Methodology — Kardashev Index

Categories and weights weights-v1

Category	Type	Weight	What it measures
Energy throughput	Measured	30%	Energy the company reports consuming in its own operations, or generating at plants it owns or operates, per year. The larger of the two is used. Energy storage deployed or shipped (e.g. battery GWh) and energy sold or resold to customers are recorded but never scored: they are not energy the company itself used or produced. A figure that is one of several unlabelled values under one heading (e.g. "Energy Consumption (kWh)" followed by three numbers whose row labels are icons) is recorded as ambiguous and not scored or summed, unless the source itself states a total.
Compute capacity	Measured	20%	Data-center capacity the company reports operating (IT or facility power, MW). Announced, under-construction or under-development capacity (including a figure whose footnote or table row says so) is shown but not scored, and so is operating capacity that contradicts the company's own reported energy use (over 10x it even at 20% utilisation).
Growth gradient	Measured	15%	How fast the company is building: capital-expenditure trend (from SEC filings when available), revenue trend, and reported energy-use growth.
Frontier acceleration	Opinion	15%	Contribution to pushing the technological frontier: new capabilities shipped, cost-per-capability reductions, open releases others build on.
Builder velocity	Opinion	12%	How quickly the company turns plans into physical or shipped reality: cadence, first-of-a-kind deployments, time from announcement to operation.
Permission to build	Opinion	8%	Public positions and actions on permitting, energy build-out, open source and regulation. This is explicitly an editorial e/acc lens and carries the lowest weight.

Measured categories: normalization

Energy and compute span many orders of magnitude, so they are scored on a log scale: every 10× is +2.5 points. Growth is mapped through fixed anchors. Anchors are fixed, not curved to the current leaderboard, so a score means the same thing next year.

Joules · 30%

Energy throughput

Inputs. Reported annual energy consumption, electricity consumption or own generation (MWh, GWh, TWh, GJ, TJ, PJ, MMBtu) with a verbatim quote from a fetched source.

P = annual energy (J) / seconds per year → average watts. Score = clamp(2.5 × (log₁₀P − 7), 0, 10). Kardashev-equivalent K = (log₁₀P − 6) / 10.

Reported	Score
10 MW avg (≈ 88 GWh/yr)	0
100 MW (≈ 0.88 TWh/yr)	2.5
1 GW (≈ 8.8 TWh/yr)	5
10 GW (≈ 88 TWh/yr)	7.5
100 GW (≈ 876 TWh/yr)	10

Watts of thought · 20%

Compute capacity

Inputs. Reported operating data-center capacity in MW or GW with a verbatim quote.

Score = clamp(2.5 × log₁₀(MW), 0, 10).

Reported	Score
1 MW	0
10 MW	2.5
100 MW	5
1 GW	7.5
10 GW	10

Slope · 15%

Growth gradient

Inputs. Annual capex and revenue: SEC XBRL company facts (API URL + accession number) first, else the cash-flow line 'purchases of property and equipment' or company-reported capex quoted from a filing, IR page or annual report. Bonds, funding rounds, deal sizes, planned spend and headlines are rejected. Energy series as above.

Each available sub-metric's compound annual growth rate (up to 3 years) is mapped piecewise-linearly through the anchors, then combined with weights capex 50%, revenue 25%, energy 25% (renormalized over what is available).

Reported	Score
−30%/yr or worse	0
0%/yr	3
+10%/yr	5
+25%/yr	7
+50%/yr	9
+100%/yr or more	10

Headline

Kardashev-equivalent

We convert reported annual energy to average power P in watts and apply Carl Sagan’s interpolation K = (log₁₀P − 6) / 10. Humanity today is about K 0.73; a 1 GW company is K 0.30. K is shown as the headline; the Energy score is the same quantity on a 0–10 scale.

Energy that is only enabled (batteries or panels shipped to others) is not counted, so that the same joules are not credited to several companies.

Opinion categories: rubrics rubrics-v1

These are an explicitly e/acc editorial lens. The model gets the rubric and a numbered list of verified quotes; each score must cite at least one of them, and if the quotes don’t support a score it must return “insufficient evidence” instead of a middle value. Confidence is capped at 85% and lowered when fewer than three quotes support a score.

Opinion · 15%

Frontier acceleration

Contribution to pushing the technological frontier: new capabilities shipped, cost-per-capability reductions, open releases others build on.

0No evidence of contributing new capability; follows others or is outside technology entirely.
2Incremental improvements to existing products; little that others can build on.
4Competitive, current products in a frontier field, but rarely first; modest cost or capability gains.
6Regularly ships capabilities that are near state of the art, or materially lowers cost per capability.
8Repeatedly first or best on important capability measures, or releases that a wide ecosystem builds on.
10Defines the frontier of its field: step-change capabilities or cost collapses that reset the industry.

Opinion · 12%

Builder velocity

How quickly the company turns plans into physical or shipped reality: cadence, first-of-a-kind deployments, time from announcement to operation.

0No evidence of shipping or building anything new in the period.
2Slow cadence; announcements routinely slip by years or are abandoned.
4Steady, ordinary cadence for its industry.
6Faster than industry norms; several meaningful launches or facilities brought online on schedule.
8Exceptional speed: first-of-a-kind deployments or record build times, repeatedly.
10Sets the global benchmark for speed at scale; compresses multi-year industry timelines into months.

Opinion · 8%

Permission to build

Public positions and actions on permitting, energy build-out, open source and regulation. This is explicitly an editorial e/acc lens and carries the lowest weight.

0Actively lobbies to restrict building, energy supply or open technology for others.
2Mostly supports restrictive positions; occasional pro-build statements.
4Neutral or mixed public record; little substantive advocacy either way.
6Generally supports permitting reform, energy build-out or open technology in public positions.
8Consistent, substantive advocacy and action for building, abundant energy and open technology.
10Leading, sustained advocate whose actions measurably expanded others' permission to build.

How a run works

01Resolve. Pin down the official name, website, ticker and whether it’s public (model with web search).
02Research. Find primary documents: sustainability and ESG reports, CDP responses, annual reports, filings, official announcements (model with web search; at most 8 searches).
03SEC EDGAR. For US-listed filers, annual capital expenditure and revenue come straight from XBRL filings — no model involved.
04Fetch. We download every proposed source ourselves. Dead links, pages that redirect to a homepage, pages that read as “not found”, and pages that look like what the site returns for a random made-up path (a soft-404 probe) are rejected. We store a hash and a text snapshot of what we read.
05Extract. The model copies figures (value, unit, period) and claims, each with a quote. We keep a quote only if it appears verbatim in the fetched text, the number appears in the quote, the unit appears next to it, and the magnitude is physically plausible.
06Compute. Units, average power, K, growth rates and the measured scores are computed in Python. If sources disagree for the same year, primary sources win and confidence drops.
07Judge. The three opinion categories are scored against the rubrics from verified quotes only.
08Aggregate and publish. Weighted in code. A run only replaces the published one if it succeeds — and a run below the ranking thresholds never replaces better data: a ranked run, an unranked run that scored more of the weight, or the legacy v0 assessment. Such runs are kept (admins see them; the page notes them) but not published. Two exceptions: a complete, non-degraded run on a newer pipeline version replaces a run measured with older rules even if it is unranked (it is then shown as Unranked with its reason), and an editor can retract a published run that is known to be wrong — it is never shown again, the most recent remaining run becomes current, and the page carries a dated correction note with the reason. Every run is kept, with timing, tokens and cost per stage.

Confidence, coverage and ranking

ΣIndex = Σ wᵢ·sᵢ / Σ wᵢ over the categories that have a score. Missing categories are left out, never filled with a default.
%Coverage = the share of total weight that was scored. Confidence = Σ wᵢ·confidenceᵢ over all categories, so missing data lowers confidence directly.
≥Ranked only when all of these hold: (1) coverage: at least 60% of the weight is scored and at least 30% of the total weight is measured — or the “no energy figure found” rule below is met; (2) measured share: more than 40% of the scored weight comes from measured categories; (3) confidence: more than 20% (confidence runs 0–100%, as shown on every page; after the “no energy figure found” reduction). Otherwise the entity is listed as “not ranked yet” with the reason. Rules (2) and (3) apply to every published run when the page is shown, including runs made before they were introduced (pipeline-v2.9); stored results are not rewritten.
ENo energy figure found: many companies never publish an energy figure. When energy throughput (30% of the weight) has no verified figure after reading at least 3 sources, and none of the energy-related sources we found failed to load, the entity may still be ranked if the scored share of the remaining 70% reaches the same 60% threshold with measured data present. It is shown with a “no energy figure found” flag and its confidence is multiplied by 0.8. It must still pass the measured-share and confidence rules above. Nothing is imputed for the missing energy.
EEnergy source couldn’t be read: if we found the company’s own energy disclosure (its sustainability or impact report or page, ESG data, a CDP response…) but could not read it — too large, unparseable, blocked (HTTP 401/403/429), a server error or a timeout — the entity is not ranked rather than scored without energy, admins are alerted and one automatic retry is scheduled 24 hours later. Dead links and “page not found” responses don’t count: those sources don’t exist. Neither do third-party articles about a report, filing indexes, press releases or news pages (unless a press release is itself about energy consumption): what matters is whether the company’s own disclosure exists and was unreadable.
cMeasured confidence: SEC filings 95%, the company’s own documents 90%, third-party sources 65%; reduced for figures older than three years, for electricity-only figures (fuels missing) and when sources conflict.

Metric definitions

Every figure must be quoted verbatim from a page we fetched and is checked in code: the quote must appear in the fetched text, the number in the quote must match (tables with “in millions” headers, parenthesized numbers and dot leaders are handled), and the wording must fit the definition. Figures that fail are kept out with the reason recorded.

1Energy consumed: Total energy the company itself CONSUMED in its own operations in a period (electricity + fuels; MWh/GWh/TWh/GJ/TJ). Scored.
2Electricity consumed: Electricity the company itself consumed/purchased for its own operations (only when total energy is not given). Scored.
3Energy generated (own plants): Energy the company GENERATED at plants it owns or operates (utilities, on-site solar). Scored as throughput alongside consumption.
4Energy sold / delivered (not scored): Energy sold or delivered to customers (retail/wholesale), including resold energy. Recorded, not scored.
5Energy storage deployed (not scored): Battery/energy-storage capacity deployed, installed or shipped (e.g. 'deployed 46.7 GWh of energy storage'). NOT energy consumed or generated. Recorded, not scored.
6Data-center capacity (operating): Data-center power capacity in operation today (MW/GW). Scored.
7Data-center capacity (planned): Announced, contracted or under-construction capacity (MW/GW). Recorded.
8Capital expenditure: Reported capital expenditure for a completed fiscal period: the cash-flow line 'purchases of property and equipment' or company-reported capex, from filings, IR or annual reports. Never bond/debt offerings, funding rounds, deal or order sizes, planned or forecast spend, or a number from a news headline.
9Revenue: Reported revenue for a completed fiscal period. Never forecasts, run-rates or valuations.

Known limits

!Companies that don’t publish energy or capacity figures score lower confidence and may not be ranked. Most private companies fall here.
!Compute counts capacity a company operates. A chipmaker whose hardware runs in other companies’ data centers is not credited with those watts.
!Research depends on what web search surfaces in a limited number of searches; a missed report means a missing figure, not a low one.
!Opinion categories are editorial. They carry 35% of the weight, are marked as opinion everywhere, and cite their quotes.
!Entities without a measured run may show their legacy v0 scores (a single model prompt, superseded), labelled as such and never ranked.

Disclosures

AIAI-assisted. Resolve, research, extraction and the three opinion categories use xAI’s Grok model (grok-4.3). Its output is validated in code as described above; individual scores are not hand-reviewed before publication.
≠Conflict of interest. xAI, which makes the scoring model, is also ranked on this index. Its figures go through the same verification and its opinion scores through the same rubrics as everyone else’s; sources and rationales are shown on every entity page. More.
%Measured vs opinion share. The nominal split is 65% measured / 35% opinion. Because unscored categories are left out, each entity’s actual measured share of the scored weight is shown next to its Index on the leaderboard and its page.
$Not investment advice. The Index is an experimental assessment, not a recommendation about any security. Errors can be reported via the corrections contact.

Mostly measured. Opinion, labelled.