OPINION
The public conversation about artificial intelligence (AI) infrastructure remains too narrowly scoped.
We measure power consumption, debate on carbon emission, and often treat water as a secondary operational detail.
That is a mistake.
Large language models depend on accelerator clusters, storage, networking and cooling plant.
Their water burden is not confined to simply the amount entering a data centre.
Water may be consumed directly through evaporative cooling and indirectly through electricity generation, semiconductor fabrication, construction and equipment supply chains.
A facility that looks dry at its boundary can still shift a substantial burden elsewhere.
This is why the phrase ‘zero-water AI’ demands discipline.
It can credibly describe the absence of routine consumptive water in a cooling process.
It cannot describe an entire AI system unless the electricity, hardware, construction and local water-scarcity consequences have also been counted.
Put each workload where it belongs
My research proposes a geographically asymmetric architecture.
High-density model training would be concentrated in climate-qualified terrestrial facilities using single-phase dielectric immersion, a plate heat exchanger and closed-loop dry cooling.
Most interactive inference would remain at metropolitan or regional gateways.
A narrow class of small, bounded inference tasks could operate on low Earth orbit nodes where local execution offers a measured advantage.
The principle is simple: training and inference do not have identical physical requirements.
Training needs tightly coupled accelerators, high-bandwidth networks and sustained heat rejection.
It can also tolerate some scheduling flexibility.
Operators can shift delay-tolerant work towards periods with more favourable electricity, carbon and water conditions.
Inference is different. It often benefits from proximity, but proximity does not automatically mean orbit.
Terrestrial gateways should handle most requests.
Low Earth Orbital execution (similar to Starlink and Amazon LEO for broadband internet) should be reserved for remote users, autonomous missions or services for which a distant terrestrial processing path creates a demonstrated disadvantage.
Removing evaporation is valuable, but not sufficient
In the terrestrial design, processors are immersed in a sealed dielectric liquid.
Heat passes through an exchanger into a separate glycol loop and is rejected to ambient air by dry coolers. No routine evaporation is required.
For a modelled 1 MW computing module, the first-order design requires substantial but physically plausible fluid circulation and heat-exchanger capacity.
Under the study’s sensitivity assumptions, power usage effectiveness ranges from 1.07 on a favourable cold day to 1.15 under design-day conditions.
Those values are not promises. They show why annual weather, extreme heat, smoke, icing, fan power, pump losses and equipment redundancy must be tested before a site is selected.
The water arithmetic is nevertheless compelling.
At a continuous 1 MW computing load, replacing a cooling system that consumes 1 litre per kilowatt-hour would avoid an arithmetic equivalent of 8.76 million litres each year.
At 100 MW, that becomes 876 million litres.
These are scenarios, not forecasts, because the true counterfactual depends on local climate and technology.
The industry should report cooling water, source-electricity water and embodied water separately.
The international WUE standard provides a useful operational measure, but a single facility ratio cannot settle the wider sustainability question.
A litre consumed in a water-stressed basin is not equivalent to a litre consumed where supply is abundant.
Low Earth Orbit is a specialised edge, not an escape hatch
Space-based inference sounds extravagant until its limits are stated clearly.
A spacecraft cannot reject heat through external convection.
It must radiate internally generated and absorbed environmental heat into space while operating within strict electrical, thermal and radiation constraints.
In my conservative 6U screening case, 16.9 watts remained for computing after a 20-watt spacecraft-bus allowance.
That narrow envelope aligns with a recent 12-watt orbital edge payload that reported a 124-day flight campaign and bounded generative-AI demonstrations.
This is evidence of limited feasibility, not permission to move general-purpose AI into constellations.
Satellite manufacture, launch, replenishment, debris, atmospheric effects, ground terminals and end-of-life obligations all belong inside the environmental boundary.
So do radiation faults, signed model updates, rollback, privacy, spectrum rules, affordability and responsibility when a service fails.
Measure first, scale later
The next step should not be a grand infrastructure announcement. It should be an evidence-based research programme.
We need a rack-scale immersion and dry-cooling demonstrator tested against measured annual weather and water data.
We need a digital twin that jointly schedules computing, pumps, fans, grid carbon and grid water intensity.
We need flight-equivalent evaluation of model accuracy, energy per token, radiation faults and thermal cycling.
We also need network-in-the-loop comparisons between orbital inference; satellite backhaul and terrestrial edge services.
Only then should life-cycle assessment determine whether scaling reduces scarcity-weighted water and total environmental harm while delivering affordable, reliable access.
The question is no longer simply which data centre is most efficient.
It is which workload belongs in which location, under whose governance, and against what evidence.
If the technology sector answers that question honestly, AI infrastructure can become materially more responsible.
If it celebrates ‘zero water’ at the facility gate while exporting the burden elsewhere, it will merely improve the accounting without serving its real purpose.
Tio Chaks is a member of the ACS Sustainability Technical Committee.
