Global Token Atlas
Global Token Atlas

Global Token Atlas · First release · Data as of 5 October 2026

AI use is spread unevenly around the world.

This first release measures chatbot usage: how often people in each country use three AI chatbots, as their providers have published it. It does not count the tokens each country consumes.

In April 2026, the United States recorded about 13 times as many sampled Claude conversations per resident as Kenya, and about 24 times as many as Nigeria.

National averages for one chatbot in one month. Not a comparison of typical users, and not a census of all AI use.

Start reading

Or jump straight to
{{sm.provider}} {{sm.period}} {{sm.kind}}
{{sm.measure}}
{{l.name}} {{l.val}}

The map could not be drawn.

The story continues below, and every country's values remain available in the explorer table. No geographic picture is shown in its place.

{{r.name}}{{r.val}}
{{l.name}} {{l.val}}
{{r.label}}
{{k.label}}
{{sl.caption}}

01

Where is AI being used?

Three providers give us different windows into that pattern. Their measures are not interchangeable, but they broadly agree on which countries use their products more intensively.

The map shows the first window. Each country's share of sampled Claude conversations in April 2026 is divided by its share of world population. A value of 1 is the world average; darker countries recorded more conversations per resident.

Select any country on the map to look it up.

{{inlineClaude.provider}} {{inlineClaude.period}} {{inlineClaude.kind}}
{{inlineClaude.measure}}
{{l.name}} {{l.val}}

The map could not be drawn.

The story continues below, and every country's values remain available in the explorer table. No geographic picture is shown in its place.

{{r.name}}{{r.val}}
{{l.name}} {{l.val}}
{{r.label}}
{{k.label}}
{{inlineClaudeLegend.caption}}

02

Three windows, a similar pattern

OpenAI publishes a ranking of ChatGPT messages per resident, covering April to June 2026.

A rank gives an order. Rank 1 is the highest of 147, but the ranking says nothing about how far apart any two countries are.

{{inlineGpt.provider}} {{inlineGpt.period}} {{inlineGpt.kind}}
{{inlineGpt.measure}}
{{l.name}} {{l.val}}

The map could not be drawn.

The story continues below, and every country's values remain available in the explorer table. No geographic picture is shown in its place.

{{r.name}}{{r.val}}
{{l.name}} {{l.val}}
{{r.label}}
{{k.label}}
{{inlineGptLegend.caption}}

Google's release groups countries into five bands of referral-adjusted conversational intensity, for two weeks in April 2026.

A band groups countries. Band 5 is the highest, and countries in the same band are not necessarily equal.

{{inlineGoogle.provider}} {{inlineGoogle.period}} {{inlineGoogle.kind}}
{{inlineGoogle.measure}}
{{l.name}} {{l.val}}

The map could not be drawn.

The story continues below, and every country's values remain available in the explorer table. No geographic picture is shown in its place.

{{r.name}}{{r.val}}
{{l.name}} {{l.val}}
{{r.label}}
{{k.label}}
{{inlineGoogleLegend.caption}}

Across the 110 countries with all three comparison measures, the country orders correlate between 0.83 and 0.91. That is substantial agreement, but it does not mean the products have equal usage or identical coverage.

{{a.label}} {{a.rho}}
01

A rank gives an order, a band groups countries, and a conversation share gives a proportion.

Spearman rank correlations on the shared set. For this calculation only, ChatGPT ranks are reversed so that higher always means more use; ranks shown elsewhere keep 1 = highest.

03

How large is the gap?

Claude April 2026 Derived comparison
{{ selectA }} {{ selectB }}

{{pair.headline}}

{{r.name}} {{r.sub}} · {{r.mult}}

{{pair.unitCaption}}

This comparison can't be calculated.

{{m}}

A ratio is shown only when both countries have an April 2026 Claude share and the population (or internet-use) figure it needs.
Conversations per

{{pair.explain}}

{{pair.bounds}}

A ranking cannot tell us whether two countries are a little or a lot apart. Claude's published conversation shares let us go further. Dividing each country's share of sampled conversations by its share of world population gives a per-resident index.

The United States accounts for 20.97% of the sampled conversations in the April table; Kenya accounts for 0.26%, and Nigeria 0.61%. Their populations differ, so the shares alone are not the comparison.

After accounting for population, the US–Kenya ratio is about 13 to one; the US–Nigeria ratio is about 24 to one.

04

What changes with the denominator?

About 35% of Kenyans use the internet, compared with about 95% of Americans in the retained 2024 context. Change the denominator from residents to internet users and the US–Kenya comparison narrows from about 13 to about 5. The US–Nigeria comparison narrows from about 24 to about 10.

That change answers a different question. It does not estimate how much of the gap internet connectivity causes, or what would happen if access improved.

Prices, payment access and the models people use may matter too, but these data do not establish their effects.

Try another pair with the menus above the chart.

Inputs for the selected pair. Claude shares and index: April 2026. Population and internet use: World Bank, year shown.
Country Share of sampled conversations Population Internet users Per-resident index (1 = world)
{{r.name}} {{r.share}} {{r.pop}} {{r.net}} {{r.idx}}

— means no published or retained value. It is never a zero.

Conversation share
The country's percentage of the sampled Claude.ai conversations in Anthropic's April 2026 table. Published to two decimals. The named countries' shares sum to 82.03%; we do not renormalize them.
Per-resident index
Our rebase: share of sampled conversations divided by share of world population (World Bank, 2024). 1 = the world average per resident. This differs from Anthropic's published working-age index.
Per-internet-user ratio
Each country's index divided by the fraction of its residents who use the internet, then one divided by the other. Shown only as a pair ratio, not as a map.

For the selected pair

{{ln}}

{{calcRounding}}

Period: 1–30 April 2026. A country average does not describe a typical user. The comparison is descriptive; it does not identify causes.

Sources: Anthropic Economic Index country data, its documentation, World Bank population and internet use.

05

What remains out of view?

{{gm.provider}} {{gm.period}} {{gm.kind}}
{{gm.measure}}
{{l.name}} {{l.val}}

The map could not be drawn.

{{r.name}}{{r.val}}
{{l.name}} {{l.val}}
{{r.label}}
{{k.label}}
{{gl.caption}}

A blank country is not a country with no AI use.

Ethiopia appears in the retained ChatGPT and Google data but has no published April Claude conversation share. China has no comparable value in these retained sources. Some places have source observations but lack population context for our rebase.

Anthropic's published April country shares sum to 82.03%, leaving part of the sample without a named country in the table. We do not redistribute that remainder or fill gaps with zeros.

82.03% in named countries 17.97% not named

Nor can conversations be turned into country token totals without knowing how many tokens those conversations contain.

The first release can show where measured chatbot use is concentrated and where its windows disagree. A global account of the quantity and capability of AI people consume still needs evidence.

Look up a country

Explorer

Look up a country

Choose a window, then select a country on the map, in the table or by searching. Each provider keeps its own measure and scale; your selected country stays selected as you switch.

{{em.provider}} {{em.period}} {{em.kind}}

{{em.measure}}

The map could not be drawn, so the explorer is showing the country table. Every value and selection works the same way.
{{l.name}} {{l.val}}
{{r.label}}
{{k.label}}
{{el.caption}}

Too small to draw at this map scale, but measured:
{{tableCaption}}
Country ChatGPT rank
Apr–Jun 2026		 Claude index
Apr 2026		 Google quintile
6–19 Apr 2026
{{r.gpt}} {{r.claude}} {{r.google}}

— means no published or comparable value; never zero. Values come from three separate measures and are not interchangeable.

Sources and notes

A dated snapshot taken on 5 October 2026. Values are retained research outputs, not a live feed. No provider or institution has reviewed or endorsed this atlas.