Gross Intelligence Product / Methodology
One gross index of machine intelligence, built entirely from statistics that already exist
GIP invents no new measurements. It takes credible, published indexes of AI power, AI quality, AI use and AI trade, and combines them with stated rules into one gross figure for each country: what it can produce, and what it actually produces.
By Post-Labor Economics · Sunday, November 1, 2026 · Methodology version 0.2 (draft for comment)
| Component family | What it measures | Existing source(s) | Frequency | Country detail |
|---|---|---|---|---|
| 1. AI power | Electricity actually consumed by AI data centres; installed AI capacity | IEA Key Questions on Energy and AI; Ireland CSO; Netherlands CBS; Norway Elhub; GB DESNZ; LBNL; China NEA; Epoch AI power capacity | Annual; Ireland quarterly; Norway hourly | About 10 countries plus world regions |
| 1a. Compute per unit of power | Computation each unit of electricity buys | Epoch AI hardware trends (FLOP/s per watt) | Continuous | Global |
| 2. AI quality | Capability of the models being run | Artificial Analysis Intelligence Index (sole quality measure) | Continuous; versioned | Global (models, not countries) |
| 3. AI consumption | Who uses AI, and how much | Microsoft AI Diffusion Report; Anthropic Economic Index; published token volumes | Quarterly / monthly | 100+ countries |
| 4. AI exports | Where AI is produced vs used | Epoch AI chip-imports data (customs); WTO digitally delivered services | Quarterly; annual | 16 customs reporters plus inferred destinations; 200+ economies |
IEA AI-focused data centres; ~32% of 485 TWh. Modelled, not metered
Epoch AI, cumulative H100e arrivals 2022 to Q2 2026
Anthropic Economic Index. One provider only
Range ~$154bn-$163bn; ads permanently excluded
- 1. Two views: Capacity and Output
- 2. AI power
- 3. Country power table
- 4. Compute per unit of power
- 5. AI quality: the Artificial Analysis Intelligence Index
- 6. AI consumption
- 7. AI exports
- 8. Combination rules
- 9. Illustrative figures
- 10. Cross-checks: dollars and tokens
- 11. The labor link
- 12. Data sources
- 13. Gaps
- 14. Sources
Headline
1. Two views per country: what it can produce, and what it does produce
A country with a giant data-centre campus but few users exports intelligence. A country full of users but few chips imports it. GIP shows both sides.
| View | Plain meaning | Combination rule |
|---|---|---|
| GIP Capacity | How much intelligence the country could produce at full use | Installed AI capacity × compute per unit of power × AA quality |
| GIP Output (produced) | How much it actually produced | AI electricity consumed × compute per MWh × AA quality |
| GIP Consumed | How much its people and firms used | World GIP Output × the country's share of AI use |
| Net AI exports | Produced minus consumed | GIP Output (produced) − GIP Consumed |
GIP Outputc,t = Ec,t × Ct × Qt
- E = AI electricity actually consumed in country c (MWh), crypto mining excluded. Family 1.
- C = compute per MWh, from Epoch AI's published hardware-efficiency trend. Family 1a.
- Q = capability per unit of compute, from the Artificial Analysis Intelligence Index. Family 2.
- Capacity uses installed capacity (Epoch GW and H100e) in place of E.
Family 1
2. AI power: count power drawn, not capacity built
A gigawatt of announced capacity is not a gigawatt of thinking. Sites wait for chips and grid connections, and inference load rises and falls with demand. So Output uses energy actually consumed, read from meters wherever a statistics office publishes it.
- 1. Metered statistics first. Ireland CSO (quarterly), Norway Elhub (hourly), Netherlands CBS (annual, from grid operators), Great Britain DESNZ (annual, meter-based). Elsewhere, IEA modelled estimates, labelled as such.
- 2. AI share of data-centre load. Default: the IEA's AI-focused share of world data-centre electricity, 155 of 485 TWh (~32%) in 2025. Replaced by a country share whenever one is published. This is an assumption, not data.
- 3. Exclude crypto mining. See the exclusion rule below.
- 4. Capacity only in the Capacity view. Epoch AI's power capacity (about 31 GW of facility power at end-2025, partly extrapolated) and chip counts feed GIP Capacity. They enter Output only through a published utilization rate, labelled an estimate.
About a third of the world's data-centre electricity now goes to AI-focused sites
Global data-centre electricity consumption, 2025, TWh. Modelled by the IEA; excludes crypto mining
Bar chart. Use the up and down arrow keys to move through 3 categories. A data table follows.
All data centres: Electricity 485
Other = 485 − 155 (our subtraction). 'AI-focused' means sites with many accelerated servers; it includes some non-AI load inside them and misses AI run in general-purpose sites.
Source: IEA, Key Questions on Energy and AI, Annex Table A.4 (Apr 2026); AI-focused figure as cited by Our World in Data (Sep 2026) Download CSV Methodology
Data table
| site | Electricity |
|---|---|
| All data centres | 485 |
| AI-focused data centres | 155 |
| Other data centres | 330 |
Utilization
Capacity and consumption differ by roughly a factor of two. The published evidence:
| Evidence | Value | What it measures | Source |
|---|---|---|---|
| Data-centre capacity factor, world | 48% (2025); hyperscale 52% | Average load ÷ installed capacity | IEA Key Questions, Table A.1 (Apr 2026) |
| Data-centre load factor | 75%-90% | Average load ÷ peak load | IEA Key Questions, ch. 8 (Apr 2026) |
| Norway, metered | Actual use is about half the sum of each site's peak | Metered load vs measured peaks | Elhub (Sep 2026) |
| AI server power when running | 70% of rated power; H100 training measured at 74% of TDP | Power draw vs nameplate | LBNL (Dec 2024) |
| AI server operating time | Training 80%; inference 40% | Share of the year in use | LBNL (Dec 2024) |
| Epoch capacity basis | Rated chip TDP × 2.5 overhead; assumes all chips at full use | Capacity, not consumption | Epoch AI (Jan 16, 2026) |
International
3. Who meters data-centre power, and who only estimates it
Only a few countries read the meters. Ireland is the gold standard: a quarterly official series back to 2015. The two largest markets, the US and China, have no metered national series.
| Country | Latest data-centre electricity | Year | Share of national electricity | AI share | Source | Quality |
|---|---|---|---|---|---|---|
| World | 485 TWh | 2025 | ~1.5% | ~155 TWh AI-focused (~32%) | IEA Key Questions (Apr 2026) | Modelled estimate |
| United States | 224 TWh (IEA); 176 TWh in 2023 (LBNL) | 2025; 2023 | 4.4% (LBNL, 2023) | Not published; GPU servers >40 TWh in 2023 (LBNL, IT only) | IEA Table A.4; LBNL Dec 2024 | Modelled estimate |
| China | 170 TWh 'computing centres' (NEA); 117 TWh (IEA) | 2025 | 1.6% (NEA) | Not published; 42 clusters of 10,000+ AI chips | National Energy Administration (May 29, 2026); IEA | Government statement, method unpublished |
| Europe | 72 TWh | 2025 | - | Not published | IEA Table A.4 | Modelled estimate |
| Germany | 21.3 TWh (expected) | 2025 | Not given | Not published | Borderstep Institute for Bitkom | Industry-research estimate |
| Ireland | 7,663 GWh | 2025 | 23% of metered | Not published | CSO (Jul 7, 2026) | Official, metered, quarterly |
| Netherlands | 5,090 GWh | 2024 (provisional) | 4.6% | Not published | CBS, from grid operators (Dec 2025) | Official, metered; excludes in-house sites |
| Great Britain | 4.5 TWh | 2024 | 2% of grid consumption | Not published | DESNZ Energy Trends (Jun 30, 2026) | Official, meter-based estimate |
| Norway | 3.14 TWh | 2025 | 2.26% | Not published; includes crypto miners | Elhub (Sep 2026) | Official metered; crypto not separated |
| Denmark | 4.3 PJ ≈ 1.19 TWh | 2025 | ≈3.6% of final electricity (our calc.) | Not published | Danish Energy Agency, Energistatistik 2025, from the EU data-centre database | Official, operator-reported (EU EED) |
| Singapore | 3.40 TWh | 2020 | ~7% | Not published | Ministry reply to parliament | Official, stale |
| Japan, Sweden, Finland | Not found | - | - | - | Searched Nov 1, 2026 | Gap |
Ireland's data centres use six times the power they did in 2015
Metered electricity consumption by data centres, Ireland, GWh per year. Eleven official annual points
Line chart. Use the left and right arrow keys to move through 11 periods. A data table follows.
2025: Data centres 7,663
CSO identifies data centres by operator name and high-consumption customers; they are not a separate class in the source data. No AI split.
Source: Central Statistics Office, Data Centres Metered Electricity Consumption 2025, Jul 7, 2026 Download CSV Methodology
Data table
| year | Data centres |
|---|---|
| 2015 | 1,240 |
| 2016 | 1,482 |
| 2017 | 1,762 |
| 2018 | 2,182 |
| 2019 | 2,490 |
| 2020 | 3,030 |
| 2021 | 4,012 |
| 2022 | 5,273 |
| 2023 | 6,339 |
| 2024 | 6,973 |
| 2025 | 7,663 |
- Official metered: Ireland (CSO), Netherlands (CBS), Norway (Elhub), Great Britain (DESNZ, meter-based). Denmark reports operator data from the EU's data-centre database, required by the Energy Efficiency Directive for sites with significant energy use.
- Estimates only: United States, China, Germany, Europe, world. China's NEA figure (170 TWh) is 45% above the IEA's 117 TWh for the same year; definitions differ and NEA's method is not published.
- No country publishes an AI share of data-centre electricity.
Family 1a
4. Compute per unit of power: Epoch AI's published trend, used directly
Each chip generation does more arithmetic per watt. We do not compute this ourselves. We take Epoch AI's published figures.
| Published series | Value | Uncertainty | Source |
|---|---|---|---|
| GPU FLOP/s per watt, growth since 2008 | 1.34x a year (doubling every 2.4 years) | 90% CI 1.29x-1.39x | Epoch AI Trends (updated Feb 5, 2026) |
| Leading ML hardware energy efficiency | ~40% a year | Not stated | Epoch AI (Oct 23, 2024) |
| Leading AI processors, performance per watt | 30%-40% a year | Range as published | IEA Key Questions, Fig. 2.2 (Apr 2026), based on Epoch AI |
| Stock of AI compute, growth | 3.4x a year | As published | Epoch AI Trends (Feb 5, 2026) |
| Facility overhead (PUE), world | 1.41 (2024) → 1.38 (2025) | Modelled | IEA Table A.1 |
Combination rule: C grows at Epoch's FLOP/s-per-watt rate, 1.34x a year (≈1.076x a quarter), with the 90% interval carried as the uncertainty. The IEA PUE series adjusts facility energy to chip energy.
Family 2
5. AI quality: the Artificial Analysis Intelligence Index, and nothing else
GIP uses one quality yardstick so that nobody, including us, can pick the benchmarks that flatter a result. That yardstick is the Artificial Analysis Intelligence Index.
- What it is: v5.0 combines nine evaluations in four equally weighted categories: knowledge work, coding, science and general (25% each). Several of the sets are private.
- What AA publishes for each model: the Index score, the cost in dollars to run the full Index (using providers' token counts), and the tokens used. It does not publish FLOP.
- So capability per unit of compute is measured as Index score per dollar of cost to run the Index, deflated by Epoch AI's chip price-performance trend (1.49x a year) so that hardware gains already counted in C are not counted twice.
- Two series, reported separately: frontier (best score at the start of each quarter) and median (the model at the median of usage). Most electricity serves non-frontier models, so the median series feeds the headline.
Qt / Qt−1 = (AA score ÷ cost to run)t(AA score ÷ cost to run)t−1 ÷ Epoch FLOP-per-dollar growth
Family 3
6. AI consumption: who actually uses it
Two large, published country datasets now exist. Microsoft counts the share of working-age people who used generative AI. Anthropic publishes each country's share of Claude use.
| Country | Working-age AI users, Q2 2026 (Microsoft) | Share of Claude use, May 2026 (Anthropic) | Claude use per capita index (1.0 = population share) |
|---|---|---|---|
| United States | 33.0% | 20.16% | 3.87 |
| India | 18.5% | 7.12% | 0.30 |
| United Kingdom | 43.7% | 3.47% | 3.35 |
| Japan | 24.7% | 3.28% | 1.91 |
| Germany | 32.3% | 2.98% | 2.40 |
| Netherlands | 43.5% | 1.07% | 3.90 |
| Singapore | 64.3% | 0.62% | 5.81 |
| Norway | 49.4% | 0.35% | 4.04 |
| Denmark | 32.4% | 0.32% | 3.57 |
| Ireland | 49.9% | 0.27% | 3.23 |
| China | 17.5% | Not published | Not published |
| World | 18.8% | 100% | 1.0 |
- Combination rule: a country's consumption share is the average of its share of world AI users (Microsoft rate × working-age population) and its share of Claude use (Anthropic). Two independent sources limit one firm's market-share bias.
- Token volumes (Google, OpenRouter) are global, not by country; they check the total, not the split.
- Microsoft's measure comes from device telemetry adjusted for operating-system share, internet penetration and population. Anthropic's covers Claude only; no China row is published.
Family 4
7. AI exports: where intelligence is made vs where it is used
Nearly every AI chip crosses a border. Epoch AI reads customs records to estimate where chips end up. Set against who uses AI, this shows who exports intelligence.
The US holds 72% of the world's AI chips, but accounts for 20% of Claude use
Share of cumulative AI-chip compute arrivals (H100e), 2022 to Q2 2026, and share of Claude use, May 2026, %
Illustrative
Bar chart. Use the up and down arrow keys to move through 9 categories. A data table follows.
United States: AI chips: Share 72.2%
Shares of 27.4m H100e (our sum of Epoch's country rows). Arrivals are not deployed compute. Claude use is one provider; no China figure. Illustrative indicator of direction, not a GIP figure.
Source: Epoch AI, AI chip imports dataset (CSV updated Oct 27, 2026); Anthropic Economic Index (release Jun 26, 2026) Download CSV Methodology
Data table
| series | Share |
|---|---|
| United States: AI chips | 72.2% |
| United States: Claude use | 20.2% |
| China: AI chips | 9.6% |
| India: AI chips | 1.6% |
| India: Claude use | 7.1% |
| United Kingdom: AI chips | 1.2% |
| United Kingdom: Claude use | 3.5% |
| Japan: AI chips | 1.6% |
| Japan: Claude use | 3.3% |
| Existing source | What it gives | Use in GIP | Limit |
|---|---|---|---|
| Epoch AI chip imports | Quarterly AI-chip arrivals by destination (trade value, units, H100e), net of re-exports, since 2022 | Capacity by country; production location | EU reported as one bloc; arrivals precede deployment |
| WTO digitally delivered services | Exports and imports of services delivered over networks, 200+ economies, 2005-2025 | Context for cross-border AI services | No AI or cloud-only split |
| Cloud region data | Which provider regions serve which users | Would map production to consumption | No public dataset found |
Combination rule: Net AI exports = GIP Output produced in the country (power-located) − GIP Consumed (use-located). Production is allocated by AI electricity where metered or estimated, otherwise by Epoch chip share.
Combination rules
8. How the indexes are combined
These are the only choices GIP makes. Everything else is someone else's number.
- Normalization. Each family is converted to quarter-on-quarter growth and to country shares of the world total. Units cancel; only ratios are combined.
- Weighting. No subjective weights. Power, compute per MWh and quality multiply, because each converts one stage into the next. Consumption shares are a simple average of two sources. Exports are a subtraction.
- Quarterly chain-linking. Indext = Indext−1 × (Et/Et−1) × (Ct/Ct−1) × (Qt/Qt−1). Inputs and the AA version in use are frozen within each quarter and refreshed at quarter end.
- Version splices. A new AA Index version or a new source series enters through a one-quarter overlap, read on both old and new; the ratio in that quarter splices them.
- Annual-only data. World and US electricity are annual. Until quarterly figures exist, quarterly links carry them as a stated assumption, flagged in each release. No chart point is ever interpolated.
- Revisions. Releases about 45 days after quarter end; the prior four quarters are revised when meter data or filings arrive; every change is logged.
Back of the envelope
9. Illustrative figures
Two of the three output terms can be filled from published numbers today. Together they say the world's AI fleet delivered about twice as much computation in 2025 as in 2024. The quality term waits for an Artificial Analysis series.
AI power and chip efficiency alone doubled delivered computation in 2025
Illustrative growth factors, 2025 vs 2024 (times)
Illustrative
Bar chart. Use the up and down arrow keys to move through 4 categories. A data table follows.
E: AI electricity (IEA): Growth factor 1.5x
1.5 × 1.34 ≈ 2.0 (range 1.5 × 1.29 ≈ 1.9 to 1.5 × 1.39 ≈ 2.1). The IEA's +50% has no published range. Quality not yet included.
Source: IEA Key Questions on Energy and AI (Apr 2026); Epoch AI Trends (Feb 2026); our multiplication Download CSV Methodology
Data table
| component | Growth factor |
|---|---|
| E: AI electricity (IEA) | 1.5x |
| C: compute per MWh (Epoch FLOP/s per watt) | 1.34x |
| E x C: delivered computation | 2.01x |
| Q: AA quality per compute | – |
| Figure | Value | How derived | Status |
|---|---|---|---|
| World AI electricity, 2025 | ~155 TWh | IEA AI-focused data-centre electricity, as cited by OWID | Modelled estimate |
| US AI electricity, 2025 | ~72 TWh | IEA US 224 TWh × world AI-focused share 155/485 (our assumption) | Illustrative; likely understated |
| Delivered computation, 2025 vs 2024 | ~2.0x (1.9x-2.1x) | E × C | Illustrative |
| US share of world AI chips (Capacity proxy) | 72.2% | Epoch cumulative H100e arrivals to Q2 2026 | Third-party estimate |
| Real GIP growth including quality | Withheld | Needs the AA score-per-cost series and its first frozen quarter | Not defensible yet |
| Real GIP index, 2023 = 100 | Withheld | No IEA AI-focused figure for 2023 is published | Not defensible yet |
| Country Output, Consumed, Net exports | Withheld | Needs working-age population weights and country AI-electricity shares | Not computed yet |
Cross-checks
10. Dollars and tokens, kept beside the headline
If GIP rises fast while revenue and tokens stall, something in the combination is wrong. These series are the alarm bells.
Four firms sell AI output at about $157bn a year; Google's share is not disclosed
Annualized AI-service revenue, $bn, latest disclosure. Ads permanently excluded. Partial coverage
Bar chart. Use the up and down arrow keys to move through 6 categories. A data table follows.
Anthropic (Q3 2026 x 4): Revenue $67.2
Anthropic = Q3 2026 revenue $16.8bn × 4; SpaceX = Q2 2026 AI solutions ex-cloud ~$594m × 4 (our calculations). Range ~$154bn-$163bn.
Source: CNBC (Oct 8 and Oct 15, 2026); Microsoft IR (Apr 29, 2026); SpaceX Q2 2026 earnings release, SEC (Aug 4, 2026) Download CSV Methodology
Data table
| firm | Revenue |
|---|---|
| Anthropic (Q3 2026 x 4) | $67.2 |
| OpenAI (end Sep 2026 run-rate) | $50 |
| Microsoft AI business (Mar 2026) | $37 |
| SpaceX SI ex-cloud (Q2 2026 x 4) | $2.4 |
| Google (in scope not disclosed) | $0 |
| Nominal cross-check | $156.6 |
| Cross-check | Latest | Rule | Source |
|---|---|---|---|
| Nominal GIP | ~$157bn annualized (range ~$154bn-$163bn) | Sales of intelligence only. All advertising permanently excluded and never imputed. Rented compute excluded as intermediate. Count each dollar once. | Company disclosures and SEC filings |
| Google tokens, all surfaces | >3.2 quadrillion a month (May 2026) | Company-reported; no country split | Google I/O 2026 |
| Google first-party API tokens | ~32bn a minute (Q3 2026) | Company-reported | Alphabet Q3 2026 call |
| OpenRouter public-ranking tokens | 188T, week of Oct 19, 2026 | Third-party compilation | The Latent |
Work
11. The labor link
The question for this blog is what machine output means for human work. So GIP is shown per human hour worked, in the US and wherever hours data exist.
GIP per hourc,t = GIP Outputc,tTotal hours workedc,t
- US: BLS nonfarm business hours worked (series PRS85006033), index 105.9 in Q2 2026 (2017 = 100). Labor's share of output was 52.8% in Q2 2026, the lowest on record since 1947 (BLS, Sep 15, 2026).
- Other countries: OECD average annual hours actually worked per worker. Total hours = hours per worker × employment; the employment step is not yet built.
- Reading: if GIP per hour climbs while labor share falls, machine output is substituting for paid work rather than complementing it.
| Country | Hours per worker, 2025 | 2024 |
|---|---|---|
| United States | 1,800 | 1,796 |
| Ireland | 1,639 | 1,645 |
| Japan | 1,598 | 1,617 |
| United Kingdom | 1,533 | 1,538 |
| Finland | 1,492 | 1,502 |
| Netherlands | 1,436 | 1,452 |
| Sweden | 1,421 | 1,433 |
| Norway | 1,386 | 1,388 |
| Denmark | 1,369 | 1,372 |
| Germany | 1,332 | 1,334 |
Source: OECD, Average annual hours actually worked per worker (dataset updated Jun 7, 2026), retrieved Nov 1, 2026. OECD says the series suits comparisons over time, not of levels between countries.
Inputs
12. Data sources
What exists today for each family, checked Nov 1, 2026.
| Family | Series | Publisher | Frequency | Latest value (date) | Quality |
|---|---|---|---|---|---|
| Power | Data-centre electricity, world and regions; capacity factor; PUE | IEA | Annual | 485 TWh 2025; US 224; China 117; Europe 72 (Apr 2026) | Modelled (intergovernmental) |
| Power | AI-focused data-centre electricity | IEA, cited by Our World in Data | Annual | 155 TWh 2025 (Sep 2026) | Modelled, secondary citation |
| Power | US data-centre electricity; AI server assumptions | LBNL | Ad hoc | 176 TWh 2023, 4.4% (Dec 2024) | Official (DOE lab), modelled |
| Power | Ireland data centres, metered | CSO | Quarterly | 7,663 GWh 2025, 23% (Jul 7, 2026) | Official, metered |
| Power | Netherlands data centres | CBS | Annual | 5.09bn kWh 2024, 4.6% (Dec 2025) | Official, metered |
| Power | Norway data centres | Elhub | Hourly | 3.14 TWh 2025, 2.26% (Sep 2026) | Official, metered; crypto mixed in |
| Power | Great Britain data centres | DESNZ | Ad hoc | 4.5 TWh 2024, 2% (Jun 30, 2026) | Official, meter-based |
| Power | Denmark data centres | Danish Energy Agency | Annual | 4.3 PJ 2025 | Official, operator-reported |
| Power | China computing centres | National Energy Administration | Ad hoc | 170bn kWh 2025, 1.6% (May 29, 2026) | Government statement |
| Power | Germany data centres | Borderstep Institute / Bitkom | Annual | 21.3bn kWh expected 2025 | Industry estimate |
| Power | Singapore data centres | Ministry of Digital Development and Information | Ad hoc | 3.40 TWh 2020, ~7% | Official, stale |
| Power (exclusion) | US crypto mining electricity | EIA | Discontinued survey | 0.6%-2.3% of US electricity, 2023 (Feb 1, 2024) | Official, preliminary |
| Power (capacity) | AI data-centre power capacity | Epoch AI | Quarterly | 31.24 GW facility power, end-2025, partly extrapolated (Jan 16, 2026) | Third-party estimate |
| Compute per power | FLOP/s per watt; compute stock growth | Epoch AI Trends | Continuous | 1.34x a year; stock 3.4x a year (Feb 5, 2026) | Third-party estimate |
| Quality | AA Intelligence Index v5.0 | Artificial Analysis | Continuous | Nine evaluations; v5 launched Oct 27, 2026 | Third-party index (sole quality measure) |
| Consumption | AI diffusion by country | Microsoft AI Economy Institute | Quarterly | World 18.8% of working-age people, Jun 2026 (Sep 2026) | Company telemetry, adjusted |
| Consumption | Claude usage share and per-capita index by country | Anthropic Economic Index | Monthly | US 20.16% of use, May 2026 (release Jun 26, 2026) | Company data, one provider |
| Exports | AI chip imports by destination | Epoch AI | Quarterly | 27.4m H100e cumulative to Q2 2026; US 72.2% (CSV Oct 27, 2026) | Third-party, from customs data |
| Exports | Digitally delivered services trade | WTO | Annual | 2005-2025, 200+ economies (Jul 2026) | Official estimate, not AI-specific |
| Labor | Hours worked, nonfarm business | BLS (PRS85006033) | Quarterly | 105.9, 2017 = 100 (Q2 2026) | Official |
| Labor | Hours per worker by country | OECD | Annual | 2025 values (Jun 7, 2026) | Official |
Transparency
13. Gaps
- Quality series not built: we have not yet pulled Artificial Analysis's history of Index scores and cost to run, so the quality term and real GIP growth are withheld.
- No country publishes the AI share of data-centre electricity. The IEA's world share (~32%) is applied everywhere, which likely understates the US.
- No metered national series for the US or China. EIA has no data-centre series; its crypto survey was discontinued.
- World and US electricity are annual; GIP is quarterly. Quarterly links carry annual data as a flagged assumption.
- Consumption: Anthropic covers one provider and publishes no China row; Microsoft measures users, not intensity. No OECD or Stanford AI Index country adoption series was retrieved for this draft.
- Exports: no public data map cloud regions to users; WTO digital-services trade has no AI split; Epoch reports the EU as one bloc.
- Utilization: no Uptime Institute, ERCOT, PJM or Dominion figure on consumed vs contracted AI load was verified. Interconnection queues measure requests, not use.
- Crypto is not separated in Norway's metered series or, as far as published, in Ireland's or the Netherlands'.
- Japan, Sweden, Finland: no national data-centre electricity figure found. Singapore's latest official figure is from 2020. China's NEA and IEA figures disagree by 45%.
- The IEA AI-focused figure (155 TWh) is taken from Our World in Data's citation of the IEA datafile; we have not opened the datafile itself.
Sources
14. Sources
- Key Questions on Energy and AI (full report, annex tables). IEA. Apr 2026.
- Key Questions on Energy and AI: executive summary. IEA. Apr 2026.
- How much energy do data centers and AI use?. Our World in Data. Sep 2026.
- 2024 US Data Center Energy Usage Report. LBNL. Dec 2024.
- Data Centres Metered Electricity Consumption 2025. CSO Ireland. Jul 7, 2026.
- Datacenters verbruiken 4,6 procent van de elektriciteit. CBS. Dec 2025.
- Hvordan bruker datasentre strøm?. Elhub. Sep 2026.
- Data centre electricity consumption in Great Britain, 2020 to 2024. DESNZ. Jun 30, 2026.
- Energistatistik 2025. Danish Energy Agency. 2026.
- National computing centres used 170bn kWh last year. National Energy Administration (China). May 29, 2026.
- Data Digits. Borderstep Institute. 2026.
- Reply to PQ on data centre emissions. MDDI Singapore. 2021.
- Tracking electricity consumption from US cryptocurrency mining. EIA. Feb 1, 2024.
- Global AI power capacity. Epoch AI. Jan 16, 2026.
- Trends in AI. Epoch AI. Feb 5, 2026.
- Leading ML hardware becomes 40% more energy-efficient each year. Epoch AI. Oct 23, 2024.
- Data on AI chip imports. Epoch AI. Oct 27, 2026.
- Intelligence Index methodology. Artificial Analysis. Live (v5.0), read Nov 1, 2026.
- Intelligence Index v5. Artificial Analysis. Oct 27, 2026.
- Global AI Diffusion, Q2 2026. Microsoft. Sep 2026.
- AI diffusion report data. Microsoft (GitHub). Q2 2026.
- Anthropic Economic Index data (release Jun 26, 2026). Anthropic (Hugging Face). Jun 26, 2026.
- Digitally delivered services trade dataset. WTO. Jul 2026.
- Hours worked. OECD. Jun 7, 2026.
- Productivity and Costs, Q2 2026 revised. BLS. Sep 3, 2026.
- Labor share at its lowest level. BLS TED. Sep 15, 2026.
- OpenAI revenue report. CNBC. Oct 8, 2026.
- Anthropic's Q3 revenue hits $16.8bn. CNBC. Oct 15, 2026.
- Anthropic revenue and burn rate. Luminix. 2026.
- Microsoft FY26 Q3 results. Microsoft. Apr 29, 2026.
- Alphabet Q3 2026 earnings release. SEC. Oct 28, 2026.
- Q3 2026 earnings call: remarks from our CEO. Google. Oct 28, 2026.
- Sundar Pichai at I/O 2026. Google. May 2026.
- SpaceX reports third quarter 2026 results. SEC. Oct 20, 2026.
- SpaceX reports second quarter 2026 results. SEC. Aug 4, 2026.
- OpenRouter weekly token volume. The Latent. Oct 2026.
- Prior work: Lee Seung-hyun on Gross Intelligence Product. Seoul Economic Daily. Mar 22, 2026.
Post-Labor Economics · grossintelligenceproduct.com · Methodology v0.2 · Last updated Nov 1, 2026. Prior work: Lee Seung-hyun (42Maru) proposed Gross Intelligence Product as a measure of AI output; this is an independent, data-driven methodology.