Disrupted
Appen and the labour that trained its own replacement
Appen · AI data services · Australia
Appen sold human data labelling to train AI. Then the value moved up the stack: Google cut a contract worth about a third of its sales overnight, and market value fell roughly 99%. The surviving business is growing — it is just a far smaller company now.
−99%
market value since the 2020 peak
~30%
of 2023 sales was the Google contract, cut overnight
+16%
ex-Google revenue growth, FY2024
Human data labelling repriced by the stack it trains.
The record
- Appen’s market value has fallen about 99% from its August 2020 peak, when the stock reached roughly A$42 and a market cap near US$4.3 billion — as the economic value in AI migrated away from human data labelling toward other layers of the stack. (CNBC, May 2024)
- Alphabet notified Appen in January 2024 that it was ending their contract — worth US$82.8 million, about 30% of Appen’s 2023 sales — effective 19 March 2024, with no prior notice; the shares fell about 40% on the news. (AIwire; CNBC, 2024)
- Revenue fell 30% in 2023 to US$273 million and a further 14.2% in FY2024 to US$234.3 million — yet excluding the lost Google work, FY2024 revenue grew 16% to US$220.9 million. The surviving business was expanding even as the headline collapsed. (Appen FY2024 results; reporting)
A business built on the input layer
Appen’s product was human judgment at scale: labelling, rating and annotating the raw data that trains and tunes AI models. For a decade that was a growth story — more AI meant more labelling — and the moat was a global crowd workforce plus direct relationships with the model builders themselves.
On paper it was an AI company on the right side of the trend. But the value it captured sat at the very bottom of the stack: the human-labelled data everything else was built on.
The value moved up the stack
The same models Appen helped train began to need less of what Appen sold. Synthetic data, model-generated labelling and in-house pipelines let the largest buyers cut their dependence on external human labelling — or bring it in-house entirely.
Alphabet’s abrupt termination in early 2024 was the sharpest instance: no notice, and about a third of Appen’s sales gone at once. It was less a pricing dispute than a buyer routing around the input layer altogether.
The freeze, not the overheat
Appen did not over-automate or relabel itself to chase a valuation. It was disrupted the other way: the part of AI it owned turned out to be the part that could be commoditized, synthesized or insourced. Its pivot toward generative-AI and sovereign work is real, and the business that remains is growing.
But it is a far smaller company than the one the market had priced, and the pivot began a beat after the repricing. That is the freeze signature — not a dramatic failure, but the quiet discovery that being central to AI is not the same as owning a defensible piece of it.
The lesson
Being “in AI” is not the same as owning a defensible layer of it. If your part of the stack can be automated, synthesized or brought in-house, demand can vanish overnight — however central you were to training the thing that replaces you.
How we’re reading this
Appen’s fall has more than one cause: heavy customer concentration (Alphabet was about a third of 2023 sales), crowd-work margin pressure and macro tightening all compounded the AI-substitution effect rather than it acting alone. The company is adapting — generative-AI and sovereign work is growing — so this is “repriced despite a pivot,” not refusal. Figures are as reported.
Sources
- 01
- 02Appen Suffers Major Blow as Google Terminates Multi-Million Dollar Contract
AIwire · Jan 29, 2024
Compiled from public filings, court records, company statements and reputable press. Figures are attributed to their sources; allegations are labeled as such. Not legal or investment advice.