Where Are AI Companies Being Founded? Evidence from ~30 National Company Registers

Matt Cortland (Prime Directive AI Ltd) and Dr John Fleming (AI researcher, University of Oxford), published by Inference Preliminary working draft, June 2026

(Global companion to the United Kingdom founder study. Same method, extended across roughly thirty jurisdictions. Figures use the survivorship-robust metrics throughout, the AI-named share of all formation, 2025 levels, the 2023 inflection, and official true-births; raw-count growth multiples are flagged as upper bounds.)

Abstract

We assemble newly incorporated company records from the bulk registers of roughly thirty jurisdictions and, for the United Kingdom and Ireland, combine them with the full persons-with-significant-control (PSC) and officer registers. We identify companies whose registered names signal artificial-intelligence activity and measure their formation against the contemporaneous population of all other new companies. Across every register with usable coverage, AI-named company formation rises sharply after the public release of ChatGPT (November 2022), with a single-year inflection in 2023. In the two countries where we hold the entire register, AI-named companies reach approximately one in two hundred of all newly incorporated companies by 2025, up roughly six- to eight-fold as a share of all formation since 2022, and the two registers land almost identically. The rise holds as a denominator-controlled share across about twenty countries (median roughly 4.5-fold), and is not an artefact of a general formation boom: official statistics show total company formation rose only about 1.2 to 1.4-fold over the same period. The companies cluster in the occupations an independent measure (Massenkoff and McCrory, 2026; Anthropic Economic Index) rates most exposed to AI, so the supply side of firm creation mirrors the demand side of labour exposure. For the United Kingdom we further characterise the founders: increasingly solo, younger, and owner-operated, with a structural break in 2023. All results are descriptive; the cohort is a name-based proxy and the analysis is bounded by register coverage.

1. Introduction

The labour-market debate about generative artificial intelligence has been studied primarily on the demand side: which existing occupations are exposed, and whether employment or wages have responded (Eloundou et al., 2023; Handa et al., 2025; Massenkoff and McCrory, 2026). We examine a complementary margin, firm formation, and ask a question with a hard data answer: since ChatGPT, are more companies being started to do AI, who is starting them, and is this a local or a global pattern? Company registers are the natural instrument. Incorporation is a dated, mandatory, near-universal record, and a handful of registers additionally expose ownership and control. We treat the population of newly incorporated AI-named companies as observable and compare it to the contemporaneous population of all other new companies, in each country and pooled.

Our contribution is twofold. First, we extend a single-country founder study (the United Kingdom) into a cross-country measurement across roughly thirty registers, so the post-ChatGPT rise can be read as a global phenomenon rather than a national one. Second, we connect the supply side (which companies are being created) to the demand-side exposure literature (which work AI touches), by mapping each company's revealed activity onto the occupational exposure scores of the Anthropic Economic Index.

2. Data

Our sources are the bulk company registers of roughly thirty jurisdictions. Two are held as complete national registers with a clean company universe (the United Kingdom, Companies House, approximately 5.7 million live companies; Ireland, the Companies Registration Office). For approximately twenty-four further jurisdictions we hold enough of the register to compute a denominator (AI-named formation as a share of all formation); for the remainder we hold the AI-named cohort only. For the United Kingdom we additionally use the PSC snapshot (approximately 15.5 million control records) and officer appointments retrieved from the Companies House REST API. Across all countries we classify companies' activity from their public websites using an automated scrape and a large-language-model classifier.

2025 is the last complete comparable year. 2026 is reported as a labelled partial year (January to May, like-for-like against the same window of prior years), never annualised and presented as a full year. Per-country data currency varies and is recorded in the methods appendix.

3. Methods

The unit. A newly incorporated company whose registered name signals an AI focus. "Company" means an incorporated legal entity, not a sole trader; every register is filtered to that unit for comparability. A strict definition requires "AI" as a standalone token; a broader "AI ecosystem" definition admits related terms (artificial intelligence, GPT, LLM, machine learning, deep learning, generative AI, Copilot, AGI, agent).

The name-based cohort, de-contaminated. Keyword tagging over-counts: in Romance languages "IA" collides with everyday words, "agent" sweeps in estate agents, "Ai" is a given name. Every candidate company in every country is therefore classified by a language-aware large-language-model as genuinely AI or a false positive, with a logged reason; only confirmed-AI companies are counted. Validation against an independent website-confirmation pass shows the classifier errs toward keeping (it rejects only about 5% of web-confirmed AI companies). Per-country false-positive rates are reported in the methods appendix.

The primary metric is the share. Wherever we hold the full register we report AI-named companies as a share of all new formation, which controls for each country's own formation trend and definition. Cross-country we compare the share multiple (each country's 2025 share divided by its 2022 share), which cancels the level definition; we report the share level ("1 in 200") only within the two clean company registers.

Survivorship. Most register snapshots, including Companies House, contain only companies currently on the register; dissolved companies are removed. Counting by incorporation year therefore counts survivors, undercounts older years, and inflates raw-count growth. The share is largely immune (numerator and denominator lose the same dissolved companies, so it cancels in the ratio), 2025 levels are near-complete, and the 2023 inflection is a step too sharp for slow survivorship decay to manufacture. Where an official gross-incorporation series exists (true births, including since-dissolved companies), we use it for any total-formation statement.

The occupational crosswalk. For each company we map its website-derived vertical to an occupation, and that occupation to its "observed exposure" score in Massenkoff and McCrory (2026): developer tools and infrastructure to computer programmers (74.5), marketing technology to market research and marketing (64.8), healthcare to medical records specialists (66.7), fintech to financial analysts (57.2), security to information-security analysts (48.6). The crosswalk is held in one place and applied identically to every country, so the exposure overlap is comparable worldwide.

4. Results

4.1 About one in two hundred new companies is AI-named (United Kingdom and Ireland)

For the two countries where we hold the entire register we can ask the harder question and answer it in two independent datasets:

By 2025, roughly one in two hundred brand-new companies is explicitly AI-named, up from about one in fourteen hundred in 2022, and the two registers land within a whisker of each other. Because total UK formation also rose over the period, the honest, denominator-controlled figure is the share rise (about 6-fold), not the raw-count multiple.

4.2 The rise is global, with a shared 2023 fingerprint

Across about twenty countries where we hold the full register, AI's share of all new formation rose a median of roughly 4.5-fold from 2022 to 2025 (range about 3 to 9-fold), with leaders including Norway (~8.7x), New Zealand (~8.4x), Chile (~8.1x), Ireland (~7.7x), Canada (~6.4x) and the United Kingdom (~6.1x). Because each country's denominator definition is constant year to year, the multiple cancels the definition and controls for that country's own formation trend.

The clearest fingerprint is the single-year jump from 2022 to 2023, immediately after ChatGPT: Ireland 3.5x, the United Kingdom 3.3x, Spain 3.3x, New Zealand 3.2x, then continued climbing to 2025. Roughly thirty very different economies, legal systems and languages show the same shape.

Absolute levels are not comparable across countries (registers count different universes); the share multiple is the portable statistic, and the level statement ("1 in 200") rests on the clean company registers. The full set of per-country series is also available interactively on the International and Countries pages.

4.3 AI is not driving company creation overall

A tempting reading is that AI is spurring company creation broadly. It is not. Official gross-incorporation statistics (true births, including since-dissolved companies) show total company formation rose only modestly over 2018 to 2024, in line across countries:

Total formation grew about 1.2 to 1.4-fold everywhere, normal business-cycle territory, while AI-named formation rose several-fold. So AI-named companies are rising as a genuine share of a near-flat backdrop, not because everyone is forming more companies. The cleanest illustration is Ireland, a full-history register with no survivorship distortion: total formation flat (about 1.2x), AI-named share up 7.7-fold. As a scale marker, the UK AI-named ecosystem is now about 14 times the size of the crypto-named ecosystem on the register (about 9,100 companies versus 655), and unlike crypto it is still growing.

4.4 2026 is on track for a record

On a like-for-like same-window basis (January to May, each year), 2026 is already ahead of 2025: the United Kingdom formed about 1,772 AI-named companies in the first five months of 2026, roughly 40% above the same window of 2025, an all-time-high 0.524% of all new companies (about 1 in 191). Holding that pace, 2026 projects to roughly 5,100 for the full year. Twenty of twenty-one comparable countries are up year on year in the first five months of 2026 alone. The surge runs right up to the eve of the Fable 5 model release (9 June 2026). These 2026 figures are a same-window slice, not an annualised projection dressed as a full year.

4.5 Who is founding these companies (United Kingdom depth)

Where the register exposes ownership and control (the UK PSC and officer registers), the composition of AI-named founders shows a structural break in 2023. These composition findings are robust to survivorship (they describe the surviving cohort, a fair sample of who is building).

• Solo and founder-led. The single-founder share of new AI-named companies rose from about 68% to over 80%; about 96% of AI-named companies incorporated since 2020 are founder-led (the same individual, matched on surname and date of birth, on both the owners and directors registers).
• Younger. The under-26 share of AI founders rose from 4.1% (2018) to 12.8% (2025), faster than the comparison group; the crossover from older-than-baseline to younger happened in 2023. People born in 2000 or later own UK AI-named companies at about 2.2 times the rate they own other UK companies.
• More international. Among AI-named companies incorporated in 2023 or later, 35.7% of owners hold a foreign nationality (up from about 31% before 2022), led by Indian, Chinese and Pakistani nationals, who are younger on average than UK-national founders. This is the nationality of owners of UK-registered companies, predominantly UK-resident; it characterises who incorporates here, not migration flows.
• Branded for a global, AI-native audience. The .ai domain is now widely adopted alongside .com and .co.uk, over-represented relative to the wider company population.

The mechanism is straightforward: before ChatGPT, building an AI company meant a team that could train a model; after ChatGPT, one person with an API key can ship an AI product. The data shows the floor on entrepreneurship dropping, and a younger, more solo, more international cohort walking through the door.

4.6 Occupational exposure: the supply side mirrors the demand side

The single most striking cross-cut connects firm creation to the labour-exposure literature. Mapping the website-derived verticals to occupations and to the Massenkoff and McCrory (2026) exposure scores, every vertical that maps cleanly corresponds to a high-exposure occupation, with a count-weighted mean exposure of approximately 65%: computer programming (74.5%), medical records (66.7%), marketing (64.8%), financial analysis (57.2%), and information security (48.6%). AI-named company formation concentrates in the same occupations that an independent measure rates as most exposed to AI.

We read this as a supply-side counterpart to the demand-side exposure literature: people are founding AI companies aimed squarely at the work that AI most affects, in Anthropic's own occupational language. The correspondence is correlational; the two datasets are independent (the formation data is ours; the exposure scores are Anthropic's) and we identify no mechanism. The crosswalk is country-agnostic, so this overlap is available for every country in the study, not the United Kingdom alone: across the enriched countries, most cluster near 30% of companies in Anthropic's most-exposed occupations.

5. Discussion

The results describe a change in firm creation coincident with, but not shown to be caused by, the diffusion of consumer generative AI. The most parsimonious reading is that the technology lowered the fixed cost of founding an AI company, and that the marginal entrant under the lower cost is younger, more often solo, more often the sole owner-operator, and more often a foreign national. That the new firms cluster in the occupations independent measures rate most AI-exposed, and that the young cohort facing slower hiring into those occupations is the same cohort over-represented among founders, is consistent with entrepreneurship operating as a margin of adjustment to automation. We caution that this links two independent datasets and is not evidence of a causal channel. What the data establishes cleanly is the shape: a several-fold rise in AI's share of new companies, shared across roughly thirty jurisdictions, inflecting in 2023, re-accelerating in 2026, and specific to AI rather than to company formation in general.

6. Limitations

• Name-based proxy. We measure companies that name themselves for AI, a clean, reproducible lower bound that misses brandable AI companies (Wayve, Synthesia) and admits a small residual of false positives after de-contamination.
• Survivorship. Live-register snapshots undercount older years; we lead with the share and official true-births and flag raw-count multiples as upper bounds (section 3).
• Cross-country level comparability. Registers count different universes (companies vs all entities vs establishments); we compare the share multiple across countries and the share level only within clean company registers. France and Brazil, whose denominators are establishments and MEI respectively, are used for their trend only, not their level.
• United States gap. The United States has no national company register; it is the largest gap in a global claim and is pending an open-data decision from OpenCorporates.
• Website coverage. The activity classification and the occupational crosswalk depend on website enrichment, which does not cover the entire cohort and is excluded where low-confidence.
• Descriptive, not causal. We document timing, composition and overlap; we make no causal claim.

7. Conclusion

Since ChatGPT, the share of new companies that are explicitly AI-named has risen several-fold in every register with usable coverage, reaching about one in two hundred in the United Kingdom and Ireland by 2025, with a shared 2023 inflection and a 2026 re-acceleration. The rise is specific to AI: overall company formation was roughly flat. The new firms cluster in the occupations independent measures rate most exposed to AI, so the supply side of company creation mirrors the demand side of labour exposure. And in the country where we can see inside the cohort, the founders are younger, more solo, more international, and own what they build. The cost of starting an AI company collapsed, and a measurable, global, and still-growing wave of people walked through that door.

References

The occupational AI-exposure scores used in section 4.6 come from the Anthropic Economic Index; the concept of scoring occupations by exposure to large language models originates with Eloundou et al.

• Appel, R., Massenkoff, M., McCrory, P., McCain, M., Heller, R., Neylon, T., and Tamkin, A. (2026). Anthropic Economic Index report: economic primitives. Anthropic.
• Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2023). GPTs are GPTs: An early look at the labor-market impact potential of large language models. arXiv:2303.10130.
• Handa, K., Tamkin, A., McCain, M., et al. (2025). Which economic tasks are performed with AI? Evidence from millions of Claude conversations. Anthropic.
• Massenkoff, M. and McCrory, P. (2026). Labor market impacts of AI: a new measure and early evidence. Anthropic. Primary source for the occupational AI-exposure scores in section 4.6.

Data sources

• Companies House (United Kingdom): bulk company data product and REST API (company register, officers, persons with significant control).
• Companies Registration Office (Ireland): company register.
• Roughly thirty further national company registers (bulk downloads), held in full for about twenty-four jurisdictions.
• Official gross-incorporation series: Companies House (UK), ASIC (Australia), INSEE (France), Tokyo Shoko Research (Japan).

Data availability

Aggregate, chart-level data underlying this draft is available in the project repository and on the International and Countries pages. Person-level officer and PSC records are not redistributed, as they constitute personal data and their bulk republication would exceed the relevant open-licence terms.

Methods appendix (to expand)

• Per-country false-positive rates and the language-aware de-contamination classifier.
• Per-country data currency and cutoffs.
• Official gross-incorporation ("true births") sources and the universe each one counts.
• Entity-type and companies-only adjustments per register.