Skip to content
aiwork.fyi

About

Methodology

aiwork.fyi is a public record of the AI training-data industry. Version 1 does two things: sourced profiles of the vendors, and anonymous reports of what workers actually earned.

Research date

Vendor profiles and the “What we know” notes on the home page are from a research pass completed on Oct 7, 2026. They are not updated live. If a newer filing contradicts a cell, the correction belongs in the cell with its source, not as a rounded replacement.

What a vendor profile contains

Each company page repeats the sourced wording for what the company sells and to whom, revenue or run rate, valuation, and worker pay range. The words are kept as reported, including disagreement inside a single cell. Scale’s 2025 revenue is one example: a company figure via Forbes and a conflicting Sacra estimate both stay in the sentence.

This site does not audit the underlying accounts. A link means the figure is attributed to that page, not that we re-derived it.

Labels

  • [reported] A figure attributed to a named report or outlet. Not independently audited here.
  • [unverified] A circulating figure we could not tie to a filing or the company. Aggregator estimates are marked this way.
  • [est.] An estimate, including a third-party estimate. Sometimes written with the estimator, as in [est., Sacra].
  • [company] The company’s own figure, including when it reached us through a news story.
  • Not public means this research pass did not find a public figure. It is not a claim that the number is zero.

Other brackets, such as [Guardian] or [WIRED], name the outlet attached to that clause.

Sorting

The vendor table can be sorted by revenue, valuation, or pay. Sorting needs a single number, and many cells contain several. The ordering aid is the latest headline revenue in the cell, a closed or last-priced valuation, and one hourly figure from the pay cell (a stated average, otherwise the midpoint of the generalist or entry band, otherwise the stated minimum).

That number is not shown, and it is not a new estimate. Cells without a comparable figure sort last. TELUS Digital’s 2024 total is not used as AI-training revenue, because AI is not broken out. Surge’s valuation talks and micro1’s unconfirmed $4B are not treated as closed prices. “Public,” with no market cap recorded here, is not a valuation.

How the home-page totals are counted

The two figures under the hero are sums of phrases in the vendor profiles. Change the phrase, and the total changes with it. The tilde on the valuation means the inputs include approximations, a greater-than, and a deal-implied price. The plus on the run rate means at least one input is a floor.

Combined valuation, ~$49B

Private companies with a priced round or a deal-implied valuation. The amount added is the figure in each phrase.

Up to ~$78B, counting reported talks

These are not closed prices. They are added only to the upper bound, at the high end of what was reported. Surge is counted at the Bloomberg figure of at least $25B, not the Reuters figure above $15B and not the Forbes estimate of about $24B. micro1's $4B is unconfirmed. Mercor's talks at $20B are not in this upper bound; only the closed $10B Series C is in the ~$49B.

Not in the valuation total

Public companies are left out. No market cap is recorded here, so none is invented for the sum.

TELUS Digital is left out. The price in the profile is for the minority stake, not a company valuation used in this total.

  • TELUS Digital — Taken private by TELUS, Oct 2025 (~US$539M for the minority stake)

No priced or deal-implied valuation in this research pass:

Combined run-rate, $6.7B+ a year [est.]

An estimate of what these vendors bill AI labs, using each company's latest figure in the list below. It mixes gross and net figures from different dates. Where a cell names more than one number, only the latest headline figure is added: Surge's 2024 $1.2B [reported], not the 2025 $1.4B [unverified, aggregator]; Scale's company figure of just shy of $1B, not the conflicting Sacra estimate; Appen's FY25 revenue of US$230.8M, not the FY26 guidance; Invisible's 2024 $134M, not the 2023 $60M.

These companies publish a revenue figure, and it is not in the $6.7B+. The cells stay on the company pages.

No comparable AI-training revenue is published for these companies, so they are not in the total either.

The third home-page figure, advertised versus actual pay, is not one of these vendor cells. It is counted from worker reports. How the pay gap is counted.

Worker pay reports

The headline worker statistic is the effective hourly rate: pay received divided by the hours the work actually took, including waiting for tasks, unpaid onboarding, rejected work, and hours that were cut. The advertised rate is what the platform listed. In the project-type table, the gap is the median of (advertised − effective) on each report in that group.

Advertised vs. actual pay

The home page compares the median advertised hourly rate with the median effective hourly rate across every approved report. It is shown only once 10 reports are approved, as a percent and a dollar gap: advertised $40/hr and actually earned $28/hr is $12/hr, or 30% less. Effective pay includes unpaid onboarding, rejected tasks, and waiting time. A company page shows the same comparison once that company has 3 approved reports. Below those counts the slot stays empty — no placeholder rate. Pending and rejected reports are not counted. Dev seed reports are fake; a local dev server may use them to fill the figure, and labels them as test data. Production does not count them.

Reports are anonymous. The form does not ask for a name. An email is optional, stored only so a report can be checked later, and never displayed.

New reports are pending. A moderator approves or rejects them. Only approved reports count. Rates and individual notes are shown for a company and project type only after at least 3 approved reports, so a single submission cannot identify someone or set the published rate. Below that threshold the page may say how many approved reports exist, without the rate or the notes.

Notes are read before approval. Moderators should reject or withhold a report whose notes name a person. The form asks people not to include names.

Submissions include a hidden field that people do not fill in, and a limit of five reports per network per hour. The site stores a hash of the network address, not the address itself.

What this version does not do

It does not license data, match workers to labs, or move money. A marketplace is a later project. Nothing here is an offer of work or a claim about what a given person should be paid tomorrow.

Corrections

If a figure is wrong, outdated, or missing its source, write to corrections@aiwork.fyi. Include the company page and the URL that should replace or sit beside the current wording. The profiles are a table of claims, and claims can be corrected in public.