The Thirst of the Machine

The word “cloud” is the most successful piece of branding in the history of infrastructure. It takes a building the size of a big-box store, packed with racks that draw the power of a small town and the water of a small city, and it renders the whole thing as vapour. Weightless. Placeless. Somewhere up there. The genius of the metaphor is that it makes the most physically demanding industrial process of the decade sound like weather.

It is not weather. It is a building, on a specific piece of land, drawing from a specific grid and a specific river. A data centre is a claim on a watershed and a power supply, made by a firm headquartered somewhere else, justified by a demand curve drawn by the AI industry. The cloud is the cleanest enclosure of physical commons since the railways, and it is being waved through with less scrutiny than a strip mall.

Compute is a physical good

The training run that produces a frontier model is among the most energy-intensive computational tasks ever performed by a private firm. The inference that follows, the day-to-day business of answering queries, is less intensive per event and vastly larger in aggregate, because the number of queries is enormous and rising. Both stages happen in data centres. Both stages turn electricity into heat, and the heat has to go somewhere, and getting rid of it is where the water comes in.

A large data centre running evaporative cooling can consume, by widely reported estimates, on the order of millions of litres of water a day. The water is evaporated to shed heat, which means it leaves the watershed rather than returning to it. The electricity it draws is measured in the same units as municipal demand, because it is municipal demand, simply concentrated behind a single private meter. None of this is hidden in any technical sense. It is hidden in the rhetorical sense, by a vocabulary that calls the facility a cloud and the consumption a footprint, as though it were a faint impression left in soft ground rather than a hard draw on a finite system.

The physicality matters because physical things are governed by physical politics. Water rights. Grid capacity. Land-use approval. Rate structures. These are old, well-developed areas of public law, and they are exactly the areas the industry would prefer to keep out of the conversation, because the conversation they want is about innovation and the conversation those laws produce is about who gets the water when there is not enough.

Quebec is the frontier

Canada has a specific and underappreciated position in this. Quebec has abundant, cheap, low-carbon hydroelectricity, which makes it one of the most attractive jurisdictions on the continent for a power-hungry data centre. The same attributes that made the province a magnet for cryptocurrency mining in the 2010s now make it a magnet for AI compute, and the province has been here before.

When crypto mining surged, Hydro-Québec found itself fielding connection requests whose combined demand was a meaningful fraction of the province’s surplus. The utility moved to ration access, to prioritize loads that produced more local economic value per megawatt, and to stop treating its grid as an unlimited free input for any firm that showed up with servers. That episode is the template for what is now happening with AI, at a larger scale and with a more politically powerful set of buyers.

The question Quebec faces is the question every hydro-rich jurisdiction faces. Cheap public power is a public asset, built over generations with public money, and it is finite. Every megawatt sold at an industrial discount to a hyperscaler’s training cluster is a megawatt not available to electrify home heating, transit, and industry, which are the loads that actually decarbonize an economy. A data centre that runs a chatbot is not, on any honest accounting, a higher public priority than a heat pump in a Montreal triplex. But the data centre arrives with a flexible cheque and a press release about jobs, and the heat pump arrives as a line in a climate plan, and the cheque tends to win.

The subsidy nobody calls a subsidy

Here is the mechanism that should bother anyone who pays a power bill. When a public utility sells electricity to a large industrial customer below the marginal cost of new supply, the difference is a subsidy. When that utility then has to build new generation to meet the load it just took on, the cost of the new generation is socialized across all ratepayers, while the benefit of the cheap power is captured by the single firm that signed the contract. Public cost, private benefit, intermediated by a rate structure that almost nobody reads.

The same logic governs the water. A facility that draws heavily on a municipal supply, or on a watershed shared with farms and towns, is consuming a common-pool resource whose price rarely reflects its scarcity. In a wet year, nobody notices. In a drought year, the data centre’s cooling demand competes directly with agriculture and households, and the competition is settled by contracts written when water looked free.

This is enclosure in the textbook sense. A commons that was governed loosely because it was abundant is converted, through long-term private contracts, into a dedicated input for a single firm, and the conversion is justified by an economic-development story that does not survive contact with the actual job numbers. A hyperscale data centre employs remarkably few people once built. It is a capital asset, not a workforce. The jobs argument is the wrapping paper, and it is thin.

What public governance would look like

The tools to govern this already exist, because we have governed land, water, and power for a long time. They are not being used, because the industry has successfully framed compute as something new enough to escape them.

A serious public framework would price industrial power at the marginal cost of the new supply the load requires, so that ratepayers do not subsidize the buildout. It would require water-use disclosure and watershed impact assessment for any facility above a threshold, the same way we require it for other heavy industry. It would condition land-use approval on a real, audited accounting of local economic benefit rather than a press-release headcount. And it would treat surplus public hydro as a scarce decarbonization resource to be allocated by public priority, not auctioned to whoever industrializes fastest.

None of this is anti-technology. It is the ordinary governance of heavy industry, applied to an industry that has been allowed to pretend it is weightless. A steel mill that wanted this much power and water would face every one of these questions. The data centre faces almost none of them, and the only reason is the metaphor.

The argument

The cloud is a building. It sits on land, draws from a grid, and evaporates water from a watershed, and it does all of this at industrial scale while being described in the language of weather. The AI boom is, in its physical reality, a vast new claim on power and water, and those are public, finite, and already spoken for by the unglamorous work of decarbonizing how people live.

Quebec is where this is most visible, because Quebec has the cheap public hydro that makes it the prize. The province has already learned once, with crypto, that a public grid is not an unlimited free input. The lesson now has to be applied to a buyer with far more money and far better lobbyists.

Price the power at what new supply actually costs. Disclose and assess the water. Allocate the surplus by public priority. Stop letting a finite commons, built with public money over generations, be enclosed one industrial contract at a time by firms that call the whole thing a cloud so that nobody thinks to ask whose river it is.