HUMAN EXTINCTION COUNTDOWN

AI (Artificial Intelligence) is software that can do tasks we used to think only humans could do — like writing text, answering questions, translating languages, recognising photos, or generating images.

Modern AI systems (such as ChatGPT, Google Gemini, or Claude) are powered by large language models (LLMs) — giant programs trained on billions of web pages, books, and articles so they can predict what words should come next in a sentence.

AI models don't read words — they read tokens. A token is a small chunk of text, roughly 3–4 characters on average. Common short words like "the" or "is" are a single token each. Longer words get split into several tokens.

Every time you send a message to an AI chatbot, it reads your message (input tokens) and writes a reply (output tokens). Both cost energy.

Inference is the moment an AI actually generates a response — when it takes your question and produces an answer, word by word (or token by token).

It's different from training (which is the one-time process of teaching the AI using vast amounts of data). Inference happens billions of times every day, around the clock, every time anyone uses an AI tool anywhere in the world.

Training is the expensive, one-time (or infrequent) process of building an AI model. It requires running data through billions of mathematical operations on specialised hardware for weeks or months. A single training run can use as much electricity as hundreds of homes use in a year.

Inference is the ongoing, every-second process of using that already-trained model to answer questions. Each individual inference is cheap, but with hundreds of millions of queries per day across every AI product on Earth, the cumulative energy use is enormous — and this is what this site tracks.

Generating each token requires running your input through a neural network with billions of mathematical multiplications on specialised chips called GPUs or TPUs. These chips are extremely powerful — and extremely power-hungry.

A single AI server rack can draw as much electricity as 20–30 family homes. Data centres housing thousands of those racks run 24 hours a day, 365 days a year. The chips also generate a lot of heat, requiring large cooling systems that consume even more electricity and water.

AI data centres use enormous amounts of water to cool their servers. When chips run at full power, they generate heat. That heat is carried away using chillers and cooling towers that evaporate large quantities of water into the air.

Microsoft's own sustainability report estimated that training a single large AI model can consume hundreds of thousands of litres of fresh water. Inference — happening continuously — adds billions more litres every year across the industry.

CO₂ (carbon dioxide) is a greenhouse gas. When power stations burn coal, oil, or natural gas to generate electricity, CO₂ is released into the atmosphere. This extra CO₂ traps heat from the sun, gradually warming the planet — a process known as climate change.

Because AI data centres consume huge amounts of electricity, and much of the world's electricity grid still relies on fossil fuels, running AI produces a significant amount of CO₂. Even data centres that use renewable energy have an indirect carbon footprint through manufacturing, water use, and grid draw during low-renewable periods.

It's complicated. AI can be a powerful tool for environmental good — helping model climate systems, optimise energy grids, or accelerate scientific research. But the current rapid growth of AI token consumption is also a significant and fast-growing source of energy and water use.

The goal of this site is not to say "AI is evil" — it's to make the hidden cost visible, so individuals and organisations can make more informed decisions about when and how to use AI, and so that pressure builds on AI companies to invest in efficiency and clean energy.

The numbers are estimates, not official measurements. No public real-time feed of global AI token consumption exists. We derive our figures from:

Published AI usage disclosures (OpenAI, etc.) · Epoch AI compute trend research · Academic papers on AI energy use · IEA electricity data · Microsoft sustainability reports.

The real-world figure could be higher or lower — the point is to communicate order of magnitude. The counter is live in the sense that it extrapolates forward in real time from a fixed anchor point at the current estimated global rate. See the Data Sources section above for the exact figures used.

The current estimated global AI inference rate is around 100 million tokens per second — generated by all the AI services (chatbots, search assistants, coding tools, image generators, API calls, etc.) running simultaneously across every country, every company, and every individual user on the planet.

That rate has grown by orders of magnitude since 2020, and is projected to keep growing as more AI products launch and more people adopt them. Even if one company's service is efficient, the sheer scale of global AI usage means the total climbs relentlessly.

Quite a lot, actually. Individual habits scale up when millions of people share them:

Write shorter, clearer prompts — the more specific you are, the fewer tokens the AI needs to generate a good answer.
Use smaller models — for simple tasks (summarising a paragraph, checking grammar), a lightweight model uses a fraction of the energy of a frontier model.
Don't regenerate unnecessarily — if an answer is good enough, use it. Every regeneration is another burst of tokens.
Tell organisations you care — write to your AI providers asking about their energy roadmap. Consumer pressure works.
Share this page — the more people understand the cost, the more pressure there is to improve efficiency.