By now you’ve all heard about Chinese AI sensation DeepSeek, the startup that built a model matching ChatGPT for just $5.6M—0,001% of Stargate’s $500B price tag. Some, like Sam Altman, are crying corporate espionage. China Daily, instead, is celebrating a “historical moment when Chinese LLMs surpass ChatGPT”. But both sides are missing the real story.
DeepSeek’s low cost isn’t only about engineering breakthroughs or stolen IP. And it’s not only about cheap labor either, though if you follow me, you’ve heard me go on about it plenty.
It’s about cheap ✯government-subsidized✯ data labor.
Sure, I’ve spent years researching the hidden labor force behind AI, and here’s the thing: these models don’t train themselves. They need data workers. They are not the “young and high-potential” talents fresh out of top Chinese universities that DeepSeek prioritizes. In my book, Waiting for Robots (University of Chicago Press, 2025) (▻https://lnkd.in/eatuRpaz), I document how AI companies have long outsourced filtering, classification, annotation, labeling of data to workers in the Global South. With my DiPLab teammates, we’ve spent years traveling across three continents, surveying thousands of workers—some earning as little as $100 a month working for the likes of OpenAI and Amazon. And workers for Chinese companies? We met only a few. Why? Because Chinese firms take a different approach: bringing data work back home.
Chinese AI companies have set up massive data-annotation bases in “low-tier cities”, less prosperous but with a large workforce. Sociologists Tongyu Wu and Bingqing Xia have documented how these annotation hubs sometimes originate from local poverty alleviation programs, but in practice, they create a race to the bottom for wages and working conditions (▻https://lnkd.in/eNiHf35v). And this isn’t just a local administrations—it’s government policy. Between 2023-2024, China’s National Data Administration announced measures to expand its data-labeling industry by 20% annually, creating annotation hubs, offering tax breaks, direct subsidies, and “data vouchers” to fuel the Chinese data annotation market.
Compare this to the U.S., where the government’s AI strategy revolves around funding gargantuan servers, sprawling data centers, and environment-draining energy. Trump is already floating the idea of “national energy emergencies” to power Stargate. But China is making a different bet—not just on servers, but on the human labor behind AI itself.
Let’s be clear: this is not a defense of Chinese AI strategy. Their policies aren’t about improving working conditions, but about maximizing profits for annotation companies. DeepSeek, after open-sourcing its model, even refuses to disclose how much human labor fuels its so-called autonomous reasoning.
We’ve seen this before. In 2023, Time exposed OpenAI’s Kenyan annotators earning less than $2 per hour. DeepSeek is just the latest example of the same old playbook: hype the technology, hide the workers.