When Prime Minister Mark Carney announced AI for All, the new federal artificial intelligence strategy, he was specific about how the country could secure its place in the defining technology of the era.
“Just as in our Defence Industrial Strategy, we will follow a Build-Partner-Buy Framework,” he told an audience at Toronto’s University Health Network. That means building in Canada first, partnering with allies where we cannot build and “only after exhausting these options, will we buy from abroad.”
The ambition is measurable: $200 billion in new growth, 250,000 jobs and a pledge to lift business AI adoption to 60 per cent by 2034 from about 19 per cent at launch of the strategy.
However, while borrowing the build-partner-buy sequence from the defence industrial strategy is a good instinct, it works only for some layers of AI.
Canada doesn’t need AI independence. It needs exit plans for critical sectors
AI is not one thing. It is a complex stack. In a simple sense, silicon chips sit at the bottom, compute and cloud infrastructure in the middle, frontier models above that, and the applications and agents that people use on top.
Build-partner-buy is the right order for the middle of that stack, the infrastructure layer, where Canada can play a meaningful role in the buildout of data centres. But for the top and bottom – the models and applications most Canadians will use and the chips Canada has no realistic prospect of producing at scale in this cycle – the order should flip to buy-partner-build.
Buy first because deploying and adopting the best available AI quickly often matters more than where it was made. Partner second to reduce dependencies, particularly beyond U.S. and Chinese supply chains. Build third where Canadian firms can create a defensible commercial moat or where government is handling particularly sensitive data, operating critical infrastructure or complying with national-security obligations.
The right idea, the wrong layers
The build-first approach works for infrastructure because data centres are the frigates of AI: capital-intensive, physical and slow to replace. Amazon, Microsoft and Google are estimated to hold 85 per cent of Canada’s public cloud market, well above their global average of 66 per cent, according to the Canadian Anti-Monopoly Project, a think tank.
Canada has the energy, land and expertise to build more data-centre capacity at home, which gives our country more leverage to ensure that data centres are built on terms that serve Canadians. This is likely compatible with continued partnership with global firms for advanced AI infrastructure.
However, when applied to AI models, building our own as the default could be self-defeating. The frontier of AI is moving faster than any procurement cycle can chase. AI models are also increasingly open-weight models that can be customized to Canadian contexts. We don’t need to reinvent the wheel to adopt AI.
Stanford University’s 2026 AI index finds the best closed models, such as ChatGPT and Claude, remain American and they still beat the best open-weight models, many of which are built by Chinese firms. However, the gap in capabilities has shrunk and open models now compete on both cost and speed.
Either way, most of the world’s most powerful models come from American and Chinese firms. Directing procurement toward Canadian firms first, as the prime minister put it, could slow the AI adoption on which the strategy depends. Companies and organizations should ultimately pick the tools that work for them and that enable them to most effectively grow and scale.
At the same time, there is a risk for Canada in failing to raise a national AI champion. We have one in Cohere, a Toronto-based enterprise AI company, and it would be wise for Ottawa and leading Canadian companies to continue backing the firm. Cohere has demonstrated strong revenue growth and capabilities, and is a credible option for Canadian enterprises.
Meanwhile, the Canadian debate has recently fallen into a binary trap: lean on American hyperscalers and stay dangerously exposed or chase a domestic answer to OpenAI and then later discover that sovereignty needs ecosystems, not isolation. Neither end of the spectrum fits Canada. We need a middle path.
In Sovereign by Design, a report I co-authored with Sean Mullin at the University of Toronto’s Munk School of Global Affairs and Public Policy, we argue that sovereignty in the AI era means freedom from coercion, not technological self-sufficiency. It means having options if key infrastructure or tools are ever turned off, not building and owning every layer of the stack.
How to buy, partner and build AI in Canada
There is momentum in Canadian AI adoption, particularly for businesses. Statistics Canada found 19.2 per cent of businesses were using AI by mid-2026, up from 12.2 per cent a year earlier.
The job now is to keep that curve bending upward as adoption goes past the early movers to the harder-to-reach majority. Reaching 60-per-cent adoption across the Canadian economy by 2034 will require sustained effort.
First, businesses and organizations can start by buying at the model and application layers. Closing the adoption gap is primarily a procurement and training problem.
The federal AI strategy’s promise to make government a strategic anchor customer should mean exactly that: buying the best available tools from Canadian firms where they win on merit, fast and at scale, across the public service, health systems and the many businesses that will never train a model of their own. Speed of adoption, not country of origin, is the measure that matters.
Second, partner where Canada needs scale that it cannot supply alone. The new sovereign technology alliance with Germany is the partner pillar in practice. Compute is one of the clearest opportunities for collaboration because no single Canadian effort will match the capital the hyperscalers deploy, while pooled allied demand can secure capacity on terms Canada helps set.
AI safety evaluation is another example of this. The Canadian Artificial Intelligence Safety Institute will gain far more by working alongside an allied network than working alone.
Finally, we can build where Canada can own a defensible slice. Some things are too important to not own. We will likely benefit from a sovereign cloud for the most sensitive tiers of public data, where protection from foreign coercion is the entire point.
The promised public AI supercomputer can help give startups and researchers access to affordable compute if it ultimately provides below-market rates for leading chips from Nvidia. This is similar to what the U.K. has done with its Isambard-AI computing cluster.
In building Canadian AI solutions, we may also be able to focus on regulated sectors where Canadian data and intellectual property are highly valuable.
For example, we have strong banks and are working on implementing open banking as a country, which could unlock valuable, anonymized financial data.
Further, we have a massive, publicly funded health-care system with a diverse population. If managed appropriately, those datasets could help us to uncover new efficiencies and opportunities to improve health-care outcomes for people across Canada.
Four mechanisms to make it real
First, classify data by sensitivity. Our federal data-classification framework is overdue for updating and often overclassifies data, which requires more management and bureaucracy.
Whether Canada buys a commercial AI model or insists on more sovereign requirements could depend on what type of data is involved. For example, a chatbot drafting blog posts is not as sensitive as the Canada Revenue Agency’s taxpayer records. General AI use does not need to be micromanaged by the government. Focus could instead be shifted to high sensitivity and important areas.
With Bill C-36 tabled to overhaul private-sector privacy law and with the looming 2026 review of the Privacy Act proposing “designated official sources” of key government data, this is the natural moment to modernize the federal data-classification scheme. That can include clear thresholds that govern cloud procurement.
Second, create a standing buy-partner-build review for major AI investments. The defence industrial strategy includes a Defence Investment Agency to weigh these decisions under one roof. Shared Services Canada could embed this test in its IT investment decisions.
Third, the prime minister’s instruction to “direct procurements for sovereign capabilities towards Canadian firms first” works only if we clearly define what counts as sovereign capability in the context of AI and digital.
A concise framework could pressure test significant federal AI investments: Can we buy this? Should we partner on it? Or must we build it? Publishing the reasoning behind major decisions might also give “sovereign capability” real meaning.
Fourth, deepen the partner pillar into an allied middle-power compact on AI. This is where Canadian diplomacy can punch above its weight. We can develop a standing arrangement among the countries that Canada already trusts on intelligence and defence to pool compute, align safety standards and co-ordinate procurement (a buyer’s club) where possible. This can help give each member buying power and bargaining leverage.
The unglamorous path to AI leadership
“AI for All” is the strategy Canada has been missing and, to its credit, there’s a federal procurement doctrine to go with it. But that doctrine is borrowed from the defence strategy and only partly matches the nuance needed for AI adoption across the Canadian economy and our governments.
Yes, we can build sovereign infrastructure where it makes sense. But for the tools most Canadians and Canadian businesses will use, first buy the best on the market, partner for scale, then build where Canada can win. We can’t do everything. We have to make strategic bets. That is how a middle power turns limited resources into genuine leverage.
As a final food for thought: where should we be building first? Quantum.
Quantum sensing and computing may be where Canada holds some of the strongest latent capabilities and could feasibly commercialize global champions.
In considering our future there, we could focus on building deeply and strategically, not broadly and thinly. The sprinkler-system approach to innovation support has not served Canada well in the past. It’s time to turn off the sprinklers.
