# After ChatGPT, AI companies' share of new companies tripled; counted by the older AI tags, it rose about a fifth

Companies on the Metix AI Platform that describe themselves as AI companies, by founding year from 2010 to 2026, with where they are, what they build, their size and funding against all companies, and where the staff of the largest came from, on September 23, 2026.

## Findings

ChatGPT launched on November 30, 2022, according to OpenAI's announcement. Of the 1,094,649 companies on the Platform founded in 2023, 2024, or 2025 (nonprofits, educational institutions, and government agencies left out, as the definition leaves them out), 17,871 (1.63%) are AI companies by this case's definition, against 10,553 of the 1,879,163 founded from 2019 to 2022 (0.56%): 2.9 times the share. By founding year the share went from 0.69% in 2022 to 1.49% in 2023, 1.69% in 2024, and 1.78% in 2025 (1.70% for 2026 so far). Every recent year holds fewer companies of every kind (474,723 founded in 2023, 308,977 in 2025, 54,296 in 2026 so far) because a new company takes time to appear, so the share of each year's foundings is the measure, not the count.

The class describes itself in new words. The definition has two branches: AI in the company's name, and technical AI tags. Taken alone, the three technical tags that were in use long before 2022 (deep learning, computer vision, natural language processing) rise only 1.18 times, from 0.29% of 2019 to 2022 foundings to 0.34% of the class; by year they go from 0.28% in 2022 to 0.35% in 2023, below the 0.38% of 2017 and 2018. Adding generative AI and large language models, the technical branch rises 2.03 times (0.45% to 0.90%). Companies with AI in their name and an AI tag rise 5.69 times (0.16% to 0.90%) and are more than half of the class (9,813 of 17,871). The probe's four terms rise 1.60 times (1.99% to 3.19% of all companies, since they keep every type) and the nine loose terms 2.16 times (3.32% to 7.18%). The audit is what decides between them: by hand reading, the definition catches about the same share of real AI companies in every era, while the name alone catches more of the newer ones and the older tags fewer of them. So the share of real AI companies among new companies did rise about 2.9 times, and the class is a different kind of AI company, one that names generative AI, language models, and agents rather than the older technical terms.

The class is concentrated where the last cohort was, a little more so, under every definition. The United States holds 38.4% of the class's companies that give a country (5,671 of 14,771), up from 31.4%; India 12.0% (from 13.3%); the United Kingdom 8.2% (from 6.6%); the United Arab Emirates 2.2% (from 1.4%), from fifteenth to seventh. Of class companies that give a city, 10.4% are in the San Francisco Bay Area (from 8.4%), 5.8% in London (from 4.4%), 4.4% in New York City (from 4.0%), and 1.9% in Dubai (from 1.2%); Bengaluru (2.9%, from 3.2%), Tel Aviv (0.5%, from 0.9%), and Berlin (1.0%, from 1.2%) fell. The Bay Area's rise holds under the probe (6.2% to 9.4% of companies with a city; the city of San Francisco alone 3.3% to 6.5%) and the loose terms (5.5% to 7.1%); London's holds too, more weakly (probe 4.5% to 5.2%, loose 4.7% to 5.4%). Shares of the whole cohort are lower because 27% of the class gives no city, against 19% before. China is nearly absent: the Platform holds 820 companies of any kind headquartered in China founded in 2023, 8 of them AI companies.

What they say they build moved from models to agents, and the move holds in the loose population as well. Among AI companies founded in 2016, 57.1% tag machine learning and 34.4% computer vision; among those founded in 2025, 18.9% and 8.9%. Among all companies matching the loose terms, machine learning falls from 38% to 12% and computer vision from 6% to 2% over the same years, so the drop is not only the name branch entering without those tags. AI agents or agentic rise from 5.4% (2016) and 7.2% (2021) to 20.5% (2025) and 27.3% of the 867 founded in 2026 so far; in the loose population, from 3% to 14% and 18%. Generative AI peaks with the 2023 class (42.4%; 11% in the loose population) and falls after. LLM tags stay near 10 to 12% of the class. Robotics, fintech, and developer tools stay at 2% or less.

The class grows fast, but so does every young company. Among class companies with 11 or more staff that state a growth figure, 41.7% of AI companies at least doubled headcount in a year (695 of 1,667), against 30.9% of all other class companies of that size (12,737 of 41,166); for the 2019 to 2022 cohort the two are 8.9% and 8.1%. Most of the gap between the eras is age; the part that is AI is about 11 points in the class and about 1 point before. AI companies are less often a single person: 11.9% of AI companies founded in 2025 list only the founder, against 21.4% of other companies founded that year, while 72.7% and 64.8% are in the 1 to 10 band.

AI companies have a recorded round far more often than other companies, and the rounds are about the same size. 11.3% of the AI class has a latest round with a date (2,023 of 17,871), against 1.3% of other class companies (14,188 of 1,076,778); for 2019 to 2022, 25.1% against 2.9%. Among United States class companies with an amount, 43.1% of AI companies last raised under 1,000,000 and 45.3% of other companies did; 4.7% of AI companies and 4.5% of others raised 50,000,000 or more. Latest rounds dated in each quarter for class AI companies rose from 45 in 2023Q1 to 228 in 2025Q3; rounds stop after 2026Q1 because the records were last refreshed in April and May 2026. Twenty-one United States AI companies founded since 2022 with 50 or fewer staff have a latest round of 50,000,000 or more that passes the checks, in `data/lean.json`; 17 are in the class and 4 were founded in 2022 (Black Ore, Deep Apple Therapeutics, Protect AI, Slingshot AI). Among them are Mind Robotics and Ricursive Intelligence (250,000,000 or more), WhiteFiber, Inferact, and General Intuition (100,000,000 to 250,000,000), and Nous Research, Essential AI, and Protect AI.

The staff of the largest new AI companies came from big tech somewhat more often than the staff of the largest older ones. Of 9,792 profiles visible through the Metix AI Platform with a current job at 93 of the class's 150 largest AI companies (each classified by hand as an AI company), 747 (7.6%) had an earlier, non-internship job at Google, Meta, Microsoft, Amazon, Apple, or NVIDIA, against 2,257 of 40,097 (5.6%) at 87 of the 150 largest AI companies founded 2016 to 2019. Leaving out the data-labeling marketplaces, whose visible staff are mostly contractors (Soul AI and four others in the class; Scale AI and six others before), the shares are 8.1% (722 of 8,949) and 5.6% (1,865 of 33,263). Google (2.4%) and Amazon (2.3%) lead in the class. Frontier labs are a small source: 34 class profiles (0.35%) came from OpenAI, Anthropic, or DeepMind, 22 of them from DeepMind, and fewer than 10 each from OpenAI and Anthropic. 288 class profiles (2.9%) have founder or co-founder in their current title, against 1.1% at the older companies, which have larger staffs.

## Population and definitions

A company counts as an AI company when its type is not Nonprofit, Educational, or Government Agency and either its name has the word AI and its keywords match one of nine loose AI terms (AI, artificial intelligence, generative AI, machine learning, large language models, LLM, deep learning, computer vision, natural language processing), or its keywords match one of five technical terms (deep learning, computer vision, natural language processing, large language models, generative AI). The full condition, the comparison definitions, every test, and the audit are in [`queries/definition.json`](queries/definition.json). Cohorts are by founded_year, which is a year only: 2019 to 2022 is before (the launch came in the last month of 2022), 2023 to 2025 is the class, and 2026 is partial ([`queries/cohorts.json`](queries/cohorts.json)). wave.json counts all companies founded in a year twice: every company (all_count, used for the lag and for the probe and loose series, which keep every type) and every company of the types the definition keeps (all_typed_count, the base of every other share). The controls in controls.json exclude the same three types, and their cohort totals equal the summed all_typed_count. A total of 100,000 or more comes back banded, so such totals are sums of disjoint parts by size band and, inside a banded band, by headcount and country.

Places use the headquarters: country compared whole, metros as a country, a list of cities, and for ambiguous names a state that is named or missing ([`queries/places.json`](queries/places.json)). Themes are keyword term lists, overlapping ([`queries/themes.json`](queries/themes.json)). Size, growth, funding bands, the controls, and the small-team rules are in [`queries/measures.json`](queries/measures.json); the staff join in [`queries/staff.json`](queries/staff.json).

## Method

The definition was audited before any measure, by reading records in whole slices: every match in one founding year with linkedin_followers in a narrow window, one slice per year from 2016 to 2025, each company classified by hand from its name, industry, and tags (AI is the product; AI is part of the core product in another field; a generalist firm listing AI among many services; not AI or not a company). A first round read 325 companies matching the nine loose terms: 42.8% were AI companies, and the probe's four terms 50.3%. Tags are a free list, generalist software houses and agencies list AI terms, and a term can match words from two different tags (deep technology innovation plus e-learning matches deep learning). keywords eq counts the same as match for every term tried, so no exact-tag rule exists. The name alone was 97% precise but kept 5 of 40 AI companies founded 2016 to 2019 and 18 of 65 founded 2023 to 2025. The chosen definition kept 17 of 40, 15 of 34, and 26 of 65, about 42% in every era. A second round read 215 fresh companies matching the chosen definition: 167 (77.7%) are AI companies, 77.5%, 74.2%, and 81.3% by era. The target was 90%, and it was not reached. The false positives are spread across industries (IT services, consultancies, agencies, software houses, a few funds and media sites); the best industry exclusion lifts precision to 84.5% on the fresh reads and loses a quarter of the AI companies. Of the 260 loose-term matches the definition leaves out, 81 (31.2%) are AI companies. The definition is a precise core, not a census. It was kept over the alternatives because it is the only candidate whose precision and catch both hold still across founding eras, which a before and after share needs: the probe's terms grow more precise and catch fewer AI companies in later years, so they understate the rise (1.60 times), and the name alone catches more in later years, so it overstates it (5.69 times). The two bracket the definition's 2.91 times. Every candidate tested, with its count, precision, catch, and era split, is in `queries/definition.json`.

Everything else is counts. `fetch.py` counts, for each founding year, AI companies, all companies, and five comparison definitions; the lag signals; 60 candidate countries per cohort (the top 15 published, the rest and no country completing each cohort); 18 metros, and the Bay Area, San Francisco, and London under the comparison definitions; 15 themes for each year 2016 to 2026, inside the definition and inside the loose population; 22 industry values; size bands per year, growth bands, funding coverage, quarters, and amount bands; and the same size, growth, funding, and amount measures for all companies. Each theme was audited on a small whole slice (158 companies); five terms that matched across tags or marked service firms were removed, as `queries/themes.json` records. The growth figure was checked on 46 records (every value a plausible percentage from two whole headcounts a year apart).

For the staff, `fetch.py` reads the AI companies of each cohort above a headcount floor, takes the largest 150, and joins only those classified by hand, with the scheme of the definition audit, as an AI company or a data-labeling marketplace. The classification (in `queries/staff.json`) left out 22 class companies and 38 older ones as generalist IT, staffing, and outsourcing houses or as not AI companies (a content page, a certification body, a training institute, a lender), besides divisions, labs, and communities that are not companies (Microsoft AI, a bank's lab, an industry council; a defense contractor in the older cohort). The Platform cannot filter people by company id, so the join is by name: names that return more than 1.5 times the company's headcount were dropped as shared (25 in the class, 21 in the older cohort, some of them real companies with common-word names such as Decagon and Cohere), and a whole slice of the kept companies' staff was read with company ids: 176 of 181 class profiles (97.2%) and 231 of 244 older-cohort profiles (94.7%) carry the right company's id, and Dify, whose four sampled profiles were all at other Difys, was dropped. An earlier job is a job entry at the employer that is not current and not an internship. Where a figure without the labeling marketplaces differs from the figure with them by 1 to 9 people, it is withheld.

The replay made 2,546 calls, read 762 records, and cost 2,481 API Credits (`data/receipt.json`). Exploration cost 10,481 API Credits: 1,256 for the probe (27), the audits, and the tests; 1,639 for a first full replay without the comparison series and controls; 823 for a replay that stopped when a connection dropped; 2,255 for a full replay before the staff companies were classified by hand; 2,172 for a full replay before the shares were set against companies of the same types; and 2,336 for a replay that stopped at an old API Credit ceiling in `fetch.py` before writing anything.

## Limits

About one in five companies the definition counts is not an AI company, and it keeps about four in ten of those that are; counts are a floor. The size of the post-ChatGPT rise depends on the words counted: by names it is 5.7 times, by the older technical tags 1.2 times; the audit's steady catch across eras is why this case uses 2.9 times, and the audit is one reader's classification of a few hundred companies from names and tags. Tags describe a company today, so an older company that took up AI later counts in its founding year, which raises the earlier years and, if anything, understates the rise. founded_year is a year, so the class cannot start on the launch date, and 2022 includes one month after it. Records were refreshed in April and May 2026, so 2026 covers four or five months, 2025 is incomplete, and no latest round is dated after 2026Q1 in any number. Growth compares ages as well as eras; controls.json shows how much. Industry values changed names at some point (Information Technology & Services holds 417 of the older cohort and 23 of the class), so industries.json compares badly across cohorts. Headcount and growth are what company pages report; the small-team list drops implausible rounds by rule rather than by checking each one. last_funding is the latest round only, and amounts outside the United States are in local currency with no currency field (South Korean teams of 8 to 36 people show rounds of 3,000,000,000 to 60,000,000,000), so amounts are compared for United States companies only. The Platform holds few Chinese companies. The country list is 60 candidates; 521 class companies are in countries outside it, so a country outside the list could in principle rank in the top 15, though no candidate below the 15th comes close. Four themes are also entry terms of the definition, so their level partly reflects how it was built; the loose-population shares beside them do not have that problem. Staff counts are profiles visible through the Metix AI Platform, joined by name, at the companies that passed the hand classification; it is one reader's judgement from names and tags. A profile left out of date still counts as current, so a count is neither a headcount nor a guaranteed floor.

## Sources

OpenAI, "Introducing ChatGPT", November 30, 2022: [openai.com/index/chatgpt](https://openai.com/index/chatgpt/). The page refused automated reads during this study; the date was confirmed from OpenAI's own indexed page.

## Rerun

```bash
export METIX_KEY=metix_xxxxxxxxxxxx   # create one at https://platform.metix.ai/api-keys
python3 cases/class-of-2023-ai-companies-2026/fetch.py
```

To make it with an agent instead, use [`PROMPT.md`](PROMPT.md).