Your AI drinks a bottle of water for every email: what on-device AI does to the footprint

Does an AI email really cost a bottle of water?

Roughly, yes. A UC Riverside analysis reported by the Washington Post found that a single 100-word email written by a large cloud model can consume about a bottle of water once you count the electricity and the cooling. The water is not in the email. It is in the power station that fed the data centre and in the cooling systems that stop the chips from melting. Multiply that by billions of prompts a day and the number stops being a novelty and starts being an infrastructure problem.

That is the real story. Every hyperscaler prompt takes a round trip to a large, distant, often thirsty data centre. On-device AI, running inference on hardware the operator already owns and sized to the actual workload, cuts out that round trip. It does not make the footprint zero. It makes it smaller, local, and measurable, which is the part a regulated buyer can actually govern.

How big is the water problem, really?

The researchers behind the analysis, Li and colleagues, project that AI infrastructure will withdraw between 4.2 and 6.6 billion cubic metres of water a year by 2027. To put that in British terms, the upper end is about half of the United Kingdom's entire annual water withdrawal, for one industry, for one use. The siting makes it worse. Of 809 planned US data centres in the study, 517 sit in drought-affected regions. That is not an accident. Data centres chase cheap land and cheap power, and those often sit exactly where water is already scarce.

Water is only half of it. The same cooling and compute burn electricity, and a lot of that grid is still carbon-heavy. So the cost of a cloud prompt is really two bills paid in different currencies: litres and kilowatt-hours. You never see either. They are absorbed into your subscription and someone else's watershed.

Why does hyperscaler inference cost so much water and power?

Three reasons stack up. First, scale of hardware. Frontier models run on enormous GPU clusters that draw serious power even at idle, and every watt of compute turns into heat that has to be cooled. Second, the round trip. Your prompt travels to a data centre, gets queued, processed, and returned, and the whole facility stays hot and thirsty whether your specific request needed that much muscle or not. Third, overprovisioning. A general-purpose cloud model is sized for the hardest possible query. Most emails, summaries, and lookups do not need that. You pay the peak cost for an average job.

The uncomfortable truth for buyers is that none of this shows up on your desk. You cannot audit a footprint you cannot see, and you cannot reduce one you do not control.

Does on-device AI actually shrink the footprint?

It shrinks the parts you can see and size. Running inference locally does three concrete things. It removes the network round trip to a distant facility. It lets you match the model and the hardware to the real workload instead of the theoretical worst case. And it puts the power draw on a meter you own, in a building you run, on a grid you can choose to green.

Here is the honest limit, because the topic deserves it. On-device AI still uses electricity, and that electricity still comes from a grid that may still burn gas and still consume water for cooling somewhere upstream. Local hardware needs cooling too. Nobody running inference is drinking rainwater. The point is not zero footprint. The point is a smaller, local, measurable footprint that you can put on a dashboard, attribute to a workload, and reduce on purpose. Measurable and owned beats invisible and rented every time a sustainability team has to sign a report.

What can a regulated organisation do about it?

Right-size the model to the task. A tuned smaller model on owned hardware often answers the same routine query for a fraction of the energy a frontier cloud model would spend. Keep the workload local so it runs on infrastructure you already power and cool. And measure it. If your AI runs on kit you control, its energy draw becomes a line you can report against, not a figure you have to take on trust from a supplier.

This is where sustainability and sovereignty turn out to be the same conversation. The reasons a defence, finance, healthcare, or government buyer wants AI inside their own walls, that is control, auditability, and no dependence on a distant provider, are the same reasons that make the footprint governable. You cannot manage what you cannot see, and you cannot green a data centre you do not run.

Where does Mickai fit, honestly?

Mickai is a British Sovereign Intelligence Operating System. It runs AI on hardware the operator owns, inside their own walls, offline, with every consequential action sealed into a post-quantum signed audit ledger. Fifty brains and around sixty studios run locally instead of calling out to a hyperscaler. That architecture, built for sovereignty and audit, has a side effect that matters here: the inference happens on your meter, sized to your workload, with no round trip to a thirsty data centre in a drought region.

We will not pretend that makes the footprint disappear. Local compute still draws power, and power still has a water and carbon cost through the grid. What changes is who can see it and who can act on it. With Mickai, the energy behind your AI is a number you own and can put in front of an auditor, not a litre you never knew you spent. For organisations that already have to report their footprint, a local, measurable, operator-controlled AI is not a green slogan. It is simply the version you can stand behind.

Mickai does not sell zero. We give regulated operators an on-device alternative where the footprint is small, local, and yours to measure and manage.

“A UC Riverside analysis found a single 100-word cloud AI email can consume roughly a bottle of water once power and cooling are counted.”

Frequently asked questions

Does a single AI email really use a bottle of water?

Roughly, yes. A UC Riverside analysis reported by the Washington Post found a 100-word email from a large cloud model can consume about a bottle of water once you count the electricity generation and the cooling that keeps the data centre running. The water is upstream, in the power plant and the cooling systems, not in the message itself.

How much water will AI use by 2027?

Researchers led by Li project AI infrastructure will withdraw between 4.2 and 6.6 billion cubic metres of water a year by 2027. The upper end is about half the United Kingdom's total annual water withdrawal. Siting makes it worse: 517 of 809 planned US data centres in the study sit in drought-affected regions.

Does on-device AI have zero water and energy cost?

No, and we will not claim it does. Local inference still draws electricity from a grid that may burn gas and consume water for cooling upstream, and local hardware needs cooling too. The benefit is a smaller, local, measurable footprint you control, not a footprint of zero.

Why is running AI locally more efficient than the cloud?

Three reasons. It removes the network round trip to a distant data centre. It lets you match a smaller model to the real workload instead of paying the peak cost of a general-purpose frontier model. And it puts the power draw on a meter you own, so you can measure and reduce it on purpose.

How does Mickai reduce the AI footprint?

Mickai runs inference on hardware the operator owns, inside their own walls, offline, sized to the workload. That avoids the round trip to a thirsty hyperscaler data centre and puts the energy cost on your own meter, where it becomes a number you can audit and manage rather than one you take on trust.

Is sustainability the same as sovereignty here?

Largely, yes. The reasons a regulated buyer wants AI inside their own walls, control, auditability, and no dependence on a distant provider, are the same reasons that make the footprint governable. You cannot manage what you cannot see, and you cannot green a data centre you do not run.