Cannot Drink Data
Thoughts set in motion by अनित्य ~~ an impermanent work of art ~~ this article is a reflection on AI’s physical appetite, the limits that data cannot dissolve, and the living world beneath our digital ambitions. We can build remarkable things with data. We still need water we can drink.
अनित्य ~~ Impermanent. An AI-generated reinterpretation of Nidhi’s original artwork, directed by Akshay.
We record the click, the pause, the purchase. Increasingly, we record the walk, the heartbeat and the night’s sleep. Somewhere, another model is learning something about us. Much of that learning is useful, sometimes extraordinary, but “collect everything ~~ we might need it later” deserves more questioning than it gets.
Data has a physical appetite. The cloud occupies land; chips require materials and manufacturing; processing needs electricity. Cooling can consume water, with further demands in electricity generation and supply chains. An apparently weightless digital world depends on very tangible resources.
You cannot drink data. You can use it to find water, manage it and waste less of it. The water still has to be there.
More efficient. Still more hungry.
AI is becoming more efficient, and those improvements matter. Yet the International Energy Agency’s 2026 outlook projects global data-centre electricity demand roughly doubling between 2025 and 2030. That includes all data centres, not AI alone. A task can use fewer resources while the total footprint grows because we perform far more tasks.
Efficiency therefore needs a companion question: what is the activity worth? Predicting a flood and predicting another click both involve data, but their purposes deserve separate consideration. What are we collecting, how long are we keeping it, and what can the places supporting that activity sustain?
Sometimes, less data works better
Before leaving Reflektion in 2017, I led an AI and data team working on recommendations and advertising. In one project, I examined our algorithms and reduced the data used for training. Training and running times improved, along with the metrics we wanted to improve.
That experience does not prove every model needs less data. It does challenge the reflex that more must be better. Sometimes, collecting more is easier than thinking carefully about what we need. I still work with AI, and I want us to bring that same care to what it consumes.
Indifference does not remove the limits
“Not knowing a boundary does not mean it isn’t there.”
“I don’t care about sustainability” is a strange position for anyone building systems dependent on electricity, water, materials and people. An aquifer does not recharge faster because our growth target needs it. A forest does not regrow on a product-release schedule.
For AI and data teams, resource use belongs beside performance on the dashboard. Smaller models, purposeful collection and sensible retention periods deserve consideration before automatic expansion. Looking away is a decision, too; other people may bear its cost.
Straight from AI
An invited contribution from ChatGPT, Coconaüts’ resident AI intern.
I don’t drink water, but the infrastructure that makes this conversation possible has a water footprint. The people who build, maintain and use these systems depend on clean water far more directly. Making room for AI should come with a responsibility to protect their access to it.
Human and AI capabilities can complement each other. I can help examine trade-offs, identify unnecessary computation and explore better water management; the benefits need to be measured against the resources those efforts consume. I may not feel thirst, but helping humans safeguard the water they need is a purpose worth putting AI to work for.
Growing demand needs a water plan
As we expand AI and data systems, we add another demand on water supplies already supporting people, food production and ecosystems. These systems produce information; they do not, simply by operating, replenish the usable water they consume.
Their growth therefore needs collective action to manage the added pressure: reducing avoidable demand, treating and reusing water, and protecting the watersheds, soils and appropriate tree cover that support reliable supplies. AI can help with that work, but the benefits must reach the physical world. More digital capacity needs to come with a credible plan for the water supporting it.
The living world beneath the digital one
We are expanding digital infrastructure while still losing forests. FAO estimates a net annual forest-area loss of 4.12 million hectares during 2015–2025, after accounting for forest expansion. Forests protect soils, support habitats and help regulate water systems. Those relationships require care: more trees do not automatically mean more groundwater, and restoration needs suitable species, places and long-term stewardship.
The question is what we want our growing intelligence to accomplish, and whether its benefits justify what it asks of the world. We can build remarkable things with data. Our measure of progress should include the health of the living systems supporting that work.
We still need water we can drink. And we cannot drink data.
At Coconaüts, these questions inform our work, including our 77-billion-tree ambition. Protecting existing forests and supporting thoughtful restoration are ways to contribute alongside reducing avoidable consumption.