Rows of NVIDIA GPU servers are operating behind locked cage doors in a data center building in Vancouver’s Mount Pleasant neighborhood, which was once mostly known for independent coffee shops and craft breweries before tech companies began to move in. Cooling systems are humming at a frequency you can feel more than hear. The servers are performing the same tasks as servers in San Jose, Dublin, and Singapore: processing data, training models, and executing AI workloads. Where the data is stored makes a difference. It remains in Canada. That’s the whole idea.
Since the mid-2020s, Canada has been under pressure to develop its own AI infrastructure due to a number of factors, including geopolitical concerns, regulatory pressure, and the obvious fact that a nation whose researchers produce a disproportionate amount of the world’s AI science lacks nearly all of the compute infrastructure needed to actually run that science at scale.
For many years, the Vector Institute in Toronto, Amii in Edmonton, Yoshua Bengio at Université de Montréal, and Geoffrey Hinton at the University of Toronto have all contributed fundamental research to deep learning. However, the majority of the compute required to carry out that study has been purchased on American cloud platforms, meaning that Canadian data, Canadian models, and occasionally Canadian government material have been handled on servers owned by US businesses that are subject to US law. That’s an increasingly awkward arrangement for regulated businesses like healthcare and financial services.
In 2024, the federal government pledged $2 billion for a Canadian Sovereign AI Compute Strategy, with the specific objective of constructing sufficient domestic GPU capacity to significantly alter that ratio. The most notable private player has been Vancouver-based TELUS, which has pledged to host Canadian enterprise AI workloads solely on Canadian infrastructure and announced a GPU cluster partnership with NVIDIA for data center facilities that place servers in British Columbia rather than Virginia or Oregon. Your AI model never processes data on foreign-owned hardware, your data never crosses borders, and your compliance team no longer needs to justify why Canadian customer data is stored in an American data center that is subject to the US Cloud Act. This is the simple pitch to enterprise clients.
BC Hydro is Vancouver’s unique edge in this situation. With around 98% of its electricity coming from hydroelectric sources, the provincial utility is by far one of the cleanest power systems in North America. Corporate sustainability reports are increasingly showing how electricity-intensive data centers with AI training workloads are; for example, training a large language model can use as much power as hundreds of transatlantic flights. The multinational corporations and regulated industries that make up Vancouver’s main enterprise clients find value in the capacity to run their workloads on power that is almost 100% renewable and to honestly state so in a sustainability report. When other factors are fairly balanced, it tips the scales, but it’s not the main reason someone installs data center infrastructure in a specific city.
Former Google Brain researchers formed Cohere, a Canadian AI model startup that operates partially from Toronto but has strong ties to Vancouver. The company has made it clear that it is designing its enterprise infrastructure with Canadian data residency in mind. Financial organizations, insurance corporations, and government agencies who have certain restrictions regarding where their data is processed can use the company’s massive language models. The Canadian government’s investment in domestic compute is specifically intended to promote Cohere’s positioning as a regulated enterprise model provider with built-in sovereignty guaranties rather than a general-purpose consumer AI product. In ways that appear less like coincidence and more like the early phases of a purposeful industrial policy really succeeding, the two tactics support one another.
The discrepancy between the current state of Canadian compute capacity and the rhetoric of Canadian AI sovereignty is still something that needs to be acknowledged. The federal investment is intended to reduce the gap between Canada’s portion of global AI research output and its share of global AI compute, which is between 1 and 2 percent. It takes years, a large amount of money, and competition from US hyperscalers who have been investing in this infrastructure for longer and on a larger scale to build data center infrastructure at the scale needed to execute frontier AI training runs. Vancouver can set itself apart in terms of renewable energy, regulatory stability, and accessibility to a highly skilled workforce in technology. It finds it difficult to compete with AWS, Azure, or Google Cloud on a raw computational scale.
The ultimate goal of the sovereign AI build-out is to provide a workable substitute for consumers who require one; it is not a replacement for everything, but rather a reliable domestic choice for government organizations and regulated industries with particular justifications for wanting their AI to run on Canadian hardware. Observing the growth of that industry from Vancouver’s tech district, it’s feasible that in five years, when the infrastructure is better developed and the regulatory framework surrounding AI data stewardship has tightened even further, what appears to be incremental now may become more substantial. Alternatively, the US hyperscalers might modify their own products fast enough to meet the same compliance standards, making the sovereignty issue less significant. That result is still up for debate.
