The artificial intelligence boom has largely been framed in terms of capability — smarter models, broader applications, and rapid adoption across industries. But a new set of estimates is sharpening attention on a less visible constraint: the sheer amount of electricity required to keep that momentum going.
A recent analysis from BestBrokers suggests that the infrastructure behind ChatGPT alone is already operating at a scale comparable to national energy systems. The numbers are necessarily based on assumptions — including user activity levels and model efficiency — but they offer a useful lens on how quickly AI demand is translating into real-world power consumption.
At the centre of the report is a simple question: what does it take, in energy terms, to answer the world’s growing volume of AI queries?
From billions of prompts to terawatt-hours
BestBrokers estimates that ChatGPT now serves around 3.2 billion prompts per day, based on roughly 900 million weekly users and average engagement of about 25 queries per user each week. That implies more than 1 trillion prompts annually — a scale that would have been difficult to imagine even a few years ago.
Processing that volume of queries is estimated to require about 60.7 gigawatt-hours (GWh) of electricity each day, or roughly 22 terawatt-hours (TWh) per year.
To put that in context, 22TWh is in the range of annual electricity consumption for smaller countries or large metropolitan regions. The report notes that this level of demand could power around 2.1 million US homes for a year, or exceed the annual electricity use of several European nations.
Even when compared with major economies, the numbers are striking. Annual ChatGPT energy use would equate to:
- Nearly 20 hours of China’s electricity consumption
- Close to two days’ worth of US power demand
- Around a week of Japan’s electricity use
These comparisons are illustrative rather than exact, but they underline a broader point: AI workloads are no longer marginal additions to global electricity demand — they are becoming a meaningful component of it.
The cost of inference
The report focuses specifically on inference — the energy required to generate responses to user prompts — rather than training large models, which is itself highly energy-intensive but occurs less frequently.
Based on estimated power usage for GPT-5-level systems, each prompt consumes around 18.9 watt-hours on average. That is significantly higher than a traditional web search, typically estimated at about 0.3 watt-hours per query.
In practical terms, generating an AI response may require more than 50 times the electricity of a standard search.
Aggregated across billions of daily interactions, the gap becomes material. Using average US commercial electricity prices of around US$0.136 per kilowatt-hour, BestBrokers estimates annual energy costs of roughly US$3 billion just to serve ChatGPT prompts.
That figure does not account for capital expenditure on data centres, cooling systems, or network infrastructure — all of which add to the total cost base of operating large-scale AI systems.
The energy footprint also translates into a meaningful carbon cost. BestBrokers estimates that ChatGPT’s annual electricity use could generate roughly 6 million tonnes of CO₂ — comparable to putting around 1.3 million cars on the road for a year, or matching the emissions profile of some smaller national economies. As usage scales, that environmental footprint is moving from a secondary consideration to a more central part of the AI infrastructure debate.
A new layer of demand for data centres
The implications extend beyond any single platform. As AI adoption accelerates, data centres are emerging as one of the fastest-growing sources of electricity demand globally.
“The AI boom is often framed as a software revolution, but behind every chatbot response lies a vast network of energy-hungry data centres,” BestBrokers analyst Alan Goldberg said.
“As AI systems become more capable — handling images, audio, and increasingly complex reasoning tasks — the amount of computation required for each interaction is likely to grow further.”
That trajectory is already visible. Newer models are moving beyond text into multimodal tasks, while enterprises are integrating AI into workflows that require continuous, high-volume processing. Each step increases the computational load — and by extension, the energy required to sustain it.
Infrastructure, not just innovation
If usage continues to scale — and if models become more complex — then AI’s growth will be tied as much to energy infrastructure as to software development. That has direct implications for sectors ranging from utilities and grid operators to semiconductor manufacturers and data centre providers.
It also introduces new constraints. Power availability, energy pricing, and grid stability are likely to become increasingly important variables in determining where and how AI infrastructure can expand.
In that sense, the AI race is not just about building better models. It is also about securing the energy systems needed to run them.
The BestBrokers analysis may rely on broad assumptions, but it highlights a shift that is becoming harder to ignore: as artificial intelligence moves deeper into the global economy, its footprint is starting to look less like a purely digital phenomenon — and more like a large-scale industrial one.