The four numbers behind emissions: the clean-energy story is about the term that didn't move

Reachability swept first: yes — every slug fetched and parsed before any number below was written. See Reachability under Open questions. Vintage checked: yes — lastUpdated / nextUpdate read per series from OWID's own metadata endpoint and recorded in the table.

Segment: The four numbers behind emissions

• Source theme(s): Batch 54 #2 (cross-refs Batch 54 #1, #13)
• Spine role: main story
• The framing people carry: that two decades of renewables have been cleaning up the world's energy, and emissions have kept rising anyway because there are more of us wanting more.
• Hypothesis (from the bank): the two efficiency terms improve "steadily and almost monotonically for decades", while the two scale terms outrun them — "the entire arithmetic of the failure to peak".
• Verdict: QUALIFIED. The scale-outruns-efficiency half is confirmed. The "two efficiency terms improving steadily and almost monotonically" half is refuted at world level: only ONE of them did. Energy intensity fell 24.9% over 2000–2021. Carbon intensity of energy fell 1.5% — and rose to a peak in 2012 before falling. Essentially all the restraint came from using less energy per dollar, almost none from making the energy cleaner.

The numbers (every row sourced — no exceptions)

The texture (what narrative flattens)

• The clean-energy story is a story about the wrong term. Renewables enter the identity through carbon intensity, and that term is the one that did not move. What moved was energy intensity — efficiency, structural shift away from heavy industry, and services growing faster than smokestacks.
• It got worse before it got better. Carbon intensity rose from 240.7 in 2000 to 247.9 in 2012. The coal build-out of the 2000s outran renewables deployment for over a decade, and the world's energy was dirtier in 2012 than when the century began.
• The recovery is recent and small. 2012 → 2024 brings it to 234.7, which is 5% below 1990 after thirty-four years.
• The identity is exact in algebra and not in these series. On LEVELS the four terms overshoot published CO₂ by a stable ~25% — see Open questions.

Datoid artifact (Tier-A only)

• Already ingested: annual-co2-emissions-per-country and population (prisma/seed.ts), plus gdp-per-capita-worldbank per country via seed-country-matrix.ts — where its World row was created as gdp-per-capita-worldbank@World, so there is no bare-slug World row for it. That script's existingWorldSlug mechanism applies to other indicators, not this one. Look up what exists before adding anything.
• NO connector work — verified by RUNNING it, after a draft of this sheet claimed the opposite. The correct externalId is the plain slug co2-per-unit-energy, which 301s to energy-mix.csv?metric=carbon_intensity&source=total. fetch follows the redirect, OWID merges the query parameters, and the redirected CSV's single data column sits at index 3 where parseOwidCsv reads it. fetchOwidPoints("co2-per-unit-energy","World") returns 77 points with 2000 = 240.71663, 2012 = 247.93112, 2021 = 236.99687 — this sheet's own numbers. The earlier claim was reasoned from the URL-building string and not executed, which is the one check this repo says cannot be substituted for.
• Stored magnitudes are NOT the raw CSV magnitudes: VALUE_DIVISORS scales population and annual-co2-emissions-per-country by 1e9. Every figure in this sheet is computed from the raw CSVs, so anything recomputed on-platform must account for that.
• Could be built: a four-panel decomposition dashboard, with the carbon-intensity panel as the reveal — after the connector question is settled.

Story justification (decide here — do NOT defer)

• Does the argument have a reveal that needs sequencing? YES — BUILD.
• Because: the punchline requires withholding one series. The reader must first accept that emissions-per-dollar has fallen a long way (true, and the familiar good-news framing), and only then see that the carbon content of energy is flat — and rose for twelve years. Shown together in one view the contrast is a legend-reading exercise; shown in sequence it is a reversal. This is the same shape as working-hours, where the second dataset is the argument.

The questions this raises (and whether we can answer them)

Stage 1 rule 6, added 2026-09-23. Everything above verifies a hypothesis; this names what a READER asks next and forces the decision before Stage 2 runs — fetch it, attribute it, or declare it open. Every slug named here was swept for reachability first.

1. If renewables grew so fast, why did the carbon intensity of energy barely move? — FETCHABLE, and largely already ingested. The electricity-mix set (seeded 2026-09-23) has world FOSSIL generation +3,628 TWh over 2010–23 against +3,605 TWh of wind and solar: new clean supply met new demand instead of displacing the dirty stock, which is why the ratio of the two stayed at 1.01. The pairing matters — a first draft set coal alone (+1,938) against all wind and solar, which makes clean look like it outran the dirty stock and argues the opposite of the finding. Not the whole answer either: carbon intensity covers all energy, not electricity alone.
2. Did energy intensity fall because economies got efficient, or because rich countries offshored heavy industry? — FETCHABLE, but NOT with the obvious test. Energy intensity is measured on production inside a country, so offshoring lowers the sender's and raises the receiver's; China's intensity falling is consistent with efficiency AND with absorbing heavy industry, so it discriminates nothing. A first draft proposed exactly that non-sequitur and called it the least-tested claim available — which it was, for the wrong reason. The series that does discriminate is consumption-based accounting: consumption-co2-per-capita (swept 200, consumption_emissions_per_capita) against production-based co2pc-<country>. If a rich country's consumption-based figure fell as far as its production-based one, the industry did not simply move. It is CO₂ rather than energy, so it is a proxy — say so. Do not assert offshoring without it. THE TWO COUNTRY SERIES END IN DIFFERENT YEARS [verified 2026-09-23, local] — production-based runs to 2024, consumption-based to 2023, for all seven pairs. Quote a gap on a COMMON year or name both years; a spoken comparison that drops them reads a one-year difference into the gap. (The WORLD rows run to 2024 on both, which is a different fact and not a counter-example — world consumption and world production are the same numbers by construction, which is why no world row is seeded here.) (energy-intensity-of-economies stops at 2021, which is why this sheet's headline window does too.)
3. Which countries have actually decoupled — emissions falling while GDP grows? — FETCHABLE, and the corpus already holds both halves per country from the #3 country matrix: co2pc-<country> (20 entities) and gdppc-<country>. Note the bare co2-per-capita is the WORLD row, not a family prefix — a draft here wrote co2-per-capita-*, which resolves to nothing while a bare lookup half-succeeds on the world aggregate.
4. How fast would carbon intensity have to fall to hold emissions flat? — DERIVABLE, not fetchable. It is arithmetic on the identity, and it becomes a projection the moment it runs past the data. State the arithmetic; never the forecast.
5. Why do the four terms overshoot published CO₂ by ~25% on levels? — OPEN (method), already under Open questions: cross-producer coverage and unit conventions, not one fixable join.

Open questions / where I'm guessing

• THE LEVEL IDENTITY DOES NOT CLOSE, and this must not become a spoken assertion. Multiplying the four World series overshoots published CO₂ by 1.25× in 2000 and 1.28× in 2021. Two independent routes to world primary energy disagree: the WB-intensity route gives 697 EJ for 2019 and the OWID-CO₂ route 557 EJ, while IEA's world TPES is ~600 EJ — so both bracket the accepted figure and neither matches it. The likeliest cause is that "World" is a different country aggregate in each producer's series; this is not established and should not be asserted as the reason.
• What IS safe is the growth-rate decomposition, which is how Kaya is normally done and where level offsets cancel. Quote the right window: over the headline span 2000–2021 the four terms sum to +39.3 pp against actual +36.8 pp, a −2.5 pp gap or −0.119 pp/yr. It falls to −0.006 pp/yr only when 2020–21 are excluded, and an earlier draft of this sheet quoted that figure beside 2000–2021 headline numbers — a twenty-fold understatement of the residual the headline figures actually carry. Every pp figure in the table is a log change, not a level product.
• Crisis years are the unreliable ones, and they are the interesting ones. 2020 carries a −1.77 pp residual, 2009 −0.97 pp. Any claim about the COVID dip specifically is weaker than the rest of this sheet.
• The window is 2000–2021 and the binding constraint is energy intensity, which OWID carries only for those years. The episode cannot speak to 2022–2024 — which excludes the energy shock following the invasion of Ukraine, the single most-asked question about recent emissions.
• OWID's primary energy uses the substitution method for non-fossil sources while the World Bank's intensity indicator uses direct primary energy. This is a known methodological difference between the two series used here and is a candidate explanation for part of the level gap. Not verified.
• Country-level claims are untested. The bank hypothesised a crossover arriving in different decades in different countries, and recent declines decomposing into carbon-intensity improvement. Nothing here tests either — this sheet is World only.
• Reachability, for the record: population, gdp-per-capita-worldbank, energy-intensity-of-economies and annual-co2-emissions-per-country return
  200. Carbon intensity is reached via co2-per-unit-energy, which 301s to energy-mix.csv?metric=carbon_intensity&source=total. co2-intensity is NOT the fourth term — it is CO₂ per GDP, which silently collapses two terms of the identity into one; an earlier sweep here mistook it for carbon intensity of energy on the strength of its name and a 200.