The Hidden Water, Energy, and Labor Cost of AI

Most conversations about AI focus on what it can do. Few focus on what it costs to run, and who pays that cost. For operators in industrial water treatment, resource recycling, and clean-tech, that's a strange gap: these are exactly the kinds of questions your own industry asks about any new process before adopting it.

Two costs rarely make it into the pitch deck for an AI tool: water and energy, and human labor.

Why AI's water and energy footprint matters for resource-management companies

Training and running large AI models requires data centers with real water-cooling and electricity demands. That demand is large enough that it's driving new fossil-fuel power plant construction in several US regions, in addition to, or instead of, renewable capacity buildout.

The scale is easy to underestimate because it's invisible at the point of use. A single AI-generated response doesn't feel resource-intensive the way a factory process does. But in aggregate, the data centers behind consumer-facing AI tools are now material enough to show up in regional power-grid planning and water-rights discussions, the same categories of infrastructure strain that our own regulatory positioning work with sustainability companies is built to help clients navigate.

What to ask an AI vendor before you adopt their tool

The second cost is harder to see, and arguably more overlooked: human labor. Every large AI model is trained and maintained using datasets that were labeled, moderated, and verified by human workers, a process the industry calls "data labeling" or, less charitably, "ghost work." Multiple documented accounts, including a 2024 open letter from 97 data labelers and content moderators in Nairobi working for US tech companies through outsourcing firms, describe pay and working conditions that would fail a basic ESG supply-chain review.

This doesn't make every AI tool equally compromised. Plenty of AI research is built transparently, on ethically sourced data, with no exploitative labor anywhere in the chain, medical imaging models trained on clinician-labeled datasets are one documented example. The difference comes down to whether a vendor can actually answer a handful of concrete questions:

1. What dataset was the model trained on, and how was that data obtained? Was it collected with consent, or scraped at scale from the public internet?
2. Who labeled and processed that data, under what conditions? Were they paid a living wage, given reasonable hours, and, if the material was disturbing, given mental health support? Is data labeling and processing ongoing, and is that necessary for the tool to function?
3. Does the tool depend on a larger model underneath it? A smaller vendor's product can still inherit a megacorp AI's labor and resource footprint if it's really just a wrapper calling that model's API.

If a vendor can't answer those questions, that's itself an answer worth weighing the same way you'd weigh an unauditable supplier in any other part of your operation.

Further reading

Researcher and writer Dr. V has put together the clearest framework we've seen for working through those three questions in practice, applied to real examples from both ends of the spectrum, from AI tools fully dependent on exploitative labor to genuinely human-safe research models. If your organization is evaluating AI vendors, or just wants a sharper vocabulary for assessing AI's actual footprint, it's worth the full read.

Read the full piece: "Exploitative AI" →

Fact-checked by Matt Maufe. Facts current as of the article's last update, September 11, 2026.

For more on how we help sustainability and clean-tech operators turn resource and regulatory questions like this into clear market positioning, see our services, or explore the data behind the industries we track on the Climate Tracker.