AI firms keep hunting for fresh data to sharpen their models. Some turn to low-paid annotators who tag endless images or transcribe awkward conversations. Others have taken a more direct route. Last week Joi AI, a startup behind NSFW chatbot companions, posted openings for ten masturbation consultants. Selected applicants would earn $2,000 a month to test audio features and file weekly reports on their experiences. The response? Over 100,000 applications in days.
Futurism first highlighted the listing. Head of brand Julie Levin told Business Insider the company received far more interest than expected. “What are we supposed to do with 100,000 applications?” Levin asked. She added that applicants should reflect on how the activity affected their lives. “We expect people to learn something about how masturbation affects their life in a good way or a bad way.”
Applicants flooded the form with quips. “I’ve been training for this my whole life.” “My therapist said I needed a hobby.” One X user joked, “Time to go pro.” The company itself tweeted about the flood of unusual cover letters. Yet behind the viral buzz sits a serious point. Even the most intimate human data still feeds the same machine-learning pipelines that now strain electric grids worldwide.
Data centers already consumed roughly 415 terawatt-hours of electricity in 2024. That equals about 1.5 percent of global supply. The International Energy Agency now projects that figure will nearly double by 2030. AI-focused facilities drive much of the surge. Their electricity demand grew faster than overall data-center use in 2025, climbing well ahead of the global average 3 percent rise in power consumption. A single large training run can burn through gigawatt-hours. Inference, the everyday queries users send to chatbots, now accounts for the majority of added load.
One typical hyperscale AI data center draws as much power as 100,000 households. The largest planned sites will consume twenty times more, according to IEA analysis. Cooling those racks sucks up billions of gallons of water each year. And the heat keeps rising. Recent facilities pack 50 to 100 kilowatts per rack. Older designs rarely exceeded 10. The difference shows up on utility bills. In several states homes and businesses paid an extra $4.3 billion in 2024 for transmission upgrades tied directly to data-center growth, per an analysis by the Union of Concerned Scientists.
But power is only one constraint. Transformers needed to step up voltage for these plants face lead times of two to five years. Domestic production of grain-oriented electrical steel, the core material, remains limited. One recent X post noted that AI data-center expansion now hinges less on raw generation and more on these long-lead components. Meanwhile capital spending by five major tech companies topped $400 billion in 2025 and is forecast to jump another 75 percent this year.
Anthropic has estimated that training a single frontier model could require five gigawatts of power by 2027. The firm also projected the U.S. AI sector alone may need 50 gigawatts of new capacity by 2028. That is roughly twice New York City’s peak demand. Former Google CEO Eric Schmidt told Congress that data centers will require an additional 29 gigawatts by 2027 and 67 more by 2030. These numbers arrive while U.S. electricity consumption had stayed essentially flat for two decades before AI demand kicked it higher.
Companies respond in varied ways. Some sign direct deals for natural-gas plants. Others explore small modular nuclear reactors. A few push hardware makers for efficiency gains. Intel, for instance, detailed new inference-focused GPUs and heterogeneous chip designs aimed at lowering energy per query. Yet efficiency improvements have so far been outrun by adoption. Power use per AI task has fallen, but total queries and agent-style workloads keep climbing. The net result? Data-center electricity demand in the United States could reach between 325 and 580 terawatt-hours by 2028, according to Lawrence Berkeley National Laboratory estimates. That would push the sector toward 6.7 to 12 percent of national consumption.
The Joi AI episode illustrates another layer. Human feedback loops remain essential even as models grow more sophisticated. NSFW chatbots in particular have drawn scrutiny. Researchers have linked heavy use to higher rates of loneliness, depression, and in extreme cases breaks with reality. One study cited by Futurism found certain chatbots correlated with elevated psychological distress. Yet the demand for ever-more-personalized data persists. Consultants writing detailed reports on intimate sessions simply represent an unusually candid version of the labeling work performed daily by thousands of contractors worldwide.
Local impacts mount. In Virginia, data centers already accounted for nearly 40 percent of electricity consumption in 2024. Residents near new builds report higher bills as utilities spread infrastructure costs across ratepayers. Water use for cooling adds pressure in drought-prone areas. And the reliance on fossil fuels in the short term undercuts corporate net-zero pledges. Goldman Sachs research forecasts AI will represent more than a quarter of the global data-center market by 2027.
So far innovation races to match the scale. Chip designers chase lower wattage per operation. Utilities scramble to site new transmission. Policymakers debate faster permitting for nuclear and renewables. None of these fixes erase the core tension. Every advance in model capability seems to multiply the compute required downstream. Even a seemingly lighthearted job posting for audio erotica testers feeds the same infrastructure that now competes with households and factories for electrons.
The applications keep pouring in. The servers keep humming. And the grid keeps tightening. Whether the consultants ultimately help produce more engaging chatbots or simply highlight the odd corners of data collection, one fact holds. The machines are hungry. And society is only beginning to measure the bill.
WebProNews is an iEntry Publication