At first glance, Canada has committed substantial resources to AI. The Pan-Canadian AI Strategy has directed more than half-a-billion dollars into the ecosystem since 2017. The $2 billion Canadian Sovereign AI Compute Strategy followed in 2024. Budget 2025 added nearly $926 million over five years for sovereign compute infrastructure and $25 million over six years for Statistics Canada’s TechStat program to measure AI adoption. Organizations such as Scale AI fund industry-led integration projects. By any measure, this is serious public investment.
The reality, however, is different. Many Canadian AI founders continue to incorporate in Delaware, and venture capital continues to flow south. In January 2026, Y Combinator announced it would no longer invest in Canadian-incorporated start-ups, requiring founders to reincorporate in the United States, Singapore or the Cayman Islands. The accelerator quickly backtracked, but the fact remains that the Canadian companies that went through Y Combinator and became unicorns had all nearly already converted into Delaware corporations. The talent pipeline tells a similar story. A study of STEM (science, technology, engineering and math) graduates from the University of Toronto, the University of Waterloo and the University of British Columbia found that one-in-four opted to work outside Canada, with the rate for software engineering graduates reaching 66 percent. The question is not whether Canada is spending enough on AI, but whether the architecture of that spending reflects the reality of building an AI company in this country.
Many founders have identified a structural gap at the centre of Canada’s approach: policy invests heavily at the research end of the pipeline and announces large figures at the infrastructure end. But the missing middle — where prototypes become products, where start-ups find their first customers and iteration speed determines survival — remains largely neglected.
Without going further, we should look at how research is set up. The conventional wisdom for a middle-power country is to select a few niche verticals and concentrate resources there, but the reality is that most entrepreneurs reject this logic. Innovation frequently emerges from non-obvious and general-purpose applications rather than from predetermined silos. Consider that two of Canada’s most valuable AI companies, Cohere and Ada, both emerged from Toronto’s foundational AI research community rather than from any government-designated vertical. Cohere builds general-purpose language models for enterprise use; Ada automates customer interactions across industries. Neither would have been predicted by a vertical strategy, yet both are now globally competitive. A country that bets on a specific vertical research risks missing technological breakthroughs and finding itself locked into a domain with limited transferability.
On the contrary, the more productive approach is to fund foundational AI capabilities, such as open-source models, shared compute infrastructure and talent development, that enable a broad range of applications. This is what American and Chinese universities do, and it is what Canadian institutions such as Mila — Quebec Artificial Intelligence Institute are positioned to do, but they lack the resources to compete at the foundational level.
This does not mean, however, that research investment should be directionless. Canada can shape priorities according to acute national needs without narrowing into silos. Health care is a case in point. Canada’s distributed population and strained public health system create demand for AI that works in environments most developers never consider: remote communities, northern reserves and clinics operating with intermittent connectivity. Developing offline large language models that function securely in these settings would address a genuine domestic crisis while also producing technology exportable to any nation with similar infrastructure constraints. The problem of building AI for a developed country but with developing infrastructure is not uniquely Canadian, but Canada is unusually well positioned to solve it.
Diverging Paths
All that said, infrastructure is where the founder perspective diverges most sharply from current policy. The $2 billion sovereign compute strategy is acknowledged as a meaningful start, but compute capacity and deployment capability are not the same thing. What Canadian AI companies lack is not awareness that the government is investing in infrastructure; what they lack is a sovereign and enterprise-ready cloud environment where they can actually host, train and deploy models under Canadian jurisdiction.
This concern has had real implications for several years. Under US legislation, notably the CLOUD (Clarifying Lawful Overseas Use of Data) Act of 2018, the American government can compel service providers to disclose data regardless of where that data is physically stored. For a Canadian start-up handling, for example, protected health information, government contracts or corporate trade secrets, building on American cloud infrastructure is an accepted but uncomfortable risk. It also means that, from a venture capital perspective, the data sovereignty of a Canadian company is only as strong as the jurisdiction of its cloud provider, which is almost always American. The absence of a credible Canadian alternative is not merely an inconvenience but rather a structural incentive to incorporate elsewhere. The European Union — notably, France — is taking steps to address its over-dependence on American cloud.
The concern about overreliance is not without precedent, even in recent Canadian policy thinking. In his 2019 Jackson Hole address, Canadian Prime Minister Mark Carney, then governor of the Bank of England, argued that structural dependence on the US dollar had created vulnerabilities in the international monetary system that no amount of sound domestic policy could overcome. Emerging economies had done everything the conventional wisdom prescribed: inflation targeting, fiscal discipline, macroprudential regulation. And, yet, it was not enough, because the system itself routed risk through a single jurisdiction. “Keeping one’s house in order is no longer sufficient,” PM Carney concluded. “The neighbourhood too must change.”
The same logic applies to AI infrastructure. Canada can fund research, train talent and announce sovereign compute strategies, but as long as the deployment layer — the cloud environments where models are actually hosted, trained and served — remains under US jurisdiction, domestic investment will remain structurally undermined. As economist Rudi Dornbusch warned, in a line Carney himself quoted, “In economics, things take longer to happen than you think they will, and then they happen faster than you thought they could.” In other words, the window is not unlimited for building sovereign AI infrastructure before dependency becomes irreversible.
Canada has natural advantages that could make sovereign infrastructure economically competitive rather than merely politically desirable. Inexpensive hydroelectric and nuclear power, combined with a northern climate that reduces cooling costs, means that running AI workloads in Canada could genuinely be cheaper than running them in Virginia or Oregon. If Canada became known as the lowest-cost jurisdiction in which to train and deploy models, the investment case for domestic infrastructure would make itself. But realizing this advantage requires deliberate policy: not just building data centres but also properly funding institutions such as Mila or the Vector Institute to produce an open-source Canadian foundation model, trained on domestic compute, governed under Canadian privacy law and available as a secure baseline for commercial development.
The most revealing frustration, however, concerns neither research nor infrastructure, but simply the government’s own reluctance to become a customer.
Canada’s public sector may be the single largest untapped market for AI adoption in the country: examples include municipal permitting processes that take more than a year; tax remittance systems whose complexity requires thousands of additional staff just to administer; and Arctic surveillance across the world’s second-largest landmass, conducted with tools designed for smaller and denser nations. These are all operational realities where AI could deliver measurable efficiency gains, and where government procurement would simultaneously create the domestic market that Canadian start-ups desperately need.
Many barriers are well understood. Bureaucracy is risk-averse by design and often lacks clarity in what AI can practically deliver. Procurement programs such as Innovative Solutions Canada are hampered by cancelled budgets and procedural complexity incompatible with the iteration speed start-ups require. Founders are not seeking massive contracts, but they need sandboxed environments within government departments where they can demonstrate capability, receive feedback and iterate. These are the basic conditions under which software companies innovate and learn what works.
Two policy proposals that emerged among AI builders deserve particular attention for their ambition. The first is the creation of Special Economic Zones oriented toward AI development; these are geographically targeted regions, with zero capital gains tax on long-term technology investments. The purpose is direct: to offer venture capital an incentive structure competitive with Delaware, so that incorporating in Canada is not a concession but a strategic choice. This strategy is not without precedent. Shenzhen, established as China’s first Special Economic Zone in 1980, used tax exemptions and regulatory flexibility to transform from a fishing village into a global technology hub now home to more than 2,600 AI enterprises. Dubai Silicon Oasis, launched in 2004, offers zero corporate tax for 50 years and full foreign ownership, and has attracted thousands of technology companies to the United Arab Emirates. These zones demonstrate that targeted fiscal policy can redirect capital flows and anchor innovation ecosystems where they might not otherwise take root.
The second is more unconventional: a proposal that a fixed percentage of every federal department’s budget be mandated for investment in Canadian AI solutions, and that this expenditure could count toward Canada’s North Atlantic Treaty Organization defence-spending commitment. The logic is that AI competence is a modern defence necessity, and dual-use technology investment is increasingly how allied nations think about meeting their security obligations. Regardless of whether one agrees with the specific mechanism, the underlying argument is serious: Canada’s AI spending and its defence spending are addressing overlapping strategic imperatives, and policy should reflect that convergence rather than treating them as separate budget lines.
These proposals are not exhaustive. They emerged from a single afternoon’s discussion among people who build AI for a living and encounter the friction points of Canadian policy daily. The value of sharing ideas was not in producing a definitive strategy but in demonstrating what becomes visible when policy formation includes the perspectives of those who must implement it.
Canada’s AI investment is not trivial, and its research talent is not lacking. What is lacking is the connective tissue between public investment and commercial reality: the infrastructure; the procurement pathways; the capital incentives; and the institutional willingness to treat AI founders as participants in strategy rather than as downstream beneficiaries of it. Until that changes, the money will continue to flow in at the top, but the companies will continue to leave from the middle.
