Who Profits From AI? Tracing Value From the Chip to the Worker

In January 2023 TIME published an investigation into workers in Kenya who were tasked with labelling tens of thousands of text snippets containing violence, hatred and abuse, in order to train a detector that would keep toxic content out of ChatGPT. According to the magazine the workers took home between $1.32 and $2 an hour, while OpenAI's contracts with the contractor, Sama, specified $12.50 an hour. Sama disputed the magazine's figures and stated that pay after tax ranged from $1.46 to $3.74. One story is not to be generalised, but it is a reminder that the "intelligence" we see on the screen has links behind it that we do not see.

Links, not a facade

AI passes through a chain: chips and energy, then cloud computing, then models, then applications, with human labour at every link. In the cloud link, Amazon, Microsoft and Google together held about 63% of a market of roughly $143 billion in the second quarter of 2026, according to Synergy Research Group. On the concentration of models and who can afford to train them, see the article "Who Owns Artificial Intelligence?". In fairness to the evidence: market share is not profit share, and we found no documented figures for the profits of each link.

How large will the gain be?

In a paper published in May 2024, the economist Daron Acemoglu estimates that total factor productivity will rise by no more than 0.66% over ten years, and by less than 0.53% once account is taken that harder tasks yield smaller benefits. He builds the estimate on the share of tasks AI touches and the saving in each. Economists do not agree on this size; for more, see the article "The Economics of AI: Why Do Economists Disagree on the Gain?".

Between workers and between capital and labour

There are two different inequalities that should not be confused. The first is among workers themselves. In a study of 5,179 customer-support agents, an assistant raised the number of issues resolved per hour by about 14% overall and by about 34% for novices and the less skilled, with minimal effect for the experienced. In an experiment with 444 professionals on writing tasks, completion time fell by about 37% and the gap between participants narrowed. Here the tool brings the weaker closer to the stronger.

The second is between capital and labour: in his model Acemoglu expects the gap between capital income and labour income to widen, though he judges that AI's effect on wage inequality will be smaller than that of earlier automation. A novice may benefit from the tool while the larger share of the return goes to its owners. That is a model's prediction, not a measurement of what has happened.

The work that goes unseen

Training and cleaning models relies on people who label data and review outputs, some of whom are exposed to distressing content. The investigation we began with is a documented example, not a comprehensive statistic. Any honest map of how value is distributed includes these workers in its accounts, because fairness in pay and faithfulness to contracts are among the requirements of trust in every industry.

Questions that protect you from being misled

• At which link does the party promising you profits stand?
• Who made the estimate, when, and with what model?
• Is the gain in productivity, in wages, or in profits?
• Who does the unseen work, and for what pay?

For more on reading numbers, see the article "How to Read AI Statistics Without Being Fooled".