The carbon math of cloud AI vs local inference: a Canadian perspective

It is Earth Day this week, which means the usual round of corporate sustainability statements. Most of them avoid the question that matters for any organization running AI: what is the carbon footprint of an inference call, and where does that footprint actually live.

For Canadian healthcare organizations, the answer is more interesting than the typical sustainability report admits. Most cloud-hosted AI runs in US data centers. Most US data centers run on grids that still burn coal, gas, or both. A query sent to a popular cloud LLM is, on the margin, fueled by a power mix that looks nothing like the Canadian grid.

Where the electrons come from

Quebec's grid is roughly 99 percent hydroelectric. Ontario's is about 90 percent non-emitting between nuclear, hydro, wind, and solar. By contrast, the AWS regions hosting most cloud AI workloads sit in Virginia, Oregon, and Ohio, where the grid mix in 2025 still included a meaningful share of coal and natural gas.

The carbon intensity gap is large. A kilowatt-hour consumed in Quebec produces about 2 grams of CO2-equivalent. The same kilowatt-hour in Virginia produces around 290 grams. That is a 145x difference for the exact same compute.

What this means for AI workloads

AI inference is power-hungry. A single conversation with a large language model can consume more electricity than a full day of laptop use. When that compute runs in Virginia, the carbon cost is real. When it runs on a server sitting in a Montreal data center, the cost is rounding error.

For healthcare organizations, this is rarely the headline reason to choose local infrastructure. Privacy and PHIPA compliance dominate the conversation. But once a self-hosted footprint is in place, the carbon math becomes a free upside. Every model invocation that stays in Canada is meaningfully greener than the cloud equivalent.

What to ask vendors

If you are evaluating an AI tool for your agency, three questions cut through the marketing: where does inference run, what is the grid carbon intensity at that location, and is there a Canadian-hosted option. Vendors that cannot answer the first question are running on cloud they do not own. Vendors that brush off the second are not tracking it. Vendors that say no to the third are choosing not to build for the Canadian market.

It is one more reason the self-hosted, Canadian-first architecture we built into Sentinel is not just a privacy decision. It is also the lowest-carbon way to run modern AI in this country.