Appen

Appen is the base rate. It is the only pure-play in the expert data vertical that has run a full cycle in public with audited numbers, and everything the private cohort is currently being marked on has already been tested here.

#The cycle

Source: stockanalysis.com. A 48% revenue decline over four years, with the equity value falling further: peak market capitalisation surpassed the equivalent of US$4.3 billion in August 2020, against A$330M on 28 August 2026 — a drawdown The Verge computes at 97%. (The atlas's own two figures do not quite reproduce that. A$330M against the ~US$4.3B peak converted to Australian dollars gives about 92%; converting the current cap to US dollars instead and comparing to US$4.3B gives about 95%. The 97% is carried here as reported by The Verge, not as computed by this page — the gap is a currency-basis artefact of exactly the sort item 7 of the atlas's contradiction list warns about, and it does not change the finding, which is that essentially all of the equity value went.)

Appen's collapse was not caused by bad execution on labelling. It was caused by buyer concentration plus a change in training technique. Both conditions are present, in more extreme form, across the current private cohort. See One customer is a binary event.

#The Google contract

At peak, 80% of revenue came from five clients — Microsoft, Apple, Meta, Google and Amazon. On 22 January 2024, Alphabet terminated a contract worth roughly US$83M, about a third of remaining revenue, as it cut thousands of search quality raters (NBC). Appen shares fell 40–41% in a single day. North American offices closed the following month and executives left through May 2024.

One buyer decision, one third of the revenue, one day. That is the mechanism the entire vertical is exposed to, and it is why Mercor's ~91% revenue share from foundation-model companies is the most important number on its page.

#FY2025 — stabilisation, and what it cost

The most recent audited year is more interesting than the headline suggests (Appen FY2025 Annual Report):

• Operating revenue US$230.8M, underlying EBITDA US$12.2M — up 250% from $3.5M.
• Gross margin 40.3%. Higher than Mercor's leaked 33%.
• Appen Global fell 21.1% to $127.9M. Appen China grew 74.8% to $102.9M. The business is now nearly half Chinese.
• Generative-AI revenue rose to 33% of total, from 22%.
• Top five customers = 74.3% of revenue, up from 67.3%. Concentration is increasing, not falling.
• Headcount 1,185; crowd of 1M+ contributors across 200+ countries and 500+ languages.
• Crowd NPS fell from 33 to 22.

That last line deserves more attention than it gets. Worker satisfaction deteriorated even as the business stabilised — the same pattern visible in falling rates at Outlier, Mercor's Musen-to-Nova cut and Handshake's Project HH. The supply pools are shared across platforms, so a deteriorating crowd is a real cost of goods, not a PR problem.

#What the market pays

~0.9x revenue — A$330M of market cap against A$361.9M of TTM revenue, both in AUD — on a 40.3% gross margin business that is growing again. Innodata, the other public comparable, trades at ~6.2x on a ~40% gross margin (49% adjusted, Q2 2026) and 40%+ growth. Both sit far below the private marks: Mercor at roughly 29–38x net, Invisible Technologies at ~15x, Handshake AI at ~7.8x net.

See What the public market pays for labour for the full restatement.