Everyone is asking what AI costs per month. Almost nobody is asking what it costs per kilowatt-hour — and for the wave of AI agents arriving next, that is the number that is about to matter.
Here is the part the pitch decks skip.
A single query is cheap. Agents do not send a single query.
A single ChatGPT-class query burns roughly 0.3 watt-hours of electricity and a few tenths of a gram of CO2. Google's own published figure for a median Gemini text prompt is 0.24 Wh, 0.03 grams of CO2, and 0.26 millilitres of water. Individually, forgettable.
Except agents do not send one query. They loop.
Independent research puts agentic workloads at 10 to 100 times the token consumption of a single chat turn — because every step is another inference: plan, call a tool, check the result, re-plan. Production teams report agents that launched at 200 tokens per request quietly climbing past 10,000 within weeks, as history, tool output, and system prompts compound.
Now point that loop at coordination — a fleet of robots, drones, or software agents that all have to move without colliding. Every agent, every tick, another decision, another inference. The token meter does not just run. It compounds with the size of the fleet.
We measured it
On an identical coordination task, we ran three arms: no coordination, our deterministic stack, and a frontier LLM making the coordination decisions one at a time. The LLM arm burned 1,134,852 tokens to complete a single mission — and finished only 20% of its missions, at more than five seconds per decision.
Our platform did the same mission for zero tokens of coordination, at 118 milliseconds per decision, completing 100%.
A million-plus tokens is the energy of roughly 2,000-plus ChatGPT-class queries — call it about half a kilowatt-hour and 0.2 kg of CO2 to coordinate one small fleet, one time. (Derivation: ~1.13M tokens ÷ ~500 tokens per query × ~0.22–0.3 Wh per query ≈ 0.5–0.68 kWh; at a world-average grid intensity of ~0.45 kg CO2/kWh, that is ~0.2 kg CO2. We use the conservative end.)
The trick is not a smaller model
It is not using a model for the part that was never a language problem.
Coordination — who goes where, who yields, who re-routes when a unit drops — is deterministic arithmetic. We let the AI do the one thing it is genuinely good at (turning a plain-English mission into a structured plan, once) and hand the moment-to-moment decisions to math that spends nothing. So the AI cost is a fixed, one-time compile — not a meter that spins faster every time you add a robot.
To be precise, because precision is the whole brand here: we are not a zero-AI system. We are an AI-once system — the model runs at the point of intent, not in the loop, forever. That single distinction is the entire efficiency story.
Now scale it
The IEA projects data-center electricity roughly doubling to about 945 TWh by 2030 — around 3% of all electricity on Earth — with AI-accelerated compute growing about 30% a year. Agentic AI is the accelerant.
So take the measured saving — ~0.5 kWh and ~0.2 kg CO2 per coordinated mission — and pick the adoption number you actually believe:
- 10 million agent-coordination missions a year → ~5 GWh and ~2,000 tonnes of CO2 avoided.
- 100 million → ~50 GWh and ~20,000 tonnes.
- 1 billion → ~500 GWh and ~200,000 tonnes.
At a billion missions, that ~500 GWh is roughly the annual electricity of 46,000 US homes, and the ~200,000 tonnes of CO2 is roughly 43,000 gasoline cars taken off the road — every year. (US-home and per-car figures from EIA and EPA averages; scaling is linear in missions.)
We are not going to tell you a billion is the right number. Nobody knows the global count of agent-coordination decisions, and anyone who quotes you one to the significant figure is guessing. That is exactly why we show the whole table instead of a single headline: the per-mission saving is measured; you supply the belief. And the benchmark mission is a small coordination task — treat the per-mission figure as a unit, not a universal constant.
Why this is not just an invoice story
For anyone building agents that coordinate at scale, the token bill and the carbon bill are the same bill. Deterministic coordination cuts both at once — not by capping capability, but by refusing to spend a language model's energy on a problem that arithmetic solves faster and more reliably.
That matters most at precisely the moment the grid can least afford another exponential. The cheapest kilowatt-hour, and the greenest, is the one you never had to spend.
If you want to check any number above, the benchmark is public, the demo runs in sixty seconds, and every claim — including the ones we retired — is on the ledger.