architectural distribution — more sites under common identity, dependency, and operator — and substitutability-weighted distribution, which the W3 continuum reads: sites with independent identity planes, independent dependency stacks, and independent operators. The HCR candidate convention (CS.02) and the W3 gate measure the latter; a count of sites alone measures the former.
European hosting is not cybersecurity. EU-region siting of a hyperscaler workload addresses data-residency and some jurisdictional-reach questions. It does not of itself change the identity-plane, supply-chain, firmware, or dependency-scanning attack surfaces. The 24 March 2026 European Commission Europa.eu AWS breach (§4.3.5) is the direct empirical point: the exfiltration happened at a US hyperscaler EU-region deployment, through a US-supplied dependency-scanning tool, against the EU's own executive body.
The distributed-architecture cascade-prevention claim is mechanistic, not empirically demonstrated. The paper's §5 cascade reading is grounded in the B1 §6 Gaussian copula and §7 five-sigma tail prism — a mathematical mechanism derived from the SSI Index v4.2 modifier substrate. Empirical demonstration that deliberately distributed compute architectures prevent the modelled cascade (as distinct from relieving transmission congestion or balancing nameplate demand) is not available in the published literature as of September 2026. The current state of adjacent evidence is summarised in the table below.
| Evidence | What it supports | Limitation for X1 |
|---|---|---|
| Thimmel et al., Energy Policy, 2019 (peer-reviewed) | Simulations using real-world data show how moving workloads between geographically distributed data centres can help balance electricity systems. | Supports the mechanism and economic rationale; does not demonstrate prevention of cascading failures. |
| Wan et al., arXiv preprint, revised March 2026 | On a modified IEEE 73-bus test network, spatial workload flexibility removed transmission overloads in the studied scenarios. Closest adjacent evidence to the X1 claim. | A modelled test network; congestion relief does not establish dynamic cascade prevention. |
| PNNL September 2026 (national-laboratory report) | Simulations show coordinated data-centre controls, batteries, and onsite generation can prevent congestion and support operation during contingencies. | Benefits depend on the combined resources and controls; they cannot be attributed to geographical distribution alone. |
| Google 2025 (operational demonstration with Omaha Public Power District) | Google reports reducing machine-learning electricity demand during three grid events in 2024. | Useful operational evidence, reported by the operator; it does not compare concentrated and distributed architectures. |
The gap — a controlled empirical comparison of concentrated and distributed hyperscale compute architectures under cascade conditions — is a research-pipeline item. X2 (Themed Analysis · February 2027) is scoped to develop the workload-flexibility and grid-service-provision angle at the depth that gap requires. The X1 finding stands as written: concentration is a measurable failure mode on the SSI Index v4.2 modifier surface. The claim that distribution is the answer is not what this paper argues; the claim is that the surface should be read, and the mesh gates applied, before the next round of hyperscaler-siting decisions is taken.34
The fourth face of the compute-siting surface is measurable, but not as a scalar; it requires a framework choice. The paper adopts the framework proposed by Reckien et al. (2023), which extends the earlier distributional-justice literature by making recognition — the political and cultural standing of affected communities — an equal-weight axis alongside distribution and procedure. The civic-equity dimension is not an add-on; it is mediated through the same modifiers that carry the other three layers.
Frankfurt-Rhine-Main and Ahrtal. The same substation cluster serving the Frankfurt-Rhine-Main hyperscaler catchment provides last-mile distribution to communities across the Ahr valley. Under Reckien 2023: recognition — the Ahr-valley communities have no formal standing in hyperscaler-siting decisions taken at the Frankfurt metropolitan-planning level; procedure — the two decision surfaces do not intersect procedurally; distribution — grid-reinforcement costs are socialised across the same network-user base that includes the flood-vulnerable communities.
Dublin cluster and rural Ireland. Grid-reinforcement investment required to accommodate Dublin hyperscaler load growth is funded through the national network-tariff structure administered by CRU. Rural non-Dublin network users contribute to the socialised cost of Dublin-centric grid reinforcement.
Boden Sweden and Sami territorial recognition. Northern Sweden's Boden hyperscaler siting — Facebook Luleå precedent — sits within traditional Sami reindeer-grazing territories. Sweden has not ratified ILO Convention 169.29 The Sami Parliament (Sametinget) has consultative but not decisional standing in Swedish national-scale infrastructure planning.
The W6 data-sovereignty gate and the W7 local-value-retention gate are the civic-equity guardrails already codified in the SSI Index framework. The W10 compute-economic-multiplier gate sits at the boundary between the civic layer and the C2 Adaptation-Justice Themed Analysis scheduled for March 2027, which will develop the R10 reading at full depth.
The paper's brief names four convergent facets: grid-resilience, compute-infrastructure, civic-equity, and climate-adaptation. The climate-adaptation face carries a distinctive analytical burden. Where the other three faces read the compute-siting surface at essentially static exposure — the substation is where it is, the hyperscaler is where it is, the affected community is where it is — the climate face reads that same surface under a moving substrate. The R6a network-coupling modifier and the R4 system-loading modifier become intermittently more stringent under heatwave conditions; R9 compound-hazard is the SSI Index modifier that carries this movement explicitly.
R9 is the SSI Index modifier calibrated on the F-01 and F-02 empirical evidence bases: heatwave-derated substation traces (F-01) and heatwave-triggered cascade traces (F-02). The Environmental Research: Energy publication (doi:10.1088/2753-3751/ae87a5) provides the R9 calibration. What R9 measures at per-substation granularity is the degree to which two or more stressors that would each be individually manageable become jointly unmanageable when their probabilities are conditionally correlated through shared infrastructure. For grid-adjacent hyperscale compute, R9 reads the joint occupancy of heatwave-derated transmission capacity and elevated compute-cooling water-and-power draw as one signal, not two.
The six-jurisdiction cohort presents a graded R9 profile. Spain and France — southern Iberia and southern-French clusters — carry the highest R9 loadings, with 2022 and 2023 heatwave-derating events well within the F-01 evidence window. Germany — the Frankfurt-Rhine-Main catchment — carries the highest R6a×R9 interaction term because the same substation cluster is exposed both to F-02 flood-cascade signatures (Ahrtal July 2021) and to F-01 summer-heatwave derating. Ireland and Netherlands — maritime-climate exposure — present moderate R9 loadings but rising trajectory. Sweden — Boden northern climate — presents the lowest present R9 but the fastest transition slope in the F-01/F-02 traces, because the northern-latitude adaptation range is comparatively narrow.
R9 heterogeneity is itself the measurement. The R9 profile across the cohort is non-uniform at per-substation granularity: the pairwise covariance of R9 across the four hyperscaler-dense Länder (NRW, Hessen, Bayern, Berlin-Brandenburg) is substantially higher than the pairwise covariance of R9 across any four substations drawn from different jurisdictions in the cohort. The data-row the reader can carry away: compute placement at hyperscaler scale implicitly makes a bet on R9 covariance; the geography of the placement determines whether the bet is on correlated or uncorrelated climate stressors. Portfolio-theory grammar applies to the measurement without being invoked as a prescription.
The B1 Systemic Layer §7 five-sigma tail prism decomposes a scenario in which a Danube-basin summer heatwave of 2003/2022 magnitude coincides with elevated compute-training demand in the Frankfurt-to-Vienna corridor. The scenario reads as a compound R4×R6a×R9 exceedance across a transmission-coupled subnetwork. This is not the paper's principal empirical claim — it is the stress-injection developed in §5 above, projected specifically onto the climate face. The reason to name the scenario here is to make explicit that climate-adaptation is not a bolt-on to the sovereignty argument: it is one of the four faces on which the same modifier surface is read, and R9 provides the direct instrument.
The W1 loss-and-damage gate and the W9 ecological-boundary gate are the anti-maladaptation gates in the W1–W10 mesh most directly relevant to the climate face. Compute-siting at hyperscale scale routinely re-siting to lower-cost or lower-latency locations without accounting for R9 compound-hazard exposure at the new siting is precisely the maladaptation the mesh is calibrated to flag. The paper flags — does not adjudicate — that the €200-billion "Choose Europe for AI Compute" wave includes no explicit R9 compound-hazard filter, and that this omission is a candidate for W1 and W9 gate review.
The structured-analytical-writing register in which this Themed Analysis is written requires a surprise injection: a compound-scenario read that stress-tests the analytical spine. The injection X1 uses is a triple concurrence: the modelled 2029 AI-training-ramp window coincides with (a) a European heatwave of the magnitude documented in F-01 and F-02 extending to a Danube-basin scale; and (b) a Volt-Typhoon-class pre-positioned cyber campaign transitioning from the pre-positioning phase to an activation phase.
The B1 §7 five-sigma tail prism reads the concurrence in a five-component additive decomposition. The interaction term is where the analytical claim lives. The concurrence probability is not the product of the three marginal probabilities. It is amplified by the fact that all three vectors are conditionally coupled through shared infrastructure.
Under central concurrence, four of six jurisdictions cross Re = 1.30 (Germany, Ireland, Netherlands, Spain). Under 1σ amplified concurrence, all six cross. Under 5σ amplified concurrence, four of six cross R = 1.50 Compound-Extreme — a band achieved historically only in the aftermath of the 2021 Ahrtal event.
The scenario is not a prediction. It is a stress-test injection.
W3 systemic-substitutability. Load-bearing for the cascade layer. Compute-supply concentration fails W3 to the degree the concentration is measurable and rising. W6 data-sovereignty. Load-bearing for the sovereignty layer. Storm-0558 is the canonical case. W7 local-value-retention. Load-bearing for the civic layer. Heterogeneous outcomes across the case-study triptych. W10 compute-economic-multiplier. Load-bearing for labour-market trajectories. Hand-off to C2 March 2027 for full development.
The paper provides the empirical surface, not the verdict. Sovereign-cloud regulation is not adjudicated. The 2029 AI-training-ramp trajectory is a T2 modelled scenario. Sovereign-cloud and edge-compute footprints outside the hyperscale visible-capacity register are not on-surface. Contested cyber-attribution elements are flagged in the text and not treated as settled.
Five triggers commit the paper to refresh within four weeks: (a) major EU AI Act implementation milestone; (b) CRA Notified Body designation milestone reached or missed (11 December 2026 target near-term binding); (c) G-SICI designation adopted at EU or national scale; (d) new hyperscaler siting decision materially altering §4.2.5 concentration reading or the §4.2.6 trajectory to 2029; (e) new cyber-cascade event within the sovereignty-traceable class of §4.3.5.
The paper opened with two vignettes on two substrates: an Ahr valley in July 2021, an Azure identity plane in May-July 2023. Physical, logical. Same coupling logic. Same modifier surface.
The three arguments. First, that the grid already prices concentrated compute at per-substation granularity. Second, that the EU regulatory stack — the AI Act as the load-bearing workload instrument, the CRA as its product-side co-anchor, NIS2 and DORA as the operator-side floor — constitutes for the first time a coordinated compute-supply-chain governance surface. Third, that fourteen publicly-attributed cyber incidents from 2022 to 2026 — anchored on the March 2026 European Commission Europa.eu AWS breach, the December 2025 Polish grid ELECTRUM campaign, the September 2025 Collins Aerospace MUSE outage across five EU airports, and the DORA aggregate 3,383 major ICT incidents recorded in 2025 — demonstrate the sovereignty-traceable cascade mechanism is not hypothetical and is not receding.
The meta-argument — four faces of one surface. The paper's brief names four convergent facets: grid-resilience, compute-infrastructure, civic-equity, and climate-adaptation. The paper's finding is that these are four projections of a single per-substation analytical surface, not four separate policy questions. Grid-resilience is the R4 system-loading and R6a network-coupling face of that surface. Compute-infrastructure is the HCR hyperscaler-concentration and SICI systemic-institution face. Civic-equity is the R10 recognition-justice face — Frankfurt-Ahrtal, Dublin-rural, Boden-Sami. Climate-adaptation is the R9 compound-hazard face — F-01 heatwave derating traces, F-02 flood-cascade coupling, the 2028 Danube-basin multi-stressor scenario. Each face is measurable at per-substation granularity from the SSI Index v4.2 canonical. Each face reads a different projection of the same modifier surface. None of the four can be resolved by policy that touches only one face.
Reader inference (Rule L). The paper does not say concentrated compute is wrong. It does not say distributed compute is right. It reads the modifier surface, documents the four faces, and flags where the anti-maladaptation mesh signals current outcomes fail specific gates: W3 for systemic substitutability, W6 for data sovereignty, W7 for local-value retention, W10 for justice trajectory. The direction of policy travel is the reader's inference — from a compass that now reads all four faces at once.
The AI Act is two months live. CRA Article 14 is two weeks live. One Dutch TSO is named in the evidence base. Six jurisdictions have the reading available now. Four faces of one surface. The surface reads differently at different granularities; the per-substation reading, now available at scale across 796,121 substations in 39 jurisdictions, is a new measurement, not a new argument. This is the 4IR civil-society surface the brief named — the surface on which the Fourth Industrial Revolution's compute substrate meets the grids, communities, climate exposures, and democratic institutions that will carry it. The compass is on. What direction to walk is not the compass's question.
Full T1 canonical register per jurisdiction covered in §3.2 above · full source URLs and SHA-256 retrieval hashes in accompanying manifest.json.
| Instrument | Citation | EIF | Full applicability | CELEX |
|---|---|---|---|---|
| CRA | Reg (EU) 2024/2847 · 23 Oct 2024 | 10 Dec 2024 | 11 Dec 2027 · Art. 14 live 11 Sep 2026 | 32024R2847 |
| AI Act | Reg (EU) 2024/1689 · 13 Jun 2024 | 1 Aug 2024 | 2 Aug 2025 (Ch. V GPAI) · 2 Aug 2026 (Art. 6 high-risk + Art. 50 transparency) · 2 Aug 2027 (full compliance) | 32024R1689 |
| NIS2 | Dir (EU) 2022/2555 · 14 Dec 2022 | 16 Jan 2023 | 17 Oct 2024 (transposition) | 32022L2555 |
| CER | Dir (EU) 2022/2557 · 14 Dec 2022 | 16 Jan 2023 | 17 Oct 2024 (transposition) | 32022L2557 |
| DORA | Reg (EU) 2022/2554 · 14 Dec 2022 | 16 Jan 2023 | 17 Jan 2025 | 32022R2554 |
| Data Act | Reg (EU) 2023/2854 · 13 Dec 2023 | 11 Jan 2024 | 12 Sep 2025 | 32023R2854 |
Full 14-incident table rendered in Figure 13. Anchor cases: Storm-0558 (2023) · Viasat KA-SAT (2022) · MOVEit / Cl0p (2023) · CrowdStrike Falcon (2024) · European Commission Europa.eu AWS breach (Mar 2026) · Polish grid ELECTRUM (Dec 2025) · Collins Aerospace MUSE (Sep 2025) · DORA aggregate 2025 (3,383 major ICT incidents). Machine-readable version in data/cyber_incident_evidence_base.csv.
Five candidate conventions introduced in this paper, all at 1/5 empirical instances per the SSI Systemic Layer convention-promotion discipline: CS.01 Compute-load-superposition · CS.02 Hyperscaler-concentration ratio (HCR) · CS.03 Grid-headroom-normalised-MW · CS.04 SICI-designation-threshold · CS.05 CRA-Art-14-as-cascade-reporting-instrument. Promotion path requires 5 instances; accretion expected via B2 (Feb 2027), C2 (Mar 2027), D1 (Apr 2027), FA-01 (Jan 2027 Flagship).
Author. SSI Index editorial team · Ikenga SL (Madrid) · with academic peer-review anchor at Information Processing and Telecommunications Centre (IPTC), Universidad Politécnica de Madrid (Rubén San Segundo Hernández; GING research group · Prof. Pedro Reviriego). Contact: ssi_index@ikenga.eu.
Disclaimer. This Themed Analysis is published under the SSI Index Systemic Layer publication programme. It expresses the analytical reading of the per-substation SSI v4.2 modifier surface and does not constitute investment, legal, or regulatory advice. T2-modelled scenarios (notably the 2029 AI-training-ramp trajectory) are lenses for scenario analysis, not predictions. Attributions on cited cyber incidents reflect the public forensic record as of October 2026 and are flagged in-text where contested. Readers citing this document should refer to the methodology pins and the sidecar manifest.json for reproducibility.
Publisher. Ikenga SL (Madrid) · Fondazione SSI Index (Naples · in establishment) · with acknowledgement to The Energy Bridge (TEB Energy Circle GmbH · Vienna) for CEE technical review and to V. Oršulić for substantive review of the pre-publication draft (Oct 2026).
Licence. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0/
Methodology pins. SSI Index v4.2 · JIPR v16 (doi:10.1186/s43065-026-00193-z) · Environmental Research: Energy (doi:10.1088/2753-3751/ae87a5) · SSI Systemic Layer B1 v1.0.1 · Reckien et al. 2023.
Regulatory citations. CRA (32024R2847) · AI Act (32024R1689) · NIS2 (32022L2555) · CER (32022L2557) · DORA (32022R2554) · Data Act (32023R2854).
Right of reply. Signed methodology critique with right of reply invited · publish alongside via ikenga.eu/reports · contact ssi_index@ikenga.eu.
Ship metadata. X1 · Themed Analysis · October 2026 · draft v0.1 · git SHA sha256:fd13a1646570 · sidecar files: manifest.json · methodology_pins.md · REFRESH.md · X1_CONTEXT.md. Zenodo v1.0.0 deposit at ship.