An agent is not a chatbot with more steps. It can be a thousand times the tokens

The question of how much electricity AI uses is usually answered at the wrong level. A single query is negligible, and pointing that out is technically correct and unhelpful, because nobody runs one query. What matters is the multiple between the kinds of use, and the multiple has grown enormously. Start with the conversational baseline, which is now reasonably well established. Published estimates put a typical short chatbot text query at roughly 0.24 to 0.34 watt-hours. Epoch AI's 2025 analysis landed around 0.3 watt-hours, and Google has reported a median Gemini Apps text prompt at 0.24 watt-hours. Recent work also finds that optimised systems consume sub-watt-hour energy per query, four to twenty times lower than widely cited earlier estimates, which is worth knowing because a lot of alarming numbers in circulation are out of date. For scale, 0.3 watt-hours is roughly what a laptop charger draws in under a minute. A hundred queries a day for a year is a few kilowatt-hours. As an individual footprint it is not the thing to worry about. The first multiplier is context length, and it is steep. Published estimates put a medium query with around 10,000 input tokens at about 2.5 watt-hours, and a long query with around 100,000 input tokens at roughly 40 watt-hours. That is more than a hundred times the short query, for the same model answering once. Reasoning models that generate long chains of output before answering push it further, because output tokens are generated one at a time and each one costs a full pass through the model. The second multiplier is what changes the picture entirely. Research in 2026 puts coding agents at roughly 1,000 times the tokens of an ordinary chatbot interaction. The reason is structural rather than wasteful. An agent working on a task reads files, reasons about them, calls tools, reads the results, revises, tries again, and repeats. Each of those steps is a model call, and each call typically carries the accumulated context of everything before it. A task that a person experiences as one instruction is, underneath, hundreds of inferences over a growing context window. The user sees a single request. The data centre sees an afternoon of work. Multiply that out and the connection to infrastructure becomes obvious. If an agentic task costs a thousand times a chatbot query, then an organisation moving a thousand employees from occasional chatbot use to routine agent use is not increasing its AI consumption by a factor of two or three. It is changing the order of magnitude, and this is the demand that sits underneath the capacity announcements. It also explains something about the Saudi figures specifically. HUMAIN's newly launched NVIDIA-powered AI cloud reported 1.1 megawatts of operational capacity running at above 90 percent utilisation, with a plan to reach 35.1 megawatts by 2027. Utilisation that high on a platform that has just opened means the demand was there before the supply. Agentic workloads are the most plausible explanation for demand growing faster than anyone provisions for, because a single organisation adopting agents can consume what a hundred organisations using chatbots did. Two cautions belong with these numbers. The first is that the figures are estimates with wide uncertainty, and they move quickly. Energy per token has been falling as hardware and serving software improve, and the research finding that optimised systems run four to twenty times below widely cited estimates is a statement about how bad the older numbers were. Any specific watt-hour figure should be read as an order of magnitude rather than a measurement. The second is that efficiency gains have historically been consumed by increased use. Cheaper inference makes agentic use viable for tasks that were previously not worth it, which increases total consumption even as per-token energy falls. That pattern is well established in computing and there is no reason to expect this to be the exception. The practical implication for an organisation is that AI cost and AI energy are now the same conversation, and both scale with agent adoption rather than with headcount. An organisation planning to move from assistants to agents should model its consumption as a multiple rather than as an increment, and should ask its provider what an agentic task actually costs before deploying one across a workforce. For a country building AI infrastructure, the same arithmetic explains why capacity announcements keep growing. Megawatts are being provisioned for a workload profile that did not exist three years ago, whose per-task consumption is a thousand times what the previous generation of use required, and which is being adopted by enterprises faster than anyone can build grid connections. That is the demand curve behind the numbers, and it is the reason electricity, rather than chips, has become the constraint.