You open a chat window, type a question, hit enter.
The answer comes back in seconds. It feels instant and free.
Most people never think about what happens in the fraction of a second between the send button and the reply. Most coverage of AI costs doesn’t either.
But a set of reports published in 2025 and 2026, including one from the world’s foremost energy authority, put a number on that overlooked moment, and the figure is larger than almost anyone outside the data center industry expected.
There are two ways AI costs money, and almost everyone focuses on the wrong one
When AI made headlines for burning through budgets, the story was almost always about training: the one-time, intensive process of building a model.
Training GPT-4, for example, required approximately $78 million in compute resources alone, according to Stanford’s 2024 AI Index Report.
That number is real. It is also a one-time event.
What happens after a model ships, the continuous process of responding to every single user query, around the clock, at global scale, has a different name: AI inference.
And inference is where the economics of AI shifted hard.
The volume of what people actually do with AI every day has become almost impossible to picture
OpenAI announced on February 27, 2026 that ChatGPT had reached 900 million weekly active users and serves more than 2.5 billion queries per day.
That number compounds every time a new model launches or a new feature rolls out. And ChatGPT is just one product among dozens.
Inference now accounts for between 80% and 90% of all AI computing activity, according to multiple industry analyses.
Unlike training, which happens once per model version, inference runs continuously and scales directly with the size of the user base.
Every Gemini suggestion, every Claude summary, every image generated on demand, each is an inference event, drawing power from a GPU rack somewhere.
The energy math behind a single query is small, but the aggregate is where the picture changes
Individual queries are not expensive. A single short AI text query still consumes meaningfully more energy than a traditional web search, but remains a modest figure in isolation.
Then you multiply it by billions of queries, every day, 365 days a year.
Scaling per-query energy estimates across the billions of AI queries processed globally in 2025, analysts have found that annual inference energy consumption already exceeds the total electricity use of tens of thousands of U.S. residential households.
The International Energy Agency’s April 2026 report on energy and AI found that data center electricity use climbed 17% in 2025, while AI-specific data centers surged 50%, both far outpacing overall global electricity demand growth of 3%.
The grid does not see a chatbot. It sees a factory that never closes.
Now analysts have put a market dollar value on inference, and the number reframes everything
The AI inference market is projected to grow from $106 billion in 2025 to nearly $255 billion by 2030, according to MarketsandMarkets research.
That is not a figure for all of AI. It is specifically the cost of running AI, answering queries, serving agents, processing requests, once the models already exist.
Global data center electricity demand reached approximately 485 terawatt-hours in 2025, a 17% year-over-year increase, while AI-specific data centers grew even faster at 50%, according to the IEA’s April 2026 report on energy and AI.
The shift in the economics is structural. The five largest technology companies collectively exceeded $400 billion in capital expenditure in 2025, according to the IEA, a figure the agency expects to increase by a further 75% in 2026.
Understanding where AI spending goes, not into building models but into serving them, is the economic insight that most product announcements still obscure. And as user bases balloon past a billion, the inference tab compounds relentlessly.
The overlooked tap on a keyboard turns out to be the engine of a quarter-trillion-dollar industry.
The efficiency gains are real, but they may not slow the meter down
The IEA notes that power consumption per AI task is declining rapidly, with efficiency improving at a rate the agency describes as unprecedented in energy history. Inference costs for models matching earlier performance benchmarks have dropped by orders of magnitude in just a few years, according to Stanford’s 2025 AI Index.
That is genuine progress. But cheaper inference per query tends to unlock more inference overall, not less.
Energy-intensive use cases, video generation, autonomous agents, reasoning models, consume hundreds to thousands of times more per query than simple text generation, and their adoption is accelerating.
Accelerated servers, the GPU-dense racks that handle AI workloads, are projected to see electricity consumption from AI-focused data centers triple between 2025 and 2030, driven primarily by inference volume growth, according to the IEA.
The good news is that the industry knows exactly where the pressure is building. Engineers and policymakers are now racing to build smarter chips, more efficient cooling and purpose-built software, all aimed squarely at the inference layer, the part of AI that runs every time anyone, anywhere, asks a question.
That single text box turned out to be the most consequential real estate in the global energy economy.
