Mechanize is the smallest company on this list and possibly the most important one, because it is the clearest test of a proposition the rest of the vertical cannot make: that you can sell the artefact instead of the hours.
Founded in April 2025 by Tamay Besiroglu, Matthew Barnett and Ege Erdil — all from Epoch AI — it raised a $9.1M seed at roughly $500M, with Nat Friedman, Patrick Collison and Dwarkesh Patel among the backers. It builds RL environments and coding evaluations, and it employs about 35 people.
The 103 days
In August 2026 Mechanize was reported to be in talks with Google at $1.5B+, structured as a technology licence plus the hiring of select staff — not an acquisition (TFN). The research notes date the talks at 103 days after the seed round.
If it closes, that is roughly a 3x step-up in a year on a company with no disclosed revenue, and about $43M per employee (TNW). It would be the largest value-per-head event in the vertical by a wide margin.
No close has been confirmed. Talks were reported in August 2026 and nothing since. Mechanize's revenue, customers and contract structure are all undisclosed. Everything on this page rests on a single reported deal that may never happen.
Why the structure is the story
Google is not buying Mechanize. It is licensing the technology non-exclusively and hiring some of the people. The stated precedents in the coverage are Windsurf (~$2.4B, July 2025) and Character.AI (2024).
That structure is a direct response to what happened to Scale AI. Meta paid $14.3B for 49% of Scale and destroyed most of Scale's frontier-lab revenue within weeks, because in a market where the buyers compete with each other, neutrality is the product. A licence-and-hire takes the capability without triggering the same collapse — and, not incidentally, without triggering merger review. See One customer is a binary event.
Why it matters to the vertical's margin problem
Every hours-based business in this market runs at a 27–40% take. Mercor's leaked gross margin is 27% rising to 33%; Prolific publishes 42.8%; Appen audits at 40.3%. That is a staffing margin and What the public market pays for labour shows what the public market pays for it.
Environments break the pattern in two ways. Epoch reports contracts at six to seven figures per quarter, a website replica at ~$20,000, a Slack-grade product clone at ~$300,000, individual tasks mostly $200–$2,000 — and, decisively, that exclusive deals price at roughly 4–5x non-exclusive (Epoch AI). A buyer paying 4–5x for exclusivity is paying for denial, not for data. That is a fundamentally different good from an hour of a doctor's attention, and it is the same insight behind micro1's claimed 80–90% margin on resold datasets and Mercor's acquisitions of Sepal AI and Deeptune.
The demand is real: Anthropic reportedly discussed spending over $1B on RL environments alone over a twelve-month period (TechCrunch).
The bear case, from the builders
The people closest to this are not uniformly bullish (TechCrunch):
- Ross Taylor (General Reasoning): "I think people are underestimating how difficult it is to scale environments."
- Sherwin Wu (OpenAI): sceptical of environment startups — the space is crowded and research moves too fast for third parties to serve labs.
- Andrej Karpathy: "I am bullish on environments and agentic interactions but I am bearish on reinforcement learning specifically."
- Will Brown (Prime Intellect), on the other side: "RL environments are going to be too large for any one company to dominate."
Epoch adds the structural threat: product companies — Salesforce, Slack, Shopify — are partnering with labs directly on environments for their own software, which cuts the vendor out of the highest-value enterprise surfaces entirely. And What better models do to each layer cuts both ways here: if in-house teams at labs and neolabs keep vertically integrating, the licence-and-hire is not a template for building a business. It is a template for being absorbed.