GIP Index / Foundations

Introducing the Gross Intelligence Product

A new index that measures how much useful machine thinking the world produces, built only from statistics other people already publish, starting with the electricity AI actually uses.

The world’s AI data centres delivered roughly twice as much computation in 2025 as they did in 2024. That is the first reading from a measure we are calling the Gross Intelligence Product, or GIP, and I think it is the most important number in economics that nobody is publishing.

GDP tells you what an economy sells. It does a poor job of telling you how much of the thinking behind that output is now done by machines. If this publication is going to say whose work machines are taking over, we need a way to count the machine side of the ledger.

So we built one. This piece explains what GIP measures, how it works, what we can already say, and what we are deliberately not saying yet.

Credit where it is due

We did not come up with the name. Lee Seung-hyun, a vice-president at the Korean AI company 42Maru, proposed a “Gross Intelligence Product” in his 2026 book AI Wealth of Nations, as a way to measure a nation’s total intelligence assets and move beyond GDP, according to the Seoul Economic Daily.

Our version is independent and narrower. Lee frames GIP as a national strategy. We are trying to build a statistic, with every input published and every rule printed, so that anyone can redo the arithmetic and tell us where we got it wrong.

The rule we started with: invent nothing

The first rule is that GIP creates no new measurements. Every number in it comes from someone else’s published series. Our only contribution is the combination rules: how those series are normalized, multiplied and linked over time.

That rule exists because AI measurement is full of numbers that are really marketing. A company announces a gigawatt campus, a lab announces a benchmark it designed, a consultant announces a trillion-dollar impact. We wanted something boring enough to trust.

GIP draws on four families of existing data. AI power is the electricity actually consumed by AI data centres, from the International Energy Agency and national statistics offices. AI quality is the capability of the models being run, from the Artificial Analysis Intelligence Index.

AI consumption is who uses AI, by country, which is pending neutral sources. AI exports compare where AI is produced with where it is used, starting from Epoch AI’s customs-based chip data.

The formula

At its core, GIP Output is three published numbers multiplied together:

GIP Output = AI electricity consumed × compute per unit of electricity × capability per unit of compute

The logic is a production chain. Electricity runs chips. Chips turn electricity into computation.

Models turn computation into useful answers. Each term converts one stage into the next, so each stage is counted once.

That last point is the reason the current version exists. Our first draft added up power, chips, tokens and models as four separate pillars with weights. Then we realized power feeds compute, and compute produces tokens, so we were counting one chain several times.

We withdrew that design. Version 0.2 measures one physical quantity, energy, and multiplies it by two efficiency ratios.

Why we start with electricity

A gigawatt of announced capacity is not a gigawatt of thinking. Sites wait for chips and grid connections, and demand for inference rises and falls through the day. So GIP counts the power AI actually draws, read from meters wherever a government publishes them.

The global figure comes from the IEA. World data centres used about 485 terawatt-hours of electricity in 2025, around 1.5% of global electricity, and AI-focused data centres used about 155 TWh of that, roughly a third, according to Our World in Data’s analysis of IEA figures. Electricity consumption from AI-focused data centres surged 50% in 2025, according to the IEA.

Those are modelled estimates, not meter readings. Only a handful of countries actually meter data-centre electricity. Ireland is the gold standard, with a quarterly official series going back to 2015.

Its data centres used 7,663 gigawatt-hours in 2025, 23% of all metered electricity in the country, according to the Central Statistics Office. That is about six times the 1,240 GWh they used in 2015.

The Netherlands (CBS), Great Britain (DESNZ) and Norway (Elhub) also publish official figures. The two largest markets, the United States and China, do not have a metered national series.

China’s National Energy Administration put “computing centre” consumption at 170 TWh for 2025 (NEA), 45% above the IEA’s 117 TWh estimate for the same year. Definitions differ, and the NEA’s method is not published.

No country publishes the share of data-centre electricity that goes to AI. Until one does, we apply the IEA’s world share, about 32%, everywhere. That likely understates the US.

One more rule: crypto mining is out. Proof-of-work mining is not intelligence. The IEA’s data-centre figures already exclude it.

Some national meter series do not, and we flag those. For scale, the U.S. Energy Information Administration’s preliminary estimate put US crypto mining at 0.6% to 2.3% of US electricity in 2023 (EIA).

Compute per unit of power

Each new generation of chips does more arithmetic per watt. We don’t estimate this ourselves.

We use Epoch AI’s published trend: GPU computing power per watt has improved about 1.34 times a year since 2008, with a 90% confidence interval of 1.29 to 1.39 (Epoch AI). Quarterly, that works out to roughly 1.076 times.

Capability per unit of compute

Computation is not intelligence. A chip running an old, weak model at full tilt is not producing as much useful thinking as one running a strong model. So the third term measures quality.

We use exactly one yardstick, the Artificial Analysis Intelligence Index. Using one measure means nobody, including us, gets to pick the benchmarks that flatter a result. Version 5.0, released on October 27, 2026, combines nine evaluations across four equally weighted categories: knowledge work, coding, science and general reasoning.

Artificial Analysis publishes each model’s score and the dollar cost of running the full index. It does not publish computing operations. So we measure quality as score per dollar of running cost, adjusted by Epoch’s chip price-performance trend (about 1.49 times a year) so hardware gains we already counted are not counted twice.

We track two versions: the frontier model, and the model at the median of actual usage. Most electricity serves ordinary models, not the best one, so the median feeds the headline.

The index has limits, and its makers say so. It is mostly text-based and in English. Some evaluations are graded by AI judges.

Components change between versions. When a new version arrives, we read models on both old and new versions for one overlapping quarter and splice the series at that ratio, so a version change never moves GIP by itself.

Keeping the hype out

The obvious worry with any index like this is that it becomes a hype machine with a decimal point. A number that says machine intelligence doubled sounds like a countdown to the end of work.

Here is how we try to keep it honest.

First, the numbers we cannot defend, we withhold. Real GIP growth including quality is withheld, because we have not yet built the Artificial Analysis score-and-cost history and frozen its first quarter. A GIP index with 2023 set at 100 is withheld, because the IEA has not published an AI-focused electricity figure for 2023.

Country-level scores are withheld, because they need country AI-electricity shares and a neutral measure of consumption by country, and neither exists yet. Each will be published once the data exist, with the method shown, and not a quarter before.

Second, GIP is anchored in physical measures, starting with electricity, not in revenue or announcements.

Third, we keep a money check beside the headline. Four companies report AI revenue we can attribute to intelligence itself, and together it comes to about $157 billion a year: Anthropic’s third-quarter revenue of $16.8 billion (CNBC) annualizes to about $67 billion, OpenAI is reported at a run rate of about $50 billion (CNBC), Microsoft reported an AI business of more than $37 billion as of March (Microsoft), and SpaceX reported AI solutions revenue, outside cloud, of about $594 million in the second quarter, or roughly $2.4 billion a year (SEC). Google does not disclose a comparable figure.

Our range for the total is about $154 billion to $163 billion. It is partial, nominal, and only a cross-check. If GIP races ahead while revenue stalls, something in our method is wrong.

Fourth, advertising is permanently excluded. That includes AI-boosted search and ad targeting at Google, Meta and everyone else.

GIP counts intelligence sold as intelligence, through models, APIs, subscriptions and AI services. Ad revenue measures attention, and it will never be imputed.

What we can say today

Two of the three terms can be filled from published data now. AI electricity rose about 1.5 times in 2025. Compute per watt improved about 1.34 times.

Multiply them and the world’s AI fleet delivered about 2.0 times as much computation in 2025 as in 2024, with a range of about 1.9 to 2.1 times based on Epoch’s interval. The IEA publishes no range for its 50%.

That figure is not GIP. It leaves out the quality term, which, if models keep getting better per dollar, will push the real number higher. Treat it as a floor on how fast machine output is growing, not a headline.

The second thing we can say is about geography, on the production side only. Summing the country rows in Epoch AI’s customs-based dataset, the US received 72.2% of the world’s AI-chip computing power that arrived from 2022 through mid-2026 (Epoch AI). Arrivals do not equal chips switched on, but the concentration is hard to miss.

The other half of the picture, where that intelligence is actually used, is not ready. Consumption by country is pending neutral sources.

Today the only large country-level usage figures come from AI companies measuring their own products, and one provider’s customers are not a country’s consumption. We will not publish a country consumption share, or the export view built on it, until an independent statistical source exists.

When it does, the export view becomes the interesting part. A country full of data centres and few users is an exporter of intelligence.

A country full of users and few chips imports it. That will matter for trade, for taxes and for where the remaining jobs are.

Why this belongs on a labor site

GIP only earns its place here if it tells us something about human work. So the first derived measure will be GIP per hour worked.

US business hours worked stood at 105.9 in the second quarter of 2026, on an index where 2017 equals 100, according to the Bureau of Labor Statistics. Labor’s share of output was 52.8%, the lowest since 1947 (BLS).

If machine output per human hour keeps climbing while labor’s share keeps falling, that is evidence machines are substituting for paid work. If both rise together, they are complementing it. That is the question this whole publication exists to answer, as we laid out in what post-labor economics means.

We will publish about 45 days after each quarter ends and revise the previous four quarters when better meter data or filings arrive. Every change gets logged. The full method, version 0.2, is published separately as the GIP methodology.

I’ll admit it is a little strange to build an index of machine intelligence by hand, one spreadsheet at a time.

What to watch

GIP is our attempt to measure how cheap and how plentiful machine knowledge is becoming. What stays scarce, and worth investing in, is the subject of a separate piece.

If you are a policymaker, push your statistics office to meter data-centre electricity and publish an AI share. Ireland shows it can be done. Business owners can track their own version: how much of your output now comes from AI tools per hour your team works, and what you are doing with the hours that frees.

And if you are a worker, watch GIP alongside the job data in our job thermometer. When machine output jumps in a field, that is your early signal to move toward the judgment, the relationships and the hands-on work that the machines are not doing. Start keeping that log this quarter, so you have a baseline before the next jump.

Sources

1. International Energy Agency, “Key Questions on Energy and AI: Executive summary,” Apr 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
2. International Energy Agency, “Key Questions on Energy and AI” (full report and annex tables), Apr 2026. https://www.iea.org/files/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf
3. Our World in Data, “How much energy do data centers and artificial intelligence use?,” Sep 2026. https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use
4. Epoch AI, “Trends in AI,” updated Feb 5, 2026. https://epoch.ai/trends
5. Epoch AI, “Global AI power capacity,” Jan 16, 2026. https://epoch.ai/data-insights/ai-datacenter-power
6. Epoch AI, “Data on AI chip imports,” CSV updated Oct 27, 2026. https://epoch.ai/data/ai-chip-imports
7. Artificial Analysis, “Intelligence Index methodology” (v5.0), read Nov 1, 2026. https://artificialanalysis.ai/methodology/intelligence-benchmarking
8. Artificial Analysis, “Artificial Analysis Intelligence Index v5,” Oct 27, 2026. https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5
9. Central Statistics Office (Ireland), “Data Centres Metered Electricity Consumption 2025,” Jul 7, 2026. https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/keyfindings/
10. Statistics Netherlands (CBS), “Datacenters verbruiken 4,6 procent van de elektriciteit,” Dec 2025. https://www.cbs.nl/nl-nl/nieuws/2025/51/datacenters-verbruiken-4-6-procent-van-de-elektriciteit
11. UK Department for Energy Security and Net Zero, “Data centre electricity consumption in Great Britain, 2020 to 2024,” Jun 30, 2026. https://assets.publishing.service.gov.uk/media/6a3ea4334c7605ab56723bfd/Data_centre_electricity_consumption_in_Great_Britain__2020_to_2024.pdf
12. Elhub, “Hvordan bruker datasentre strøm?,” Sep 2026. https://frzq0sxltynr.objectstorage.eu-frankfurt-1.oci.customer-oci.com/n/frzq0sxltynr/b/insight-prod-open-data-osb/o/download/publications/elhub_analyse_hvordan_bruker_datasentre_strom_202609.pdf
13. National Energy Administration (China), statement on computing centre electricity use, May 29, 2026. https://www.nea.gov.cn/20260529/18c09743776c4dddb55288ceea38b8b1/c.html
14. U.S. Energy Information Administration, “Tracking electricity consumption from U.S. cryptocurrency mining operations,” Feb 1, 2024. https://www.eia.gov/todayinenergy/detail.php?id=61364
15. Seoul Economic Daily, “AI emerges as national strategic asset, new book argues,” Mar 22, 2026. https://en.sedaily.com/technology/2026/03/22/ai-emerges-as-national-strategic-asset-new-book-argues
16. CNBC, “Anthropic’s Q3 revenue is said to hit $16.8 billion ahead of IPO,” Oct 15, 2026. https://www.cnbc.com/2026/10/15/anthropics-third-quarter-revenue-hits-16point8-billion-ipo.html
17. CNBC, “Nvidia, Oracle, other AI stocks sink on OpenAI revenue report,” Oct 8, 2026. https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html
18. Microsoft, FY26 Q3 earnings press release and webcast, Apr 29, 2026. https://www.microsoft.com/en-us/investor/earnings/fy-2026-q3/press-release-webcast
19. SpaceX, “SpaceX reports second quarter 2026 results,” SEC filing, Aug 4, 2026. https://www.sec.gov/Archives/edgar/data/1181412/000162828026052515/earningsreleaseq22608042.htm
20. U.S. Bureau of Labor Statistics, “Productivity and Costs, Second Quarter 2026, Revised,” Sep 3, 2026. https://www.bls.gov/news.release/prod2.htm
21. U.S. Bureau of Labor Statistics, “Labor share at its lowest level, 52.8 percent, in second quarter 2026,” Sep 15, 2026. https://www.bls.gov/opub/ted/2026/labor-share-at-its-lowest-level-52-8-percent-in-second-quarter-2026.htm