The Real Cost of AI: What the Industry Isn’t Telling You

Every time you ask a chatbot to write your email or summarize a contract, a data center somewhere draws more electricity than your home uses in a day. The model running that request was trained on hardware that consumed water equivalent to filling an Olympic swimming pool — not once, but repeatedly across months of training runs. This is the story the industry does not put in its press releases.

The Energy Equation Nobody Is Solving

The International Energy Agency reported in early 2024 that data centers already account for roughly 1–2% of global electricity consumption. That figure is expected to double by 2030 as AI workloads scale. To put it plainly: the AI boom is arriving at the same moment the world is trying to decarbonize its grid.

The irony is sharp. Companies like Google and Microsoft have published ambitious net-zero pledges. Microsoft’s 2023 sustainability report showed its water consumption increased by 34% year over year — a period that coincided exactly with its heavy investment in OpenAI infrastructure.

“We are essentially building a new industrial revolution on top of a grid that was never designed for it. The question is not whether AI uses energy — everything does. The question is who pays.”

— Dr. Kate Crawford, author of Atlas of AI

Water: The Invisible Input

AI training and inference requires cooling. Cooling data centers at scale requires water. A single large training run for a frontier model like GPT-4 is estimated to have consumed between 500,000 and 700,000 liters of water — in regions already facing drought stress.

Water Consumption by Region (Estimated, 2023)

The Human Cost: Data Labeling and Content Moderation

Before a model can learn what a harmful image looks like, a human has to look at thousands of them and label each one. This work — called data labeling and content moderation — is outsourced to contractors in Kenya, the Philippines, and Uganda earning between $1.32 and $2 per hour.

A Time investigation published in January 2023 documented workers at a Nairobi firm contracted by OpenAI who were shown graphic descriptions of violence and sexual abuse in order to train the safety filters on ChatGPT. Many reported lasting psychological harm. Counseling sessions were available, but limited to a single session per worker.

What Needs to Change

1. Mandatory energy disclosure — AI companies should be required to publish per-model energy and water consumption, the same way food products list calories.
2. Fair wage standards for data labor — The ILO has called for platform work to fall under national minimum wage law. Several governments are now drafting legislation.
3. Open-source efficiency research — Smaller models trained on cleaner data often match frontier model performance on specific tasks at a fraction of the compute cost. This research needs funding.
4. Locating data centers near renewable sources — Iceland’s geothermal grid and Norway’s hydroelectric capacity make them natural candidates. Incentive structures currently point elsewhere.

Is There a Cleaner Path?

Some researchers argue that efficiency gains will outpace the growth in demand — a version of Jevons Paradox applied to compute. As models become more efficient, the argument goes, the total energy cost per useful output falls.

The counterargument is simpler: demand is growing faster than efficiency. Every percentage point of efficiency improvement is immediately absorbed by a tenfold increase in the number of queries, applications, and users.

The AI industry is not uniquely villainous. Every industrial revolution has had a resource cost that was visible only in retrospect. What is different this time is that we have the data, the science, and the policy tools to act before the damage compounds. Whether we use them is a choice being made right now — mostly in boardrooms, mostly without public scrutiny.