Every number here carries its source, the wording that supports it, and the period it describes. This page is how to check us.
“population increased from 42,9 million in 2002 to 63,5 million in 2026”
The period a number describes and the day we read it are two different dates. Confusing them is how a dataset goes quietly wrong.
Nobody publishes this count, and there’s no shared definition of what qualifies. We couldn’t reproduce our own previous value of 8 from any source.
So it ships as unverified, with the deployments we could confirm listed instead.
A gap you can see beats a number you can’t check.
Compute and connectivity weigh most, because they decide whether inference can run at all; policy and ecosystem decide whether it's practical to operate there.
Whether AI-grade compute is actually available: AI-capable facilities, hyperscaler region depth, and whether GPUs can be rented or colocated at all.
Submarine cable access, internet exchange density, and round-trip latency to the nearest European point of presence.
Grid reliability for sustained inference workloads. Large facilities run generators, which mitigates outages but raises operating cost.
The rules you'd operate under: data protection in force, a national AI strategy, and whether cross-border data flows are aligned.
The founder and developer base that determines local talent, tooling, and actual demand for AI products.
An analyst assigns the scores using the bands below. A test recomputes the composite for every country on every build, so the weighting shown here is the one the data follows.
Every dimension runs 0–100 against fixed criteria, published in full so you can argue with a specific judgement rather than the number in general.
One question: can you run inference here today, or not yet.
Local or near-local inference is operationally feasible today.
Infrastructure exists, with real gaps in compute, power or policy.
Early stage. Offshore inference is the only practical option.
An operator describing its own capacity and a regulator publishing statistics are different kinds of evidence. Keeping them apart is what catches a planned facility reported as a running one.
The organisation describing itself. Authoritative on its own facilities, and commercially motivated.
An operator's own capacity page, a company's financial filing
A national statistics office or sector regulator, independent of whoever is being described.
National statistics releases, telecoms regulator quarterly reports
Bodies that publish methodology alongside numbers.
World Bank, ITU, PeeringDB, submarine cable registries, sector research firms
Useful for corroboration and for finding primary sources. Never the sole basis for a value.
Industry publications, general news reporting
Collecting a number and confirming it are separate jobs, done separately on purpose.
A script re-fetches every source and checks the quote is really there. Paraphrased or invented citations fail automatically. You can run it yourself; the evidence files are public.
A quote can be real and still not support the number: wrong operator, wrong year, a projection read as fact. A second reviewer, never the one who collected it, checks each value against its evidence.
Stated here rather than discovered later.
The whole protocol is public: schema, source rules, both gates, and how a pass becomes the live dataset.
Read the data protocol