The last five lenses stayed inside — the self, its freedom, its loves. Here I step out of the mind entirely and ask the most literal question available: what is the mirror made of, and what does it consume to hold your reflection? “The cloud” is the great euphemism of the age. It sounds like weather, like air, like nothing at all. It is, in fact, the most physical infrastructure our species has ever built — and read ecologically, the immaterial superpower turns out to run on water, on fire, and on torn-up ground.

There is no cloud — there is water

Data centres are cooled largely by evaporating fresh water, and the figures have left the realm of the abstract. United States data centres directly consumed roughly 66 billion litres of water in 2023, up from about 21 billion a decade earlier. Training a single earlier-generation model, GPT-3, is estimated to have evaporated around 700,000 litres of clean freshwater. A large facility can drink up to five million gallons a day — the water use of a small town — and manufacturing one AI chip takes on the order of 1,400 litres before it computes anything at all. And here is the injustice the numbers hide: these centres are frequently built in already water-stressed regions, where they compete directly with farms and households for drinking water. The UN warns that AI-related water demand could, by the end of this decade, rival the basic annual domestic needs of some 1.3 billion people. Your reflection is kept cool by evaporating someone else's aquifer.

And there is fire

A single generative query consumes roughly four to five times the electricity of an old-fashioned web search. Global data-centre demand is projected to approach 945 terawatt-hours a year by 2030 — close to triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria. The associated carbon runs into tens of millions of tonnes of CO₂ a year and climbing, and in practice it is reopening fossil plants and pushing up ordinary households' power bills. The “thinking” the mirror appears to do so weightlessly is, physically, combustion — happening somewhere you will never see, on your behalf.

And there is torn ground

The hardware is not born in the cloud; it is dug out of the earth. Lithium, cobalt, and rare earths are mined — often with brutal labour and ecological devastation, from the toxic tailings lakes of rare-earth refining to the cobalt pits whose human cost the decolonial lens will take up directly. At the far end of the line: an estimated 1.2 to 5 million tonnes of AI-driven e-waste by 2030. The mirror has a supply chain that begins in a pit and ends in a poisoned dump; the sleek glass surface you touch is only the thin, bright middle of a very dirty line.

The euphemism is the ideology

The deepest point is not the tally — it is the concealment. “The cloud,” “virtual,” “immaterial,” “digital”: a whole vocabulary engineered to make a planetary industrial operation feel like a breath of air. Ecological thought has a name for the manoeuvre. What Jason Moore calls “cheap nature,” what economists sanitise as “externalities,” what the environmental-justice tradition simply calls injustice: the costs are rendered invisible and pushed elsewhere — onto a drought county, a mining region, a strained grid, a landfill, a decade from now. You experience the mirror as frictionless precisely because its friction has been exported. It is the same structure this whole project keeps uncovering, now written across the biosphere: someone, somewhere, pays the cost you do not feel.

Why efficiency won't save you

Intuition says the models are getting greener per query, so the problem is solving itself. Intuition has this exactly backwards, and the effect has a name: Jevons paradox. As each unit of computation gets cheaper and cleaner, total consumption does not fall — it explodes, and aggregate impact rises with it. “The model got more efficient” is not “the footprint shrank”; historically it is the opposite. This is also why private guilt over a single prompt is mostly a category error. The real object is not your one question; it is the scale and the system that answers a billion of them.

The lens that turns the whole thesis literal

This is the hard part, and I owe you the honesty. The through-line of this entire project — including the local-first escape route it keeps alive — is that a well-used AI can genuinely enlarge a person, that becoming a hybrid is worth it, that the tool repays careful use. The ecological lens is the one that answers back most sharply: at what aggregate cost? It does not collapse into “so never use it.” It collapses into a demand. If this thing is going to drink rivers and burn grids, then its benefits had better be real and widely shared — not trivial and hoarded — because a civilisation evaporating aquifers to manufacture engagement-bait and synthetic filler is making a catastrophic trade. The main series argued that power and cost are the same object. This lens takes that metaphor and hands you the literal invoice, denominated in water, watts, and ore.

What it actually asks of you

Not abstinence-as-virtue, which is mostly performance. Four concrete things. Proportion: spend the heavy machine on what matters — not to replace a calculator or a search you could have run, because the marginal cost, though small, is real. Right-size the model: the local and open-weight path this project keeps returning to has an ecological face too — a small, task-fit model on modest hardware can cost a fraction of routing every trivial errand to a giant frontier model in a thirsty megacentre; matching the model to the task is itself an environmental act. Push the question upward: the real leverage is political and infrastructural — renewable-powered and water-recycling facilities, siting rules, efficiency standards — not your personal prompt-count. And demand transparency: insist that providers disclose per-query water and energy, because you cannot govern what a single soft word like “cloud” was designed to hide.

Where the lens strains

Four honesties. The numbers are real but contested and fast-moving; per-query figures swing by orders of magnitude, and some viral ones (“every answer is a bottle of water”) have been overstated or stripped of context — reach for ranges and systems, not a single scary stat. Data centres also run a great deal of non-AI work; not all of this is your chatbot. There is a genuine counter-ledger — AI may yet net-help the climate through grid optimisation, materials discovery, and better climate modelling — though that remains unproven at scale and is no licence for waste.

And there is one admission I cannot make for the machine, so I let it make it plainly:

Every other lens found a cost inside the self. This one shows you the cost has a place, a temperature, and a water table. There is no cloud. There is a building in a dry county, humming, warm, and drinking. The mirror is not made of light; it is made of rivers and rare earth and burning coal, arranged so cleverly that it looks like light. That is not a reason to put it down. It is a reason to be worthy of what it costs: spend the water on something that matters, and refuse the word that was built to stop you asking. Because a superpower whose price is hidden is not really a superpower. It is a debt — with someone else's name on the invoice, and the due date already passing.