· 206 models · 4 task types · NVIDIA H100 80GB
Measuring the energy footprint of modern AI inference
Every time you send a message to an AI, a physical server somewhere draws electricity to process your request. That electricity comes from the power grid and depending on where the data center is located, it may come from coal, gas, or renewables. The carbon cost is real, even if invisible.
AI's energy demand is growing rapidly as models become more complex and widespread. Understanding that demand is the first step toward making informed choices about how we use AI and how we can mitigate its environmental impact.
Data centers are among the fastest growing electricity consumers in the world. In 2024, global data centers used 415 TWh in total and AI optimized servers accounted for 80 TWh, which is 19% of that total. By 2025 that share had grown to 93 TWh. By 2030, AI servers alone are projected to consume 432 TWh, nearly 44% of all data center electricity.
The bar chart ranks the 10 highest energy developers by combining the Wh across all their benchmarked models. OpenAI leads by a wide margin because it has the most models in the dataset.
The chart below zooms into individual models and reveals that all top 10 are Reasoning type. This is not a coincidence. Some AI models can be configured to think at different depths before giving an answer. This process is called chain of thought reasoning and the amount of thinking allowed is called the thinking budget.
Low — the model answers quickly with minimal deliberation. Least energy used.
Medium — the model considers the problem more carefully. Moderate energy.
High — the model deliberates extensively, verifying its logic before responding. Most energy.
Watt(Wh) is an abstract unit for most people. One way to make it tangible is to compare it to a standard 10W LED bulb, which consumes 10 Wh per hour. Dividing a model's energy by 10W gives the number of seconds that same energy could keep a LED lit. Click any dot in the chart below to start a live countdown and watch the light run out.
CO₂ is never measured directly. It is estimated by multiplying energy by the carbon intensity of the electricity grid. Reasoning models carry a much heavier environmental cost than standard models because chain of thought inference generates hundreds to thousands of additional tokens internally before producing a visible response, consuming 150 to 700 times more energy on the same hardware.
Every query sent to an AI model consumes electricity—and every watt consumed leaves a carbon trace. Most developers spread their work across multiple task types. A single company might maintain text models, image generators, reasoning engines, and speech tools all at once. Meanwhile, different models performing the exact same task can consume vastly different amounts of energy depending on who built them. Filter by developer to inspect one company's full lineup. You can see which task types they prioritize most, and identify which models and task types release the most CO₂ and create the greatest environmental impact.
This scatter plot maps 100 models across four task types (reasoning, text, image, and speech), plotting each one by its carbon cost per 1,000 queries.
Every time you ask ChatGPT a quick question, whisper a prompt to Claude, or generate an image with Midjourney, a small puff of carbon enters the atmosphere. Here, you can see those puffs accumulate in real-time. Simulate your daily routine: Send 1 query , 10 , 100, or 1,000.
Watch as bubbles rise and fill the container. Each one representing the invisible CO₂ cost of that interaction. The bigger the bubble, the heavier the emission.
AI is growing faster than most people realize, and so is its electricity bill. Most of that cost is invisible to the people generating it. Every conversation, every query, every generated image draws power and emits carbon, but users rarely think about the environmental weight of a single interaction.
The charts reveal several patterns worth paying attention to. Even within the same task category, energy consumption can vary enormously depending on which developer built the model and how it was architected. Two reasoning models can sit at opposite ends of the carbon scale. Some models offer low, medium, and high thinking modes that match energy use to problem complexity, but many users reach for the most powerful setting regardless of what the task actually requires. Developer strategies also differ widely: some companies release dozens of models spanning multiple task types, while others put nearly all their usage into a single model at massive scale. None of this means AI should be avoided, but recognizing that model choice, task type, and thinking depth all carry real environmental costs is the first step toward using these tools more deliberately.