Canada’s AI Ambition Needs a Working Stack

An impressive AI demonstration tells us very little about whether a northern clinic can use the system reliably. That takes a connection, usable patient information, dependable computing and a clinician who can check the result. Canada’s place in AI will be decided partly by those less glamorous things, and by who can make them work together.

Canada has earned its place in artificial intelligence research. The Pan-Canadian AI Strategy supports Amii in Edmonton, Mila in Montréal and the Vector Institute in Toronto, institutions that connect researchers, training and work with Canadian businesses. There is a substantial foundation here to build on.

The difficulty comes after the discovery. A researcher needs computing time. A promising company needs customers and enough money to grow. A hospital needs a tool that fits its records and clinical responsibilities. Somewhere along that route, a Canadian invention can become a Canadian business, an imported service or an idea whose largest returns accrue elsewhere.

That is what Nasty Canuck’s series, Canada’s AI: From Power to Payoff, will examine: the chain that turns AI capability into something Canada can use, maintain and sell. Our position is that Canada should build enough control into the systems it depends on, then use its strengths to compete in markets where customers have a reason to choose it.

Research needs a route to market

The familiar brain-drain story captures only part of the problem. Vector’s 2025–26 annual report says 91% of employed graduates from its recognized programs remained in Ontario. That is a specific group, not a census of Canadian AI talent. Still, it is a useful correction to the idea that everyone worth keeping has already left.

People can stay while much of the commercial value goes elsewhere. In its 2026 venture-capital review, BDC identifies financing from Series A onward as a persistent obstacle to scaling Canadian companies, alongside heavy reliance on foreign capital at later stages. This is a broader technology-financing problem, with obvious consequences for AI businesses trying to grow.

Foreign investment can help those businesses reach customers they could never serve alone. The Canadian interest is in what the arrangement leaves here: decision-making, skilled work, intellectual property and the ability to build the next company. A successful acquisition can reward founders and recycle capital. It can also relocate the decisions that determine where future investment goes. The terms matter.

The larger AI economies have different starting points. Stanford’s 2026 AI Index records US$285.9 billion in U.S. private AI investment in 2025. It also finds that U.S. and Chinese models have been trading the performance lead, while China leads in industrial robot installations. Private investment alone understates China’s state-supported effort; it is not a complete national spending comparison.

Europe is pooling resources. In July, the EU launched a call for up to seven AI gigafactories, supported by proposed public funding and intended private investment. Those are planned facilities, not seven new operating supercomputers.

Copying American spending would be a poor Canadian strategy. Canadian model developers should pursue ambitious products where they see a commercial opening. The national plan also needs to work for a manufacturer buying software, a clinic with unreliable connectivity and a startup that cannot afford its next training run.

The stack starts outside the screen

An AI model is software trained to recognize patterns or generate outputs. “Compute” is the processing capacity needed to train and run it. Neither does much for a user without the systems around it.

For this series, the stack includes:

Layer	What has to work
Power	Reliable electricity, transmission, connections and cooling
Compute and hardware	Processors, servers, data centres and cloud services
Networks	Fibre, mobile connections, internet exchanges, satellite and backup routes
Data and models	Usable information, lawful access and software suited to the task
People and companies	Research, engineering, operating skills, capital and customers
Deployment and accountability	Workflows, security, oversight and evidence that the service helps

These layers depend on one another. More processors will not repair incompatible medical records. Excellent software will not rescue a connection that keeps dropping. Cheap electricity is less useful if a project cannot connect to the grid where it wants to build.

Ottawa’s AI for All strategy acknowledges many of these weaknesses, including nascent sovereign cloud capacity, dependence on overseas GPU fabrication and constrained electricity infrastructure. The useful response is to test delivery against those admissions.

Canada is building capacity at several speeds

There is operating infrastructure to point to. TELUS says its Rimouski AI factory opened in September 2025 and had sold out by its May 2026 expansion announcement. The independent TOP500 listing records the Rimouski system and its HPE and NVIDIA equipment. The machine is real. TELUS’s proposed B.C. cluster is a different proposition: more than 60,000 GPUs and 150 megawatts targeted by 2032, with facilities arriving in stages. Those figures describe planned scale. They cannot be added to an inventory of capacity available today.

Bell offers an equally useful distinction. Its September agreement with Saskatchewan proposes up to 900 megawatts of additional capacity, creating a path to a 1.2-gigawatt hub. The memorandum is non-binding. Development depends on customers, commercial agreements, power generation, permits and approvals.

This is a serious pipeline of possible investment. Calling it completed infrastructure would make it less useful to understand. An announcement, a signed contract, construction and a service accepting customers are different milestones. We will keep them separate.

Power deserves the same care. Canada’s hydro and nuclear resources give it a favourable starting point. Provincial grids, however, have their own supply mixes, connection limits and competing demands. A national advantage has to become an available connection at a particular site.

Ontario’s system operator forecasts electricity demand growth of 65% by 2050. In its reference scenario, data centres account for 8.6% of demand that year. These are planning forecasts for all data-centre demand, not a measurement of AI’s current power use.

Who pays for the added infrastructure is already a live question. The federal Responsible Data Centre Development Principles say proponents should cover project-driven electricity-system costs. The principles establish expectations; they do not prove that every project’s financing and utility arrangements meet them. Households have a reasonable interest in the difference.

Sovereignty has to survive a supplier problem

A data centre’s postal code tells us where the building is. It leaves a great deal unanswered about the service inside. Who owns and operates it? Which laws and contracts govern access? Who holds the encryption keys, and who can authorize their use? Can the customer move its data and workloads elsewhere? Can Canadian staff keep the service running if a supplier withdraws support?

The answer can vary by layer. Bell’s June collaboration with Cohere, Hypertec and BUZZ HPC combines Canadian infrastructure, enterprise AI software and hardware manufactured here, while using NVIDIA’s platform. That can build useful domestic capability. It also retains dependencies that the word “sovereign” should describe honestly.

Canada should judge sovereignty by the control a particular service needs and the consequences of losing it. A critical public service may justify paying for domestic operations, stronger access rights and a tested fallback. A less sensitive commercial task may be served well by a foreign platform. Canadian companies can earn export revenue using international computing infrastructure.

There is no virtue in making a capable Canadian firm buy a worse tool merely to satisfy a slogan. There is considerable value in knowing how an essential service could continue when its preferred tool becomes unavailable. Switching may be expensive, and some hardware dependencies will take years to diversify. A credible plan includes those costs.

That leaves room for Canadian ownership where it changes consequential decisions, and for trusted foreign suppliers where partnership makes sense. The same test should expose an expensive domestic monopoly offering little practical control in return.

Health care tests the whole chain

The human case for getting this right is especially clear in health care. The 2025 OurCare survey estimates that 5.9 million adults in Canada lack regular access to a family doctor, nurse practitioner or primary care team, an improvement from 6.5 million in 2022. The remaining gap is large.

An early opportunity is to give clinicians time back. The CMA’s account of a McMaster Health Forum evidence synthesis describes modest, consistent evidence that AI scribes reduce administrative burden. It also identifies integration problems, training needs, bias and accuracy risks. A tool that drafts a clinical note still needs appropriate checking.

Less paperwork could support more appointments or reduce the after-hours work that makes practice harder to sustain. Whether saved time becomes additional patient access needs to be measured. A documentation result cannot, by itself, establish that the doctor shortage has shrunk.

Rural care exposes another dependency. Remote specialist support needs reliable communications, and some services need local processing or a way to function when the connection fails. Telesat’s planned Lightspeed landing stations in Quebec and Saskatchewan show how a satellite network has to connect to ground-based fibre and internet exchanges. The announcement describes infrastructure under development, rather than a service already available to every northern clinic.

The data layer is just as consequential. Ottawa’s June announcement of $100 million for the VITAL health-data platform addresses information fragmented across hospitals and systems. VITAL is building access to de-identified hospital data for research and innovation. It should not be confused with a finished national patient-record system.

Before expanding a tool, a health system needs to establish whether it is validated for the patients using it and fits clinical work. A qualified person must remain accountable when the software is wrong.

Adoption is where the investment earns its keep

Statistics Canada’s second-quarter 2026 survey found that 19.2% of businesses reported using AI to produce goods or deliver services in the preceding 12 months, up from 6.1% in the equivalent 2024 survey. Use ranged from 42.3% in information and cultural industries to 4.5% in agriculture, forestry, fishing and hunting.

The third-quarter planning survey, released October 1, found that 25.2% intended to use AI over the following year. Intent and completed adoption are separate measures. Both suggest momentum; neither tells us how much productivity has improved.

Industrial expertise gives Canada places to test useful products: mine maintenance, crop management, energy operations, manufacturing and transport. These are candidate markets, not an automatic list of Canadian winners. A company still has to solve a costly problem well enough that someone pays for the solution.

That is also how we should assess public procurement. Shared Services Canada says it has begun deploying the GC AI Platform with selected departments and agencies, with expansion planned through 2026 and 2027. Government is becoming an actual deployment site. The next questions concern service quality, costs and whether suppliers gain commercial references they can use beyond that first customer.

Procurement can help a young business prove itself. It can also trap one in a bespoke system built for a single department. Competitive contracts, workable integration and a product that other customers want would make a stronger foundation for growth.

Trust is part of the machinery

Canada’s privacy authorities already set out principles for lawful, accountable uses of generative AI. Their guidance places responsibility for decisions with the organization using the system. A confident answer on a screen cannot accept liability on its behalf.

Security has two sides. The Canadian Centre for Cyber Security assesses that AI is making malicious activity easier to scale and more persuasive. It also identifies AI-focused organizations as increasingly prominent targets. Giving software access to records or authority to carry out tasks makes the permissions and monitoring around it consequential.

Data sovereignty also has to respect the people whose information makes the system useful. The First Nations Information Governance Centre describes First Nations’ authority over the collection, use and sharing of their community data. Calling a dataset Canadian does not settle who is entitled to decide how it is used.

Workers deserve an evidence-based account too. Statistics Canada’s early employment analysis, covering trends through December 2025, found that employment generally grew across AI-exposure categories, with weaker outcomes for some younger workers. It could not isolate AI’s effects from other economic forces. That leaves real questions about entry-level opportunities, training and who receives the gains from faster work.

Performance in French, accessibility and the treatment of people poorly represented in training data belong in deployment decisions. These are tests for whether a service works for its users. A national strategy that counts installations while ignoring those users would miss its own purpose.

Keep a Canadian scoreboard

The case for a Canadian AI strategy should be judged against results that readers can follow. How much computing capacity is actually available, at what price, and to whom? Which companies are growing revenue and finding customers abroad? Where are essential services gaining a usable fallback? Are workers and patients seeing improvements they can recognize?

Start with critical public services and industries where Canada already knows the work. Give those users reliable connections, usable data and access to computing, and buy against measurable outcomes. Build larger infrastructure against credible demand, while paying for the redundancy an essential service needs. Help Canadian suppliers turn successful deployments into products they can sell to the next customer.

Delivery deserves scrutiny because delays have already mattered. ISED’s evaluation of the Pan-Canadian AI Strategy found one-to-two-year delays across key activities in the PAICE research-compute project. That is a finding about a particular rollout, not every Canadian compute initiative. It shows why promised access and usable access need different entries on the scoreboard.

This series will return to three questions: what Canada should control, where it can compete, and where AI can improve Canadians’ lives. The answers will vary by industry and by the risk a service carries. They should change when the evidence changes.

Winning would leave Canada with firms that sell useful technology, public services that work better and enough operating knowledge to keep critical systems running. That requires us to be demanding customers of Canadian ambition.

A clinic with more time for patients is a better result than another Canadian AI announcement. Build the stack that gets us there.

Editor’s note: This opens Canada’s AI: From Power to Payoff, Nasty Canuck’s series on AI capability, control and practical benefits. The next article will examine the electricity system behind the build.