Methodology

Sources

• Y Combinator directory, via the open yc-oss dataset, refreshed daily. Gives the largest consistent set of startups with locations, founding cohorts and tags. Only active and public companies appear on the map; closed and acquired ones still count in the founding trends.
• Unicorns: Wikipedia's List of unicorn startup companies (CC BY-SA 4.0), with the latest reported valuation and its date. Companies already on the map get a valuation badge; the rest are added with a headquarters city checked by hand where we are confident of it.
• Curated research labs: frontier labs, corporate research groups, academic institutes and safety organisations, added and checked by hand.
• Startup mafias: companies founded by alumni of the same employer. Each link comes from public reporting (for example Sifted and UKTN, using Accel's alumni counts) or the founders' own published histories. We only name a founder when the link is documented.
• Communities: builder communities, accelerators, campuses and fellowships, curated by hand and shown as their own layer.
• Submissions from companies and labs, reviewed before they go live.
• Wikidata (CC0): companies founded since 2008 that list a tech industry, a website and a headquarters with coordinates, refreshed weekly. The stated industry decides the field. Wikidata only adds companies or fills gaps; it never overrides YC or hand-checked details.
• Hugging Face: download counts for the 10,000 most used open models, grouped by publishing organisation, refreshed weekly. Personal accounts are left out.
• Companies House (Open Government Licence v3.0): the monthly snapshot of every live UK company, used to count newly registered companies with AI in their name under software, IT, data or R&D activity codes. Counts only; no individuals' details are stored.
• OpenAlex: AI papers by country and year, for the research comparison.
• Funding: rounds announced in public startup news, read from headlines and stored with a link to each source article. Undisclosed amounts are left out, so totals are a floor.
• Places: city coordinates from GeoNames (CC BY 4.0). Pins are placed at city level and spread around the city centre; a company that claims its listing can set its exact office.

Fields and AI-native

Each listing counts in one field. Specific domains (legal, health, finance, climate, hardware, security, education, property, quantum) take priority, so an AI contract tool counts as Legal. Separately, a company is AI-native when its tags or its own description say AI is central to the product. That is why AI-native companies (43.6% of startups) outnumber the AI & models slice.

City score

Score = 40% × the city's share of the world's AI-native startups + 30% × its share of research labs + 30% × its share of new AI startups founded in 2025–2026. The highest-scoring city is set to 100 and the rest scaled to match.

Growth and momentum

Growth compares AI startups founded in 2025–2026 with those founded in 2023–2024, using founding cohorts. Momentum is that change divided by the earlier count, with a floor of three so a city going from zero to one does not top the list.

Known bias

YC is US-centred, so the United States and the Bay Area are over-represented compared with ecosystems YC rarely funds. Each extra source, and each submission, narrows this. Read country comparisons as a lower bound outside the US.

Open models

Hugging Face reports how often each model was downloaded in the last 30 days. We take the 10,000 most downloaded models, add up their downloads by the organisation that published them, and keep organisations rather than personal accounts. Where we know the organisation, it links to its listing and counts towards its country; the rest show as not yet placed. A download is a file fetch, so servers and pipelines count as well as people: it measures reach, not users.

Logos

Company names and logos belong to their owners. We show a logo only on that company's own listing, to identify it, and never to suggest the company endorses Labspie.

We use a logo the company published itself: one it uploaded when claiming its profile, the one it gave the Y Combinator directory, or the icon its own website declares (including its web app manifest and favicon). For companies whose site turns automated requests away, we use the logo file their Wikidata entry links to on Wikimedia Commons, which only hosts freely licensed or uncopyrightable files; the company page credits that file and its licence. Our fetcher identifies itself as LabspieBot, follows robots.txt, takes one small image per company, and serves a resized copy from labspie.com so we never load files from anyone else's servers. Where no suitable logo exists we show none rather than make one up.

Want your logo removed or replaced? Email privacy@labspie.com from your company domain, or claim your profile and upload the one you prefer. Removals take effect at the next morning's refresh.

Removals and personal data

Listings describe companies, built from public sources. Where a profile names people, such as founders, it uses what the company or a public source already publishes. To remove a listing, a founder name or photo, or anything else about you, email privacy@labspie.com with the page link. We act on removal requests within a few days and keep a removed company out of future refreshes.

For anything else, including corrections, write to hello@labspie.com. For advertising, write to ads@labspie.com.

Back to the Index · 8,547 listings in this build.