The real cost of AI isn’t on your invoice

The real cost of AI isn’t on your invoice

I can’t tell you what my AI usage costs.

I tried to work it out. I opened the billing pages, added up my subscriptions, and felt briefly in control. Then I realized I had answered one question out of five. I knew what AI costs me in euros. I had no idea what it costs in electricity, in water, in attention, or in freedom.

Most days, AI is open on my machine from morning to evening. Claude Code for my apps, Codex when I want a second opinion, an AI service wired into my Obsidian vault through MCP. It has become infrastructure. And infrastructure has a way of disappearing from view when it becomes essential.

This is not an article against AI, and it is not a defense of it either. I use it too much to pretend to be neutral about its benefits, and I’ve read too much to pretend it’s free.

What I want is simpler: to lay the costs on the table, be honest about what we know and don’t, and leave you with five questions worth carrying into your next session.

The ledger nobody shows you

When we talk about the cost of AI, we usually mean the subscription or the token rates. Those arrive on an invoice. The other costs show up in our working hours, in shared infrastructure, or later.

I count five columns:

  • Energy: the electricity that powers the data centers.
  • Water and carbon: the water consumed and greenhouse gases emitted along the way.
  • Money: the visible cost, plus the invisible one (review time, rework, stacked subscriptions).
  • Cognitive: what happens to our own thinking when we delegate it.
  • Dependency: how much of our work and knowledge becomes hostage to a vendor.

I can put a price on some of these. Others need a different kind of question. I want them in view before I decide what the tool is worth.

Energy: a small number multiplied by a large one

I need a sense of scale before I can judge my own use. In its April 2026 report, the International Energy Agency puts worldwide data center electricity use at about 500 TWh in 2025. Its central projection sees demand roughly doubling by 2030, reaching around 3% of global electricity use. Those totals cover AI and other digital services.

The same report estimates that answering all conventional web searches with simple AI text queries would consume less than 1% of current data center electricity use. A short request can use little electricity, while total demand grows as people use AI for more demanding work.

The kind of use matters. A short question, a reasoning model working for minutes, and an agent chaining dozens of calls to refactor a codebase ask for different amounts of work. My own workflow sits on the demanding end of that spectrum.

Question to carry with you: when I say “AI is cheap,” am I thinking about a prompt, or about the way I use it?

Water and carbon: what the numbers leave out

The bottle-of-water comparison needs its denominator. A study first released in 2023 estimated roughly half a liter for 10 to 50 medium-length GPT-3 responses, depending on when and where the model ran. That estimate included cooling and electricity-generation water. It was a modeled result for a particular system, not a universal meter attached to ChatGPT.

Google offers a much smaller figure: about a quarter of a milliliter of on-site cooling water for a median Gemini Apps text prompt, based on May 2025 usage. But the accounting boundary is narrower: it excludes the water used to generate electricity. Google also says the estimate has not been independently verified. I can’t read those two figures as a direct measure of progress between GPT-3 and Gemini.

A 2025 inference benchmark estimates much larger footprints for some long requests and reasoning models. Its results depend on assumptions about hardware and infrastructure. They help explain why the footprint varies; they don’t give me a reliable water bill for my own prompts.

What I take from this is that a per-query figure needs context: the model, the response length, the cooling system, and what the calculation includes. Location matters too. As a 2026 study of water accounting emphasizes, the same physical volume has different consequences depending on local water stress. An average across a global fleet can hide the pressure on a particular community.

Carbon needs its own line of thought. Electricity use translates into different emissions depending on how the power is generated. And a figure for serving a prompt doesn’t automatically include training the model or manufacturing its hardware. The water research also points to a trade-off I find easy to overlook: a choice that lowers carbon emissions can increase water consumption. I need to know what a footprint counts before I compare it with another one.

Question to carry with you: do I know where my queries are answered, and would it change anything if I did?

The rebound nobody budgets for

When a task becomes more efficient, it can also become easier to do more often. That is the rebound effect: extra use can absorb some of the savings, or even outweigh them. The 2025 benchmark discusses this risk for AI. I read it as a reason to watch total usage alongside efficiency, rather than assume that either trend settles the question.

A subscription makes that easy to miss. An extra request may leave the monthly bill unchanged while still using resources. With an agent in my workflow, the useful question becomes: how much work am I asking it to do, and does the task justify it? There is no single effort setting that answers that for every model and every job.

Question to carry with you: has my usage grown faster than the efficiency gains I’m counting on?

The cost that doesn’t show up on any meter

This is the column I find hardest to write about, because the evidence is young and the headlines are loud.

In the 2025 preprint “Your Brain on ChatGPT,” researchers studied 54 participants writing essays over three sessions: with ChatGPT, with a search engine, or without either. The ChatGPT group showed the weakest EEG connectivity during the task, struggled more to quote their essays, and reported less ownership. The authors describe the pattern as cognitive debt.

I find the result worth considering, with limits. EEG connectivity during an essay task doesn’t establish a lasting loss of ability. Only 18 participants completed the fourth session, which switched some participants between AI and unaided writing. Those who wrote unaided first showed stronger engagement when they later used AI. That makes the order of work an interesting lead, with a small sample behind it.

A separate Microsoft and Carnegie Mellon study surveyed 319 knowledge workers. Higher confidence in AI was associated with less self-reported critical thinking. The researchers also found that the work shifted toward checking information and integrating answers. They measured people’s reported experience, so this doesn’t demonstrate a decline in their abilities over time.

Neither study tells me what years of AI-assisted coding or writing will do to my thinking. They do give me a practical question: am I still doing the work that lets me judge the answer?

Question to carry with you: what part of my thinking do I never want to delegate, and am I protecting it?

The invoice is only one line

The visible cost is the subscription. The invisible costs sit around it.

There’s the time spent reviewing output before I can trust it. There’s rework when the first answer is plausible and wrong. There’s subscription fatigue, as each tool adds another monthly line and another login. And there’s dependency, which I think about more than any of the others.

I learned that one the practical way. My first iOS app began as a vibe-coding experiment. It worked, and it needed a serious refactoring later (a proper architecture, testable code, versioned data migrations) before I could trust it. The initial speed was real. So was the bill that came later, in a different currency.

Dependency works the same way at a larger scale. If your notes, your habits, and your workflows live inside one vendor’s platform, the real price is what it costs you to leave. That’s why I keep my vault in plain text and my process in files that any tool can read.

Question to carry with you: if my main AI tool disappeared tomorrow, what would I lose, and what would I merely have to relearn?

What I see in my own practice, and where I’m unsure

I won’t turn this into a list of rules. I don’t have any that I follow consistently.

What I can offer is what I notice. Matching the model to the task is one choice I can control. Running a local model, as I did in an earlier experiment, gives me more control over where my data goes. Its environmental benefit needs a separate comparison: my machine still uses electricity, and the model, hardware, and workload all matter. Doing the thinking first and bringing in the AI second is another choice, and it is the hardest to keep up.

I also can’t claim any of this offsets anything. One person adjusting a setting is not an energy policy.

About the numbers

I’ve kept a few numbers because they establish scale. They come with different kinds of evidence: the IEA estimates demand and projects future scenarios; Google reports an estimate for a defined service and period; researchers model environmental footprints or study particular tasks. Those distinctions matter more than an extra decimal place.

The links point to the underlying reports and studies. None gives me a complete footprint for my own workflow, and the cognitive findings leave long-term effects open. I can take the questions seriously without pretending I have a precise total.

I’m also a heavy user writing about a tool I benefit from daily. That’s a bias, and you should weigh my conclusions accordingly.

The grid of questions

These questions feel most immediate in a workflow like mine, with an assistant open much of the day. Occasional short text requests give you less reason to focus on aggregate consumption than sustained, demanding use. But frequency alone doesn’t describe a workload, and I can’t give you a universal threshold below which every cost disappears.

Over the years I’ve learned that tools aren’t good or bad. What matters is whether we keep choosing them. AI is the first tool I’ve used that makes it easy to stop choosing.

So here is the grid, in one place:

  1. Scale: am I thinking about a prompt, or about my real usage?
  2. Place: do I know where my queries run?
  3. Rebound: has my usage outgrown the efficiency I’m counting on?
  4. Thinking: what will I never delegate?
  5. Exit: what do I lose if this tool disappears?

The real cost of AI goes beyond what it charges us. It includes what we stop noticing while we use it.

If you had to put a price on your AI use today, what would be on the invoice?

Sources

Subscribe