The Digital God Doesn’t Run on Faith; It Runs on Water

AI is not running in the cloud. It is running on infrastructure, and that infrastructure has a cost. Nobody talks about what it costs to think at scale. That silence is doing work. AI is being sold as a technology story. It isn’t. It is a resource story; and the ledger is missing. Every response is backed by systems drawing power without interruption, consuming water at industrial scale, and expanding without a clear accounting of its cost.

I work in law; precedent, procedure, judgment built over years. I am told, with increasing confidence by people who have never read a case law, that AI will replace what I do inside a decade. Maybe it will. What is never addressed is what it takes to run the systems making that prediction; what they consume, where that cost lands, and whether anyone has done the math before building the narrative.

The Free Pass Is Over, Digital God Doesn’t Run on Faith

For years, computing got a free pass on the environmental question. Data felt clean. It wasn’t steel, it wasn’t diesel, it wasn’t a smokestack you could photograph. It was light and fast and it lived somewhere called the cloud, which sounds deliberately weightless. That framing served the industry well for a long time; it let infrastructure expand without the kind of public scrutiny that follows a refinery proposal or a pipeline application. The product was invisible, so the cost felt invisible too. That assumption no longer holds, and the industry knows it, which is part of why the sustainability language has gotten so elaborate while the underlying numbers stay so hard to find.

Here’s what actually sits behind a response: physical servers drawing continuous power, cooling systems pulling millions of litres of water daily depending on design and climate, transmission infrastructure upgraded on the public dime to meet private load, and a semiconductor supply chain that burns through ultra-pure water and high-energy chemical processing before a single server ever gets racked. None of that appears in the product announcement. The gap between what gets marketed and what gets consumed is where the real conversation needs to happen.

Efficiency Is Not a Ceiling

The efficiency argument is the one the industry reaches for first, and it’s not wrong; it’s just incomplete. Better chips help. Smarter scheduling helps. Shorter training runs help. What they don’t do is set a ceiling. They lower the floor. And when the floor drops, use expands to fill the space. This is not speculation; it’s the entire history of computing. Processing power improves, demand rises to meet it, total consumption climbs regardless of per-unit gains. AI doesn’t break that pattern; it accelerates it by embedding computation into decisions that used to require no computation at all. A search query that once returned ten blue links now triggers inference at scale. A document that once got a human skim now gets AI analysis as a default step. Multiply that across enterprise deployments running continuously and the load doesn’t look like a series of requests; it looks like a baseline.

The Water Question Nobody Is Asking

Water is where the accounting gets uncomfortable, partly because it sits outside the energy conversation most people are already having. Cooling is a physical necessity at data centre density; heat has to go somewhere, and at the volumes modern AI infrastructure generates, it goes into water. Liquid cooling at scale means real water volumes; not metaphorical, not offset, actual water drawn from actual regional systems and returned to those systems warmer than it left. In arid regions that equation becomes a competition. In temperate regions it becomes a municipal planning question that rarely gets asked until the facility is already built and drawing.

In Canada, as of 2026, more than thirty long-term drinking water advisories remain active in First Nations communities. Some of those communities have been navigating boil-water conditions since the mid-1990s; not as a temporary disruption but as a sustained, documented, politically acknowledged failure that successive governments have committed to fixing and haven’t. I’ve worked directly with First Nations communities for seven years. The water advisory isn’t an abstraction you encounter in a government dashboard; it’s a conversation you have at a kitchen table with someone who fills a pot from a jug because the tap isn’t safe. Industrial cooling systems in the same country operate continuously without interruption. I’m not suggesting a direct causal link between a data centre in a southern Ontario industrial park and a water advisory in a northern community. I am saying the resource is the same resource, the infrastructure gap is measurable, and the allocation question deserves to be asked plainly rather than buried in a sustainability report nobody reads past the executive summary.

Where the Costs Actually Land

Site selection is where these trade-offs crystallize into something concrete. Data centres follow cheap power, available land, and acceptable latency. In Canada that pull has concentrated development toward regions with hydroelectric capacity or flexible generation; good for cost, good for the renewable energy narrative, consequential for the grids and water systems in those regions. The clustering that results is efficient from a performance standpoint. From a grid standpoint it concentrates demand on the same nodes, the same water intake points, the same transmission corridors that were not designed to carry continuous industrial load from facilities that never go offline. Local infrastructure gets upgraded to absorb that demand. The engineering costs are real. They land on the public side of the ledger while the compute capacity lands on the private side.

Upstream effects rarely get counted because they happen before the facility exists. Semiconductor fabrication requires ultra-pure water, aggressive chemical processing, and sustained high-energy inputs. The environmental cost of a chip is incurred at the foundry, often in a different jurisdiction with different reporting requirements, long before it reaches a Canadian data centre. Drawing the boundary of analysis at the facility gate understates the total burden by a considerable margin. Lifecycle accounting would tell a different story than point-of-use accounting. The industry prefers point-of-use accounting.

Transparency Is Not Optional

The transparency problem compounds everything else. Companies disclose regions and announced capacity. They rarely provide continuous data on actual water consumption, real-time grid draw, or regional strain during peak periods. Without consistent metrics, comparison is impossible and accountability is mostly ceremonial. Renewable energy contracts make the picture look cleaner on paper without necessarily changing physical grid reality; a company can be technically running on renewables while drawing from a fossil-heavy grid during peak hours, with the accounting reconciled later through certificates that describe a financial transaction, not an electron. The gap between the certificate and the grid is where the credibility problem lives.

Regulatory frameworks haven’t caught up. Environmental assessment processes built around single-project impacts struggle with facilities that cluster, that expand incrementally, and that draw from shared regional infrastructure in ways that only become visible in aggregate. Grid interconnection rules that evaluate each facility on its own merits miss the cumulative load picture entirely. Water use reporting tied to permitting conditions would help; so would grid impact assessments that account for what’s already drawing from the same nodes. None of this is technically complicated. It requires political will to impose reporting obligations on an industry that has benefited from their absence.

Conclusion

Sustainable AI isn’t an engineering problem. It’s a choice problem; about what gets built, what gets deployed, and whether anyone in the chain is actually doing the math on what it costs. Right now, nobody is. Regulators approve facilities in isolation while cumulative load climbs. Procurement officers sign cloud contracts without asking a single question about water consumption. Developers ship into continuous inference without modelling the baseline. Every one of those decisions is a vote for the status quo, cast quietly, without a recorded division.

The industry will not fix this on its own. Transparency doesn’t emerge from goodwill; it gets imposed, or it doesn’t happen. Reporting obligations tied to real consumption data, permitting conditions that reflect actual grid and water impact; these aren’t radical asks. They’re the minimum conditions for treating shared public resources as exactly that; shared, and finite.

We are at the point where the next round of expansion decisions will either entrench this pattern or begin to correct it. That window doesn’t stay open indefinitely. The conversation the industry would rather defer is the one that needs to happen now; before the constraint arrives externally and the choices get made for us.

Marc-Roger Gagné MAPP

@ottlegalrebels

See more breaking stories here.