Ales Frelih

Today, the European Union made it illegal for artificial intelligence to pretend it is human. Chatbots must say they are machines. AI-generated images, video and text must be labeled. Deepfakes must be disclosed. I build with AI in the open, so my reaction was simple: good. But if we are going to be honest about what AI is, we have to be honest about what it costs — and that bill is paid in electricity and water, not just euros.

So here is the part almost nobody advertising AI wants to say out loud. The real numbers. The bad and the good. You can decide for yourself.

The damage, in numbers that don’t flatter my own industry

• Electricity. The world’s data centres used roughly 448 TWh of electricity in 2025. The International Energy Agency projects that climbs to about 945 TWh by 2030 — more than the entire electricity consumption of Japan today — with AI the single biggest driver of the growth.
• Water. Google’s own 2026 environmental report admits it consumed 10.9 billion gallons of water in 2025 — up 34% in a single year, and more than double its 2021 level. Amazon disclosed 2.5 billion gallons. That water cools the servers answering your prompts.
• The trajectory. A UN-backed report warns data centres could drink 9.3 trillion litres of water a year by 2030. In Texas alone, projections show data-centre water use rising from 49 billion to 399 billion gallons by 2030 — a 714% increase — in a state that already fights over water.
• Your single prompt. One ChatGPT-class query uses roughly 0.3 watt-hours — about ten times a Google search — plus a small sip of cooling water. Send somewhere between 5 and 50 prompts and the system has quietly used about a 500ml bottle of water. Multiply that by a billion queries a day.

On June 23, 2026, UN Secretary-General António Guterres launched an initiative demanding that every major AI company measure and publicly disclose its carbon, water and land footprint, and run on renewable energy by 2030. That is not a fringe activist talking. That is the top of the United Nations telling my industry to stop hiding the meter.

The other side — because doom with no nuance is also a lie

• Efficiency is improving fast. Per-query energy for mainstream models has fallen sharply as hardware and software mature. The cost of a single answer today is a fraction of what it was two years ago.
• AI pays some of it back. The same technology is being used to optimise power grids, cut energy waste in buildings and logistics, speed up drug discovery, and improve climate and weather modelling. In several cases the energy AI helps save outweighs the energy it burns.
• Water can be designed out. A growing share of data centres are shifting to recycled or non-potable water, closed-loop cooling, and renewable power. The 10.9-billion-gallon headline is real — so is the fact that it is now being measured and reported, which is the first step to shrinking it.

Why I’m telling you this instead of selling you magic

There are two dishonest stories on your feed right now. One is the hype machine: AI is pure magic, infinite, free, consequence-free. The other is the doom machine: AI is nothing but theft, pollution and lies. Both are selling you something, and both are lying by leaving half the facts out.

I’m doing the third thing. I build with AI — openly, labeled, honestly — and I’ll tell you the whole ledger, including the numbers that make my own work look worse. That’s the entire point of this place: where rules, money and machines collide, with no PR gloss on top.

Transparency isn’t just a machine admitting it’s a machine. It’s telling you what the machine costs the planet — and then deciding, with your eyes open, whether it’s worth it. Today the law made AI confess what it is. The harder confession is what it uses. There are the numbers. Now the conversation can actually be honest.

Sources

• IEA, Energy and AI (data-centre electricity 415–945 TWh)
• UN University / UN AI Environmental Transparency Initiative, June 2026 (9.3 trillion litres by 2030)
• Google 2026 Environmental Report (10.9 billion gallons, +34% YoY); Amazon disclosures
• HARC / University of Houston (Texas data-centre water projections)
• Academic estimates of per-query energy and water footprint (~0.3 Wh; ~0.32 ml/query)