How Compute Institute scores are built
Compute Institute publishes two ratings: the Compute Institute Jurisdiction Competitiveness Index (JCI) — a 0–100 score of how favorable a jurisdiction is for siting a ≥100 MW AI compute campus on the evidence currently in force — and the Project Risk Score (PRS), a 0–100 health score for whether a tracked project will deliver its capacity roughly when and as announced. Every weight, anchor, tier boundary, and override rule on this page is the disclosed configuration the scoring engines actually run. Transparency is the moat: a methodology a journalist can cite must be one a critic can audit.
Design principles
Both methodologies follow practices modeled on Moody’s (disclosed scorecard weights, bounded notching), MSCI (public methodology books, versioned changelogs), and the EIU Democracy Index (honest framing of expert-judgment composites):
- Every datapoint traces to a source. Each input carries a citation, retrieval date, and an evidence-quality grade (E1–E5).
- Enacted ≠ announced. Base scores reflect only what is legally in force or contractually binding. Announcements, draft bills, and op-eds move the Outlook, never the base score.
- Disclosed weights, absolute anchors. All factor and sub-factor weights are published below. Anchors are absolute benchmarks, not cohort-relative, so scores stay comparable as coverage expands.
- Bounded judgment. Analyst judgment is allowed only as explicit, documented notching of at most ±1 point per sub-factor, with written rationale. No silent overrides.
- Versioned methodology. Changes ship as new versions with a public changelog; material changes trigger a full historical re-score.
- No overclaiming. v0 scores are rubric outputs, not statistical predictions.
- Subscriber-pays, never rated-entity-pays. We accept no payment from scored jurisdictions or project sponsors, avoiding the issuer-pays conflict of credit rating agencies.
The Evidence Ladder (E1–E5)
Every input is graded E1 (strongest) to E5 (weakest). The gating rule is strict and mechanical: only E1 and E2 evidence can move a base score. E3 and E4 evidence — announcements, draft bills, press releases — affects the Outlook only. E5 items go to the watchlist and trigger a verification task. In short: announcements move Outlook; only enacted law or executed contracts move scores.
|
Grade
|
Definition
|
Examples
|
Affects
|
|---|---|---|---|
| In-force law, regulation, or executed contract | Statute in effect; signed interconnection agreement; closed financing; recorded deed or permit | Base score | |
| Passed/approved but not yet in force, or funded official program | Law passed awaiting effective date; approved tariff; appropriated incentive fund | Base score (flagged) | |
| Formal proposal in an official process | Draft bill introduced; filed permit application; docketed rate case; LOI disclosed in a filing | Outlook only | |
| Official announcement outside a binding process | Government press release, executive op-ed, sponsor press release, earnings-call statement | Outlook only | |
| Third-party reporting / unverified | Media reports, analyst notes, local press | Watchlist — triggers a verification task |
Each scored entity also carries a Data Confidence grade: A (≥80% of sub-factors evidenced at E1–E2), B (50%–79%), C (<50% — published with an explicit caveat or suppressed). Confidence is displayed next to every score.
Part A — Jurisdiction Competitiveness Index (JCI)
Question answered: if you were siting a ≥100 MW AI compute campus today, how favorable is this jurisdiction, on the evidence currently in force? The unit of analysis is the jurisdiction at the level where siting decisions actually differ — in the US that is the state/grid-operator level, so the US is split. Each sub-factor is scored 0–10 against published absolute anchors using E1–E2 evidence; a dimension is the weighted average of its sub-factors; and JCI = Σ(dimension × weight) × 10, reported on a 0–100 scale with a Data Confidence grade.
Dimensions and weights
|
Dimension
|
Weight
|
Sub-factors (weight within dimension)
|
|---|---|---|
| Power Availability & Deliverability | 20% |
|
| Power Cost | 12.5% |
|
| Speed to Build | 17.5% |
|
| Regulatory & AI Policy Environment | 12.5% |
|
| Fiscal & Incentives | 10% |
|
| Capital & Ecosystem Depth | 12.5% |
|
| Stability & Execution Risk | 15% |
|
Dimension weights sum to 100%; sub-factor weights sum to 100% within each dimension. Weights are disclosed hypotheses, set by analyst judgment informed by investor discovery interviews and published siting literature. Power availability carries the largest weight (20%) because the 2026 binding constraint is deliverability — deals die on energization dates, not tariffs. A published sensitivity table will show each jurisdiction’s rank under ±5pp weight perturbations, so readers can see which rankings are weight-fragile.
Absolute anchor examples
Anchors are absolute benchmarks, not cohort-relative. Published examples (full anchor tables ship in the methodology book appendix):
- Time-to-power — Power Availability & Deliverability
-
- 10 = Firm power offer <18 months typical
- 5 = 3-4 years
- 0 = >6 years or de facto moratorium on new large loads
- Industrial electricity price — Power Cost
-
- 10 = Sustained <$45/MWh all-in with contractable 10-yr terms
- 0 = >$140/MWh or non-contractable
- Permitting regime — Speed to Build
-
- 10 = Credible 18-month greenfield shell
- 0 = >5 years typical or judicially blocked
- AI-specific law in force — Regulatory & AI Policy Environment
- No AI law in force is NOT automatically a 10; absence of law scores 5-7 depending on predictability. An enacted favorable regime scores higher than a vacuum.
Score construction and Outlook
- Score each sub-factor 0–10 against published absolute anchors using E1–E2 evidence.
- Analyst notching: max ±1 per sub-factor, written rationale required, disclosed in the jurisdiction profile.
- Dimension score = weighted sub-factor average; JCI = Σ(dimension × weight) × 10, reported with a Data Confidence grade.
- Tiers assigned from the boundaries below (lower bound inclusive).
- Outlook (Positive / Stable / Negative / Developing) is driven by E3–E4 evidence and pending processes: each pending signal’s implied JCI-point delta is computed against the current base, and a net delta of ≥ +0.5 points is Positive, ≤ −0.5 Negative, otherwise Stable. Outlook can change any time; base scores update quarterly plus event-driven re-scores after a qualifying E1/E2 event.
- Every published score links to its evidence register: source, date, E-grade, excerpt.
Tier boundaries
|
JCI range
|
Tier
|
|---|---|
| 75–100 | |
| 60–74 | |
| 45–59 | |
| 30–44 | |
| 0–29 |
We use tiers rather than letter grades deliberately, to avoid confusion with NRSRO credit ratings.
Part B — Project Risk Score (PRS) (prs-v0.1)
Question answered: will this project deliver its capacity roughly when and as announced? Every tracked project receives four dimension scores — Schedule, Regulatory, Infrastructure, and Capital health, each 0–100 from observable, source-graded factors — and an Overall Project Health Score (0–100, 100 = healthiest) with a Data Confidence grade and an evidence register. Health bands: 80–100 On Track · 60–79 Watch · 40–59 At Risk · 20–39 Distressed · <20 Stalled / Likely Cancelled.
Stage-dependent weights
A project’s binding risks change as it matures, so dimension weights vary by lifecycle stage: announced projects mostly die of money and permits; construction projects mostly die of schedule and grid. Overall Health = Σ(dimension health × stage weight).
|
Stage
|
Schedule
|
Regulatory
|
Infrastructure
|
Capital
|
|---|---|---|---|---|
| Announced / pre-development | 10% | 25% | 30% | 35% |
| Permitted / pre-construction | 20% | 15% | 35% | 30% |
| Under construction | 35% | 10% | 30% | 25% |
| Energization / commissioning | 40% | 5% | 40% | 15% |
Hard caps (override rules, fully disclosed)
Certain single facts cap the overall score regardless of weighted arithmetic — caps prevent averaging from hiding fatal flaws (the “watermelon dashboard” failure mode):
- Critical permit denied → Health ≤ 35
- Sponsor insolvency event → Health ≤ 25
- No interconnection application on file 18+ months after announcement → Health ≤ 55
- Project status “cancelled” → Health ≤ 10
- Project status “stalled” → Health ≤ 35
Updates, staleness, and corrections
- Event-driven recompute: any new E1–E2 datapoint recomputes the affected factors; each project page shows a score history with the triggering evidence.
- Staleness decay — silence is information: when a dimension’s newest evidence exceeds its freshness window (Schedule 6 mo, Regulatory 9 mo, Infrastructure 9 mo, Capital 12 mo), its score drifts 25 points toward the stage-default prior of 50 and its Data Confidence degrades.
- No retroactive smoothing: published score history is immutable; corrections are issued as dated corrections, never silent edits.
The two methodologies link: JCI regulatory and power dimensions provide priors for project-level Regulatory and Infrastructure health in that jurisdiction (project evidence always overrides the prior), and aggregated project outcomes — observed construction velocity, queue times — feed back into the JCI’s speed-to-build and power-availability sub-factors.
Calibration honesty
What these scores are: structured expert-system assessments. Each score is the output of a published rubric applied to cited evidence. The rubric encodes analyst judgment about which observable facts matter and how much.
What these scores are not: statistical predictions. As of v0 we have no outcome history — no set of projects we scored that have since succeeded, slipped, or died — so we cannot state that “Health 50” corresponds to any particular delay or cancellation probability, and we do not.
Path to calibration: every score, sub-score, and evidence item is archived at issuance. As outcomes accumulate (target: ≥100 resolved milestone predictions and ≥30 terminal outcomes), we will publish calibration analyses (score deciles vs. realized slippage/cancellation), re-derive factor weights empirically, and version the methodology (v1 = first empirically informed weights). Until then, treat Health bands as ordinal rankings within stage cohorts, not probabilities.
This framing is deliberately conservative. The fastest way to lose credibility as the ratings layer for AI infrastructure is a debunked probability claim in year one — so we label judgment as judgment.
Conflicts & Transparency
The ratings firewall
Compute Institute sells benchmarking and advisory engagements to governments (the Jurisdiction Ratings Program). Those engagements are walled off from the public index by a non-negotiable policy:
Engagement revenue cannot influence scores. The methodology is the only path to a score change.
A government engagement can never touch a public index score — not its inputs, not its weights, not its timing. The only way any jurisdiction’s score moves is the way every jurisdiction’s score moves: new E1/E2 evidence processed through the published rubric above. Score history is append-only with frozen input snapshots, so anyone can audit that scores never drifted around engagement dates.
Public government-engagement registry
Every government engagement is disclosed here — jurisdiction, scope, and dates — displayed next to that jurisdiction’s full score history. Audit us.
No government engagements to date.
Methodology changelog
- v0 — 2026-06-11
- Initial publication: JCI v0 (7 dimensions, disclosed weights, Tiers 1–5) and PRS prs-v0.1 (4 dimensions, stage-dependent weights, hard caps), gated by the E1–E5 Evidence Ladder.
Material changes (weight shifts >5pp, new or removed dimensions) require a published consultation note and trigger a full historical re-score so time series stay consistent.