I thought I was just being productive. Writing code, debugging, generating boilerplate. Every time I hit Enter on a ChatGPT prompt or asked GitHub Copilot to finish a function, I felt like I was getting things done faster.
Then I started tracking my token usage. Three months later, I had a number that stopped me cold: nearly 1 MWh of electricity. That’s not a typo. My personal AI coding habit consumed as much energy as an average American household uses in a month. And that doesn’t include the AI I use at work – which is probably 10x more.
Here’s the thing nobody tells you: every prompt you type has a physical weight. It’s not just bits floating in the cloud. It’s megawatts of electricity and liters of water, pumped into data centers to keep your AI co-pilot running.
I built a simple local tracker called AI Meter to measure my own usage. It reads logs from coding tools, counts tokens, and estimates the energy and water impact. The numbers are rough – based on published research models – but they’re directionally terrifying.
The estimate breakdown: each million tokens costs about 0.39 kWh of electricity. Every kilowatt-hour of IT load, after accounting for cooling overhead, consumes about 0.45 liters of water for direct cooling. That means every time you ask a model to write a SQL query, a data center somewhere evaporates enough water to fill a small bottle.
We’re so focused on the existential risks of AI – the alignment, the sentience, the job displacement – that we’ve completely ignored the immediate, compounding environmental tax of our daily prompt engineering. We debate hypothetical futures while lighting up a real megawatt.
And the cost isn’t going to stay hidden. Tech companies are subsidizing it now, but as AI usage scales, those energy bills will come due. Either you’ll pay for it directly through tiered pricing, or the planet will pay for it through resource depletion.
What can you do? Start by knowing your number. You can’t manage what you don’t measure. AI Meter is open source, local, and privacy-respecting – no login, no tracking. It’s not a guilt trip; it’s a reality check. The next time you hit ‘generate,’ remember that convenience has a weight. And it’s heavier than you think.
FAQ
Q: What would I actually do with my AI energy usage data?
A: Awareness alone is the first step. Knowing your footprint lets you make informed choices – like choosing lighter models, batching less critical queries, or advocating for transparency from your AI provider. As energy costs rise, this data will become a bargaining chip for corporate governance and personal usage habits.
Q: Is the real environmental impact from training models, not inference?
A: Training is a massive one-time cost, but inference – the daily prompts you send – is growing exponentially and already dominates energy consumption for deployed models. The cumulative effect of millions of users running billions of prompts per day is a constant, compounding drain that training alone doesn't capture.
Q: Isn't this just guilt-tripping users? The real problem is tech companies, not individual prompts.
A: True, systemic change is needed. But the hidden cost is deliberately abstracted away from users to keep adoption frictionless. By making the invisible visible, we shift the conversation from 'AI is magical and free' to 'AI is physical and has a price.' That pressure is what forces companies to invest in efficiency and transparency.