The next durable AI companies sell into markets where the bottleneck is permission, not the model.
In regulated industries that permission is certification, liability and a buyer's trust; in physical AI it is the safety case, procurement and real-world data that cannot be scraped. The moats are the same in both: compliance work competitors must repeat, proprietary operating data, and slow buyers whose switching costs run for a decade. Both have to pass through the same layer that grants the permission: evaluation, security, simulation and certification. And before any of it, the models have to be taught the work by the people who already hold the permission.
41 companies across five orbits, at least $10B raised between them. Two have agreed to be acquired this year.
Edition 6, 16 November 2026. Next check 23 November 2026.
Size is capital raisedRound in the last six weeksAgreed or completed exit
This edition
What changed since the last check, each line linked to the page that says so.
RoundMercor raised $750M (Series D) at $20B, led by Not named; company did not name a lead investor. Participants include Benchmark, Felicis, Fundomo, funds affiliated with Apollo, G Squared, General Catalyst, Glade Brook Capital, NVIDIA, Salesforce, Sapphire Ventures, StepStone Group, XTX Ventures, and ZoomNov 2026 mercor.com
Week to 16 November 2026: 41 companies checked; 1 change applied, 0 held for a human because their source didn't say so.
The orbits
Each orbit is a sector in the thesis file. A fund with a different thesis gets different planets.
Regulated vertical AI
AI built for licensed professions and regulated workflows: health, legal, tax and accounting, financial services, veterinary and dental.
The permission layer: evaluation, AI security, governance and compliance, simulation and certification that regulated and physical AI must pass through.
4 companies, $101M raised
Complyance, Cogent Security, Pillar, BeyondMath
Training data and evals
Expert data, RL environments and evaluations that teach and test frontier models on professional work. The same doctors, lawyers and bankers whose sign-off grants the permission are now the ones teaching the models.
Helsing and Harvey hold 32% of the capital recorded on the map. The log scale keeps the seed and Series A companies, where the thesis is aimed, readable beside them.
Total disclosed capital raised, converted to dollars at approximate rates for placement only. Hollow dots are the latest round alone, where earlier rounds aren't yet totalled; brass dots are exits. Hover a dot for the figure as reported.
Scout's first shortlist of new candidates arrives with the first scheduled run.
How it stays current
The model proposes; the source decides. A change is applied only when the page it cites says so, and anything else waits for me to check.
One interests file
A belief, the orbits, the signals worth trusting, how to score, and any companies to leave off. Swap the file and every step below re-orients.
Refresh, each week
Each tracked company is searched for a new round, an exit or a shutdown since it was last checked. Pending names get their first figures.
Source check
A change is applied only if the page it cites states the amount or the acquirer right next to the company's name. Anything else is held for a person; 1 are held now.
Scout and radar
Web searches shaped by the interests file, plus up to twelve other sources (some need a free key), find new candidates and score them on its rubric. They go to an inbox, never straight into the map.
Publish
The tracker and a dated edition are committed to the repository. This page reads them within the hour.
Schedule
Mondays, early morning UK time
Figures matched to their source
35 of 42, 16 Nov 2026 (7 sources couldn't be fetched)
Lenses applied to new candidates
Permission test, Moat taxonomy, Real-world data loop, Services as software, Who buys it, Procurement path
Data published
16 Nov 2026
Make your own map
Everything here comes from one file of interests. If you'd like a map of your own, fork the engine, add a model key (Anthropic, OpenAI or OpenRouter), rewrite config/thesis.yamlwith your own orbits, and point a copy of this page at your fork's data/observatory.json.
Edition 616 November 202641 checked, 1 applied, 0 heldMercor raised $750M (Series D) at $20B, led by Not named; company did not name a lead investor. Participants include Benchmark, Felicis, Fundomo, funds affiliated with Apollo, G Squared, General Catalyst, Glade Brook Capital, NVIDIA, Salesforce, Sapphire Ventures, StepStone Group, XTX Ventures, and Zoom.
Edition 59 November 202641 checked, 0 applied, 1 held
Edition 45 November 202641 checked, 1 applied, 0 heldCrosby, the AI-native law firm, bid for Weil Gotshal's data and brand; terms weren't disclosed.
Edition 326 October 202641 checked, 0 applied, 0 held
Edition 220 October 202643 checked, 5 applied, 0 heldLabelbox, which runs the Alignerr expert network, raised more than $100M in a Series E led by Infinitum Partners. Halluminate raised a $30M Series A led by Oak HC/FT to build finance training environments. micro1 reportedly raised more than $100M at a $4B valuation. New orbit: Training data and evals, the experts and environments that teach the models (11 companies).
Edition 113 October 2026Re-oriented to the current thesisWexler agreed to be acquired by Legora. Workday agreed to acquire C1, the identity platform for people and AI agents. Harvey raised $550M at $15.6B, co-led by Lightspeed and Diffusion. Comand AI raised a €32M Series A led by Blossom Capital, with Saab as a strategic investor.