The Data Center Workers Nobody Writes About

There’s a sentence that appears in nearly every AI product launch: “Powered by state-of-the-art infrastructure.” The sentence is true, technically. What it doesn’t mention is that the infrastructure is powered by people — often poorly paid, often invisible, often working in conditions that the companies building the AI would not want photographed.

Two categories of worker keep AI functional. The tech press almost never writes about either of them.

The AI industry has become extraordinarily good at producing a certain kind of narrative. Benchmark improvements. Impressive demo videos. Thoughtful blog posts about responsible development. The vocabulary of the field — “frontier models,” “inference clusters,” “compute efficiency” — is thoroughly technical, thoroughly abstract, and thoroughly free of human beings doing physical work for wages. This is not accidental. Industries that prefer not to discuss their labor supply chains develop languages that make those chains invisible. Fast fashion talks about “supply network partners.” The AI industry talks about “data pipelines.” The pipeline has workers in it. The workers mostly go unmentioned.

The first category is the physical data center workforce. When an AI company builds a data center in Prineville, Oregon, or Clarksville, Tennessee, or outside Columbus, Ohio — these locations are chosen deliberately. Cheap electricity from hydropower or coal. States with weak labor organizing traditions. Land that’s distant enough from urban centers that the facilities don’t get scrutinized. The workers who staff these facilities — maintaining cooling systems, swapping failed drives, managing the physical hardware — are not well-compensated engineers from Stanford. They’re technicians earning $20-25 an hour, often employed through staffing agencies that sit between them and the companies whose infrastructure they maintain.

The agency structure matters. It means that when Amazon Web Services or Microsoft Azure builds a data center, the workers keeping those servers running are not Amazon employees. They’re employees of a contractor, who is employed by a sub-contractor, who has a service agreement with the hyperscaler. This layering is not accidental — it places legal and moral distance between the tech company and the labor conditions inside its facilities.

The physical conditions in data centers are genuinely difficult. Server rooms run hot — 35 to 40°C in the aisles between racks is normal during peak load, and some facilities push higher before thermal throttling kicks in. Workers spend extended shifts on raised floors under fluorescent lighting, in elevated noise environments from cooling fans running at full blast. Repetitive motion injuries from drive swaps are common. The buildings themselves are often remote enough that there’s no practical public transit, meaning car ownership is a prerequisite for employment, which quietly eliminates workers at the lower end of the wage range.

There are roughly 10,000 data centers of significant scale operating globally. The largest — hyperscale facilities operated by AWS, Google, Microsoft, and Meta — can occupy two million square feet and house hundreds of thousands of servers. Each of those servers needs periodic maintenance: thermal paste replacement, failed component detection, physical cleaning, cable management. None of that happens autonomously. Robotics has made progress in warehouse logistics, but server rack maintenance at the component level remains a human job. It will remain so for a long time. These are not abstract statistical workers. They are specific people, working specific shifts, inside buildings whose existence is mentioned in earnings calls as a competitive advantage and whose workforce is invisible in those same calls.

The second category is harder to look at directly.

Content moderators and data labelers are the people who make AI models usable. Before a large language model can decline to produce child sexual abuse material, someone has to label thousands of examples of it. Before a model can recognize violent content, human workers have to view violent content and annotate it. Before an image classifier can flag hate symbols, someone sits in a small office in Nairobi or Manila and classifies hate symbols — one by one, hour by hour, at wages that would be illegal in most of Europe and all of California.

TIME magazine’s reporting in January 2023 revealed that workers employed by Sama, a firm contracted by OpenAI to label data for ChatGPT, were paid between $1.32 and $2 per hour to review some of the most disturbing content on the internet. Workers reported symptoms consistent with acute stress responses and PTSD. Sama terminated the contract early — but not before the reporting made the arrangement public. The arrangement itself was not unusual. It was standard practice.

The geography of this workforce is not coincidental. Kenya, the Philippines, Venezuela, India — these countries supply the majority of content moderation and data labeling labor because wages are low enough to be profitable at the volumes AI training requires. A well-funded AI company might need millions of labeled examples for a single training run. At $2 an hour, even a modest labeling workforce can produce that at acceptable cost. At $25 an hour — the minimum to make this work ethically in a US context — the economics break down entirely.

The psychological toll is significant and well-documented. A 2022 paper by researchers including Thilo Hagendorff surveyed content moderators and found rates of PTSD symptoms substantially higher than in comparison populations. The problem is not obscure — it is discussed at conferences, in academic literature, in investigative journalism — and the response from the industry has been roughly nothing. Some companies provide “wellness resources,” defined as a telephone hotline number in an HR document. That the work causes psychological harm is treated as a regrettable feature of a necessary process rather than an avoidable design choice.

The structural comparison that fits here is fast fashion. In 2013, the Rana Plaza building in Dhaka, Bangladesh collapsed. 1,134 garment workers died. The building housed factories producing clothes for H&M, Primark, Benetton, and others. The fashion companies did not own the factories. They had contracts with manufacturers who had contracts with the factory operators. Legally, the companies maintained, they were customers, not employers. The legal argument was correct. The moral argument was not.

AI companies are in the same position with their content moderation and data labeling supply chains. They don’t employ the workers. They have contracts with firms like Sama, Appen, Scale AI, iMerit — firms that in turn hire workers in low-wage countries. The arrangements are legal. Whether they’re ethical is a different question, and one the industry prefers not to have to answer.

Scale AI raised $1 billion at a $7.3 billion valuation in 2021. Its business is, fundamentally, paying large numbers of people small amounts of money to do cognitively and sometimes psychologically taxing work so that AI companies can train their models. This is not a criticism of Scale AI specifically — they’re operating within a system that the entire industry has built and depends on. But the gap between the valuation and the wages paid to the workers whose labor justifies that valuation is worth sitting with.

The data center workforce has its own version of this dynamic. The companies operating AI infrastructure at scale — Google, Microsoft, Amazon, Meta — are among the most valuable corporations in human history. The workers maintaining their physical infrastructure are, in many cases, a staffing agency phone call away from unemployment and lack health benefits through their actual employer. The hyperscalers have successfully externalized not just the work but the employment relationship itself.

None of this is secret. It’s just not mentioned. The AI industry communicates primarily through product launches, benchmark announcements, and blog posts about safety and alignment. The working conditions of the people who make the products possible don’t fit into that communication style, so they don’t appear. The effect is that the public discussion of AI is almost entirely about the technology and almost entirely silent about the labor.

There’s a useful contrast with the early computing industry. When IBM was building the System/360 in the 1960s, the workers on the factory floor in Poughkeepsie were IBM employees, with IBM benefits, in unionized or near-unionized conditions. This wasn’t altruism — it was the labor market structure of the era. The current AI industry has access to global labor markets that 1960s IBM did not, and it has used that access to separate the returns of AI from the workers producing it.

What would changing this look like?

It would look like AI companies treating content moderation and data labeling workers as direct employees, not as the workforce of a contractor three steps down the supply chain. It would look like data center workers having employment relationships with the companies whose infrastructure they maintain. It would look like wages and psychological support commensurate with the work being done — which means trauma counseling for content moderators, not just a hotline number in a contract.

None of this is technically difficult. It’s expensive. The companies have made a calculation that the expense is avoidable, that the workers are invisible enough that the calculation won’t attract sustained attention, and that the public conversation about AI will remain focused on capabilities and governance rather than labor.

That calculation has mostly been correct. A handful of journalists — TIME, The Guardian, MIT Technology Review — have reported on these conditions with rigor and specificity. The stories generated attention for a few days and then the conversation moved back to benchmark scores and geopolitical competition.

The contrast with the discourse around AI safety is instructive. The AI safety community — alignment researchers, governance specialists, biosecurity-focused AI thinkers — produces voluminous written material about risks that are mostly speculative and future-tense: what might happen when AI systems become more capable, what the risks are from systems we don’t yet have. These are legitimate questions worth taking seriously. The workers being harmed by the AI industry right now are not a speculative future risk. They are a present fact. The discourse gives approximately one hundredth the attention to present harms as to speculative future ones. That’s a choice about what counts as a serious problem, and the choice is not neutral.

If you buy and use AI products — ChatGPT, Gemini, Claude, Midjourney, any of them — you are a consumer of services whose supply chains include psychological harm at $2 an hour and physical labor in 40°C server rooms with staffing agency employment. This is true. The products are still useful. Both things can be simultaneously true. What can’t be true is that you know this and continue to expect the companies to make it visible without external pressure. They won’t. The incentive to disclose the conditions of invisible workers is negative: disclosure attracts scrutiny, changes nothing legally, and provides no competitive advantage.

The workers who review the content that makes AI models safe, and the workers who maintain the servers that make AI models run, are not incidental to the AI industry. They are constitutive of it. The industry is structured, deliberately, so that this fact doesn’t have to be reckoned with. Reckoning with it is a choice the industry hasn’t made. The question is whether the people who buy and use AI products are willing to demand it.