Invisible is the least marketplace-shaped company in the expert data vertical, and consequently the one whose numbers can be read at face value.
It was founded in 2015 as a personal-assistant bot routing tasks overseas, grew through what its coverage describes as twentyfold revenue increases before taking venture money, and became a training-data supplier to OpenAI — running RLHF work since 2022. Matthew Fitzpatrick, ex-McKinsey, succeeded Benjamin Plummer as CEO in January 2025 (Sacra).
It discloses profit, which almost nobody here does
| Year | Revenue | Growth | EBITDA |
|---|---|---|---|
| 2023 | $60M | — | not disclosed |
| 2024 | $134M | +123% | ~$15M (11% margin) |
Source: Sacra.
An 11% EBITDA margin on 123% growth is an unremarkable services profile and a completely legible one. It is also more than Mercor, Surge AI, Scale AI, Handshake AI or Turing have ever put on the record about their own profitability. Mercor disclosed $6M of profit in H1 2025 and nothing since; the rest disclose nothing at all.
The $100M round led by Vanara Capital, a TPG spin-out, in September 2025 at $2B+ (BusinessWire; SiliconANGLE) prices that at roughly 15x 2024 revenue, up from about 8.3x at the $500M mark.
That 15x is worth sitting with. Innodata, a public company doing the same category of work at a ~40% gross margin (49% adjusted in Q2 2026) and 40%+ growth, trades at 6.2x. Appen trades at 0.9x. Invisible is being marked at more than double the best public comparable on numbers that are structurally similar. See What the public market pays for labour.
The staffing ratio tells you what it is
3,000+ agents across 35+ countries against 350 full-time employees (Sacra). That is roughly one FTE per nine agents.
Compare: Surge AI runs ~130 FTEs against ~50,000 contractors, and Mercor ran 30 US FTEs plus 20 contractors in India at $50M of ARR. Invisible carries an order of magnitude more internal staff per unit of supply, which is the signature of an operations-as-a-service business rather than a marketplace — it runs the process, it does not merely match. Marketplace, staffing firm, BPO or agency is the distinction, and it has direct consequences: higher switching costs and longer contracts, but the revenue is booked closer to gross-with-delivery-cost and the margin is capped where BPO margins are capped.
The customer list reflects that: OpenAI, Amazon, Microsoft, Cohere and DoorDash. DoorDash is not a frontier lab, and the presence of non-lab enterprise buyers is the one thing on this page that mitigates the One customer is a binary event risk running through the rest of the vertical.
M&A
Invisible agreed to acquire WeCP, a technical-assessment company, in March 2026; terms were not disclosed. That is the same move Handshake AI made with Cleanlab and Mercor made with Sepal and Deeptune — buying verification and quality tooling rather than more supply, which is a reasonable inference about where these operators think their cost problem is. See Who is actually on the other end.
No 2025 or 2026 revenue figure for Invisible could be found. The 15x multiple above is computed on a FY2024 number against a September 2025 valuation, so it overstates the multiple if the business kept growing at anything close to 123%.
Take rate, contractor pay rates and customer concentration are all undisclosed. Because Invisible employs or directly contracts its operators rather than running a payout marketplace, the gross-versus-net question is less acute here than elsewhere in the vertical — but the revenue basis has still never been stated explicitly.