Latest in the Canadian AI Sovereignty Series

Canada’s data-centre policy was designed around an intelligence-scarcity model that ceased to hold on July 27.

There is significant public opposition to data centre development in Canada and a lot of industry support. But do we even know how much is required? As the market matures, how much frontier-tier reasoning do we really need? Is the country at the point where we are exceeding demand? What’s the practical data-driven argument for and against? Canada’s National AI Strategy says commercial players may require 5.5 GW by 2030. But with open weights, those numbers have changed.

Every AI prompt sent in Canada in a day, by every citizen, every business and every public servant, draws about 16 megawatts on average and fits inside roughly 51 megawatts of provisioned capacity. Canada already operates about 337 megawatts of AI data-centre capacity. The country has more than 20 gigawatts under planning or development. The gap between what Canadians consume and what is being proposed is a factor of roughly four hundred, and no published demand model accounts for it.

Roughly seventy data-centre proposals have been announced in Canada since 2024, and a handful have broken ground. Do we need them?

Repricing the asset

At the start of the Iran conflict I argued that Gulf data centres had become strategically critical and strategically exposed, and that the concentration of frontier compute in a small number of jurisdictions created leverage for whoever held it. The concentration risk was real and remains real. What has changed is the price of the thing being defended.

Compute is still a “binding constraint” when models are built. Anthropic tightened peak-hour limits for paying customers in March 2026 on capacity grounds, then contracted in May to rent capacity from xAI’s Colossus cluster at $1.25-billion a month through 2029. When a frontier lab buys infrastructure from a direct competitor to keep its consumer tiers online, it settles much of the question at the training layer.

Serving a national population its daily prompts sits at a different layer, and that layer commoditized. Moonshot’s Kimi K3 reached rough parity with Fable 5 and GPT-5.6 and shipped its weights publicly. Google reported a thirty-three-fold reduction in the energy of a median Gemini text prompt over twelve months. The Gulf states bought concentration risk in exchange for leverage that open weights have partially dissolved, which actually makes them more exposed today than when I wrote about them, just differently.

What the country actually consumes

The calculation is available to anyone willing to do it. Google published a median of 0.24 watt-hours for a Gemini text prompt in 2025. Epoch AI puts a standard query near 0.3 watt-hours, a query carrying ten thousand input tokens near 2.5 watt-hours, and reasoning-model responses in the range of 18 to 40 watt-hours. Statistics Canada reported in March 2026 that 41.6 per cent of Canadian workers had used at least one AI or automation technology in their main job over the preceding year, and business adoption reached 19.2 per cent in the second quarter of 2026.

Assume half of Canadian adults are active consumer users at eight prompts a day, and that working users average twenty-five. Assume the task mix the usage data describes, with ninety per cent routine prompts, eight per cent medium-context and two per cent reasoning-intensive.

Provisioning for peak concurrency at two and a half times average, with redundancy at a further thirty per cent, gives about 51 megawatts. The federal government’s entire prompt volume, the workload that the phrase “sovereign AI data centre” is meant to protect, runs under one megawatt.

Every assumption above can be moved. Doubling consumer intensity, doubling worker intensity and doubling the reasoning share together produce a figure still comfortably inside existing capacity. Should build not be driven by projected need and demand?

Where the load concentrates

Two per cent of prompts carry fifty-five per cent of the energy. The reasoning slice is small in volume and dominant in consumption, the math behind the pattern visible in market share: the models with the largest user bases are rarely the strongest, and the strongest models serve a thin population of users and firms doing work that requires them.

Agentic workloads are the one variable that moves the national figure by an order of magnitude. An agent task consuming the equivalent of two hundred chat turns, run twenty times a day, produces 67 megawatts of provisioned demand (!) if five per cent of AI-using workers adopt it, 333 megawatts at twenty-five per cent, and about 1.3 gigawatts at full saturation of every AI-using worker in the country. Add a sovereign training capability, which at GPT-4 scale runs near 25 megawatts continuous, and the ceiling on Canadian domestic demand under maximal assumptions lands around 1.5 gigawatts.

That ceiling is seven per cent of the announced pipeline. The remaining ninety-three per cent is export capacity, and it should be argued for on those terms.

The June order and the old chokepoint

On 12 June 2026 the U.S. Commerce Department required Anthropic to restrict access to Mythos 5 and Fable 5 for any foreign national inside or outside the United States. Anthropic disabled both models for all customers within hours. Commerce lifted the restriction on 30 June. In July, reporting described a centralized programme under which companies would need Washington’s approval before admitting partners to a model launch.

A Canadian data centre serving a U.S. model through an API went dark on 12 June alongside everyone else. Rack space in Sturgeon County would have changed nothing, because the instrument was an export licence on model access.

Kevin Yin, writing in The Globe and Mail, reads the same episode as a case for building compute, and reaches for the Strait of Hormuz. The analogy locates the chokepoint in the wrong place. Oil is the scarce commodity and the strait is the constriction. In AI, compute is the commodity, purchasable in a dozen jurisdictions, and model access is the constriction. A country cannot build past a licensing regime. It can download past one.

The sovereignty objection is that downloadable frontier-tier weights are largely Chinese, and that this trades one dependency for another. This is an inaccurate characterization. A weight file is a possession rather than a service, and possession survives a letter from any commerce ministry. Provenance becomes a question of inspection and testing, which is tractable. Access is a question of another government’s jurisdiction, which is not. Yes, we still need compute and inference and power but these are different issues than frontier weights, which are suddenly abundant thanks to a deluge of Chinese innovation.

A hosting industry judged on hosting terms

Yin’s second argument survives all of this intact, noting that economists usually care where capital sits rather than who owns it, and that data centres invert this because computational services can be supplied from abroad while the durable benefit comes from learning to build, operate and improve the facilities. That is an industrial policy case for domestic participation, and it stands on its own merits.

It is not a sovereignty case. Once the two are pulled apart, the pipeline has to answer on industrial terms, and Yin has already conceded most of that: data centres raise local energy prices, generate noise, employ far fewer people than comparable capital projects, and produce profits that are difficult to tax when the intellectual property sits in another country. Angus Reid polling this year found broad support for domestic AI infrastructure alongside more than two-thirds opposing construction near their own communities.

Governance at collective scale

In an amazing initiative, Indigenous communities are assembling complete sovereign stacks, uniting globally to cover models, data governance and hardware. The First Languages AI Reality initiative, run by IndigiGenius in partnership with Mila, develops automatic speech recognition for languages that lack the hundreds of hours of transcribed audio conventional systems require, according to my collaborator Laszlo Lakatos-Hayward, engaged on a project. “They have built their own open source custom models, specifically for language so it is not english is the primary base to work from. From New Zealand, to Kenya, to S. America, Hawaii communities are creating their own and only connecting to other communities,” says Hayward, bypassing nation states, and releasing the resulting tools as open source. Māori developers took Qwen, stripped roughly ninety-seven per cent of it and rebuilt the remainder in house so that English stopped being the base the model reasons from, he added. The hardware is bought outright and networked between communities. Sovereignty here is exercised through possession of the weights, control of the training corpus and ownership of the machines, on a timeline that ran ahead of any provincial strategy and never depended on one.

Canadian First Nations developed the OCAP framework in 1998, asserting ownership, control, access and possession over their own data, and the First Nations Information Governance Centre has stewarded it since. Te Hiku Media in Aotearoa built its own content distribution platform in 2013 rather than depend on Spotify or YouTube for archival control, then moved from manual transcription to an in-house speech model in 2016 governed by its own licensing terms. Neither approach required owning a gigawatt. Control was exercised through the conditions attached to the data and the terms under which models trained on it may be used.

Some First Nations have made a different call. The Logic reported a band membership voting in July to approve a data centre on its land, expanding its tax base, one instance of communities capturing value from the construction boom directly. Both positions are exercises of self-determination, and a compute policy that recognizes only one of them is speaking over the other. The lesson available to Ottawa is that sovereignty over data has been asserted for nearly thirty years through instruments that had nothing to do with megawatts.

The moratoria want a number

There’s clear resistance to data centres that is increasingly organized and based on everything from environmental to health to economic issues. And that resistance is getting results. Mississauga council voted unanimously on 29 July to direct staff to prepare an interim control bylaw prohibiting data centre approvals for up to one year, extendable to two, with the bylaw returning for a final vote on 16 September. The application behind it is a proposed fifty to one hundred megawatt facility at Tenth Line West and Argentia Road, sitting under three hundred metres from housing. Adopted, it would be the first citywide moratorium of its kind in Canada. Hamilton council defeated a comparable motion weeks earlier. Vaughan approved a hyperscale facility that is nearly built. Toronto residents are pressing for a pause on projects in Etobicoke and Scarborough, and New York has already imposed one at the state level.

Mississauga’s motion carried a second instruction that has drawn less attention. Staff were asked to develop a protocol for evaluating applications, which requires a defensible figure for how much capacity the country needs and for what. That figure does not currently exist in any Canadian planning document, which is why the fight is being conducted entirely through noise studies, water tables and setback distances.

The opposition movement argues from local externalities. The argument here is a demand argument, and the two reach overlapping conclusions from unrelated premises. The distinction matters because the demand argument survives every externality being solved. Hourly power matching, closed-loop cooling, full property-tax capture and generous community benefit agreements would leave the arithmetic untouched: the capacity Canadians consume already exists.

A national prohibition would foreclose the capacity the sovereignty case actually requires, which is Canadian-operated, Canadian-jurisdiction compute at a scale between fifty megawatts and, under maximal agentic adoption, something near a gigawatt and a half. The conclusion available from the numbers is an allocation with a ceiling.

British Columbia already produced it

British Columbia now requires every data-centre project approved over a two-year window to compete for a fixed allocation of 400 megawatts. A province decided that compute is rationed against a defined budget rather than built to whatever the pipeline requests. That figure sits above the 51 megawatts Canadian demand currently justifies and well inside the 1.5 gigawatt ceiling that maximal agentic adoption would produce.

The sovereign requirement is jurisdictional. Capacity operated under Canadian law, by Canadian-controlled entities, running weights Canada possesses outright, satisfies it at a scale the country has already built. ThinkOn remains the only Canadian company cleared to house federal data, the only Canadian-owned cloud provider approved under Shared Services Canada’s secure-workloads framework for Protected B workload, which is a much bigger sovereignty issue than any megawatt figure and one that no new construction in Alberta addresses.

Method note: load figures derive from published per-prompt energy estimates (Google 2025, Epoch AI 2025) applied to Statistics Canada adoption data, with a 90/8/2 routine/medium/reasoning task mix, peak provisioning at 2.5x average and 1.3x redundancy. Per-prompt energy figures are provider-reported and only two firms have published them. Image and video generation are excluded. The agentic multiplier is the largest source of uncertainty and could be wrong by a factor of five in either direction.