AI in Canada and Why Compute Matters

Canada is recognized internationally for its contribution to Artificial Intelligence (AI), as it was among the early jurisdictions to introduce legislation, and to establish a national policy and regulatory framework for AI, with efforts now underway to develop internationally-aligned industry standards for AI. Canada was also the first to launch a national plan for AI development, the Pan-Canadian AI Strategy (PCAIS), led by the Canadian Institute for Advanced Research (CIFAR). These and other initiatives have contributed to Canada’s position as a leader in AI research, talent, and start-up growth. 

The country is home to trailblazers in the AI field, notably Geoffrey Hinton and Yoshua Bengio, who worked extensively on neural network algorithms,3 and Richard Sutton in the field of reinforcement learning. These names are among the most high-profile in a robust ecosystem of researchers, which leads in AI research publications per capita among G7 nations as of May 2023.4  Canada’s domestic AI workforce is estimated to be between 200,000 and 220,000, composed of data science, software engineering, and development professionals, or one percent of Canada’s workforce in 2023 (compared to just 0.5 percent of the workforce in the US in 2019).5 6 Other studies also suggests Canada’s leading position in AI talent. 7 8

Canada also boasts a vibrant AI start-up ecosystem, including over 1,500 AI companies, with over 150 firms raising a total of $2.5 billion in venture in 2023.9 10 Canada ranks fourth among the top ten nations in venture capital investment, total funding raised, Generative AI (GAI) companies founded, and AI patent filing growth per capita as of mid-2023.11 

Canada's AI Strategy is funded by the federal government’s Innovation, Science and Economic Development (ISED) department under the Pan-Canadian Artificial Intelligence Strategy (PCAIS, or the Strategy). Established in 2017, the Strategy has allocated over $573 million to support AI organizations, focusing on commercialization, standards, talent, and research until 2031. CIFAR, supported by ISED, administers PCAIS with $160 million since 2017, extended with an additional $208 million in 2022. CIFAR supports AI research institutes like Alberta Machine Intelligence Institute (AMII), Montreal Institute for Learning Algorithms (Mila), and Vector Institute, along with Canada's Global Innovation Clusters (CGIC), namely Scale AI, and the Standards Council of Canada in developing AI adoption standards.

As Canadian researchers and firms are faced with a race to develop and adopt AI in a rapidly advancing field—and policymakers seek to strike the right balance between enabling conditions and guardrails against harm—one critical issue has received relatively little attention in this country: Canada’s “AI Compute” (AIC) capacity.

Canadian AI start-ups face challenges in scaling innovation to commercialization.12 Previous research found that despite a strong start-up ecosystem, adoption of AI by general businesses is low, with fewer than 1 in 25 Canadian firms reporting AI being used in their business, among the lowest of 35 OECD countries, and the second lowest in the G7.13  A recent report by Scale AI, Canada’s AI Global Innovation Cluster, finds that Canada is at a critical juncture: despite strength in domestic research and R&D and AI start-ups, Canada is missing key ingredients to allow AI to responsibly offer benefit to the economy. Canada risks falling behind leading international peers that are translating R&D and industry adoption of AI into productivity improvements, weakening Canadian competitiveness. 

Figure 1. AI at Scale report graphic depicting Canada’s AI ecosystem trajectory14

Recent reporting suggests that AIC resources form a significant share of an AI firm’s capital cost, as high as half of many companies’ annual capital spending.15 16 Even the largest companies, despite spending tens of millions on computing infrastructure, cannot afford, let alone access, the fastest AI-specific computing infrastructure and cloud computing.17 In February 2023, the OECD highlighted the lack of targeted plans for national AI Compute capacity in Canada’s national AI strategy as a “policy blind spot [that] may jeopardize domestic economic goals.”18 While policymakers at all levels have started focusing on the issue, Canada’s domestic supercomputing infrastructure lags in the number of compute units available (capacity) and performance (effectiveness) in AI training and inference.  

What is Artificial Intelligence, and Artificial Intelligence Compute?

Before we delve further, we discuss the definition of Artificial Intelligence (AI) and Artificial Intelligence Compute (AIC):

AI systems require computing power to generate inferences from their inputs, which are generally conducted by one of two types of systems: 

(a) parallel computing infrastructure such as supercomputers, including publicly-owned supercomputers used primarily for academic and scholarly research purposes; or

(b) cloud-based computing systems, typically operated by large publicly traded US-based private companies such as Microsoft (Azure), Google (Cloud), and Amazon (Web Services). To combine these multiple contexts, we rely on the OECD’s working group definition21 of Artificial Intelligence Compute (AIC):

Public funding for Canadian AI Compute 

Only $40 million of the $573 million of the Pan-Canadian AI Strategy’s funding is earmarked for AI Compute initiatives directly. The majority of AI computing infrastructure is the responsibility of the Digital Research Alliance of Canada (DRAC) and its research partners. DRAC is the successor to the high-performance computing ecosystem that began in the 1990s and evolved through research-focused initiatives like SHARCNET and Compute Canada. DRAC is now funded through PCAIS, and collaborates with the national research and education network operators, CANARIE (formerly the Canadian Network for the Advancement of Research, Industry and Education), to provide computing resources to researchers, start-ups, and educational institutions. 

Investment in Canada’s computing infrastructure has been largely limited to government and research institutions. Moreover, the current infrastructure’s lacklustre computing performance effectively inhibits the ability to handle AI workloads within domestic borders.

There are no Canadian programs that focus on the AI Compute needs of private-sector actors, especially companies seeking to scale, or to dramatically grow public AI Compute infrastructure. Canada is already behind, as we’ll explore in the next section, so this is a policy gap that needs to be addressed urgently.

Key Insights and Policy Scenarios

In the previous section, we identified the following key issues:

• AIC capacity: Canada lags behind all other G7 countries in AIC infrastructure. The available AIC infrastructure is not sufficient to support frontier-based research. Less than 10 percent of domestically situated AIC infrastructure is owned by the private sector. Investment to build domestic infrastructure is comparatively low, and the supply chain to support such projects is only in its nascent stage.
• AIC effectiveness: Canada primarily uses its research AIC capacity in biomedical science, climate science, and aerospace engineering. Private companies have to rely on cloud-based solutions to access AIC outside of Canada, and price-related pressures decrease their use. 
• AIC resilience: Canadian public policy that shapes the ability to build a resilient AIC relies on broad-based programs that encourage the adoption of technologies in general, and no tailored policies exist that specifically target AIC access. This is in contrast to other jurisdictions with specifically-targeted policies.

Public policy objectives and trade-offs

We presume as a public-policy objective that Canada needs to expand access to AIC to close the AI Compute gap, retain AI firms and talent, and promote greater AI development and adoption. This objective can promote a larger public-policy goal of greater national productivity and competitiveness. Like the Internet before it, AI capacity could power a wide variety of areas of economic growth that require immediate, medium, and long-term solutions. AIC has the potential to be treated like a utility-like infrastructure. 

Public policy must be designed so that public investment benefits the broadest array of AI ecosystem firms and stakeholders, that public investment does not simply replace private investment, and that any public involvement is well governed and adaptive to quickly-changing policy and industry circumstances.

The following options summarize the limited range of scenarios that we see available to the Government of Canada. The options are ordered from shorter to longer time horizons, in consideration of the significant present needs of Canada’s AI ecosystem.

Scenario 1 | Centralize and subsidize AIC via federal procurement from cloud providers 

Proposal: AI Compute resources can be purchased via enterprise cloud computing services. To help close the AI Compute gap, this scenario proposes a national effort to purchase capacity at as large a scale as possible to address immediate AIC access barriers for the domestic AI ecosystem. The scale of immediate AIC access needs could be assessed through a request-for-proposal process and operationalized through cloud-platform-allocation programs. 

Providing direct support for cloud-computing costs, similar to the recent programs administered by France, would address the considerable need for computing resources endemic to Canadian innovators and nascent AI firms. 

Ideally, the federal government would acquire the cloud AIC resources in bulk to increase the likelihood of accessing the expedient and scalable capacity at the lowest cost, based on demand, growth potential, and equity considerations for AIC. 

Implementation: Given the urgency of the matter, the government would need to allocate funding and propose any legislative provisions through the 2024 Budget process, with an intention to have a market in place for Canadian firms to buy from as soon as possible. While the total demand from Canadian firms and their ability in the short term to switch providers is currently unknown, a purchasing program on the scale of hundreds of millions annually would both reduce overall costs and improve access to the domestic AI ecosystem.61

The fiscal cost to the government would depend on the extent of the subsidy or discounts provided to buyers, and which types of organizations would be eligible (e.g. research, start-ups, SMEs, large enterprises).

An alternative option is to provide direct subsidies vis-a-vis cloud service provider discount programs with firms already registered in existing technology-related programs. This direct subsidy reduces the public allocation and procurement burden and can be used to gauge the total market demand for cloud AI Compute by start-ups, SMEs, and researchers. The uptake and cost of the direct-subsidy approach would be covered by public resources, and the discount rate could be adjusted accordingly based on firm type and size in future iterations of the programs. 

Pros: Centralized procurement and direct cloud-service-provider purchases are the most immediate path to increased affordable Canadian AIC access. This policy would be of substantial benefit, especially to early stage AI companies.

Cons: These subsidies will further increase Canada’s dependence on the US and cloud computing suppliers for access. Even with federal purchasing power, our market size may not afford sufficient economies of scale. This policy carries a fiscal cost, though this can vary with decisions around the size of the subsidy available.

Scenario 2 | Work with key trade partners to jointly purchase AIC at scale

Proposal: Canada can acquire even larger quantities of AIC and increase its negotiating power by joining existing international collaboratives, and through trade collaboration, by buying with some combination of European, British, Japanese, and other partners. Canada is uniquely positioned to lead such an initiative based on our international trade, diplomacy, and AI expertise. These partnerships would drive down the costs of AIC infrastructure procurement and access to additional cloud computing from existing public and private high-performance computing infrastructure internationally through joint purchasing, procurement and research agreements. 

Implementation: The Canadian Government could expand on recent agreements with the UK and chipmaker Nvidia, and older multilateral work with the Global Partnership on Artificial Intelligence, among others, to support the development of national AIC Compute capacity through friend-shoring and expanding our ties to various players across the AI supply chain.62 63

Canada is concluding negotiations in 2024 to accede to Pillar II of the Horizon Europe research and innovation program. As a result, Canada could explore joining the European High-Performance Computing Joint Undertaking (EuroHPC JU), and joining other non-EU states such as Israel and Iceland in this undertaking. Canada should also support and take advantage of massive US investments in the CHIPS and Science Act and our unique relationship with the US to negotiate greater access to long-term AIC hardware procurement. 

Despite some loud voices to the contrary, Canada should not be building domestic chip manufacturing capacity (i.e. should not build chip foundries).64 65 66 It would be an inefficient use of federal dollars and would not, from security, economic development, or other potential perspectives, deliver a satisfactory return on investment. By contrast, proactive procurement of future AIC hardware and infrastructure over long time horizons could ensure Canada has sustainable supply chains relative to its international peers. Multi-firm procurement of AI infrastructure and cloud services may require more complex upfront legal and contractual obligations, but will reduce the risk of long-term lock-in to subpar technology and services.67 While the federal government has recently hinted68 at the need to work with allies to look at “joint opportunities” in providing more computing power, more details need to be thought out before implementing this idea. 

Pros: While trade agreements and multi-national initiatives usually take time to build, they should be faster than building domestic AIC capacity and they play to Canada’s unique advantages. Scale and purchasing power are of the greatest importance for AIC due to their impact on price, after the immediacy of access. Robust multi-partner procurement reduces some costs for the domestic AIC ecosystem. It also has the potential to make more overall compute available to Canadians, compared to peers who might have more domestic computing capacity. 

Cons: The way technology is governed internationally is changing fast, and the possibility of joining alliances may go down over time. This approach may also result in Canada having multiple partnerships with either interoperability issues, or lower performance per dollar in the long run. 

Scenario 3 | Build domestic AI supercomputing capacity

Proposal: Canada procures pre-built high-performance computing hardware from third-party AIC computing vendors, with performance specifications similar to South Korea’s Olaf supercomputer, which would provide deployment of additional AIC hardware for the AI ecosystem over a three- to five-year time horizon.69 In addition, a public-private AIC resource allocation program, similar to Finland’s business research program, could be introduced to offset the upfront investment and address domestic AIC access challenges for industry.

Currently, Canada’s domestic AIC infrastructure has less than half the capacity it needs to reach parity with international competitors. For example, for Canada to triple its capacity and add 123 PFLOPS of Nvidia H100 AIC-specific compute infrastructure and reach internationally comparable AIC performance capacity, our preliminary estimates suggest it would require upwards of $400 million in AIC-specific hardware procurement, integration, and operation costs.70 71 72 73 Yoshua Bengio has estimated that the cost of sufficient high-performance computing infrastructure as a whole is about one billion dollars.74

Implementation: The uncertain future of AIC cost-to-performance makes it difficult to forecast the clear option for future domestic infrastructure investments. Vendors that handle the increasingly complex landscape of large-scale computing infrastructure in conjunction with a portfolio of AIC chipset design firms could offset the risk of potential AIC supply chain shortages. It would be prudent to secure multi-year, interoperable infrastructure agreements to build out domestic AIC capacity that could be offset by reselling or repurposing aging and incompatible legacy AIC infrastructure for industry needs and educational institutions. While the immediate intention is to ensure specialized AIC infrastructure is being procured over a longer horizon, the need for multi-firm public-private partnerships ought to be considered to ensure that resilient, long-term supply chains are established for unforeseen innovation in Canada’s AI ecosystem. 

A large domestic build-out of AI Compute could result from a public-private partnership. The federal government could help convene and join a consortium of major Canadian capital providers who could build supercomputing capacity. The capacity could be made available to a variety of private-sector actors as well as the federal government at globally market-competitive rates. The federal government could then use its membership in the consortium to help subsidize or make access available to others, including start-ups, to help achieve objectives to ensure that more AI talent and organizations stay and grow in Canada.

Pros: This scenario gets Canada closer to self-reliance, to greater performance parity with our international peers, and to more domestic sovereignty in AI capacity. Computing allocation can be more cost-effective, closely monitored, and distributed more securely among research and industry stakeholders.

Cons: Potentially the highest fiscal cost to the government, and dependent on many factors including foreign nations and supply shortages. This option has a relatively longer time horizon to procure and operationalize with existing domestic computing infrastructure compared to cloud service providers' resources.

Policy considerations under all scenarios

The three options are not mutually exclusive and can be sequenced. For example, “Centralize and Subsidize” (Scenario 1) and “Work with Key Trade Partners” (Scenario 2) can be pursued in the near term to ramp up AIC capacity more quickly and offer a bridge to the domestic “build” (Scenario 3) of new capacity over a longer horizon.

The aforementioned scenarios are focused in particular on AIC capacity and do not address allocation mechanisms (i.e. AIC effectiveness) of expanded domestic AIC capacity. Since the majority of private enterprises access AI Compute resources internationally, understanding the aggregate demand for such services would require considerable AIC ecosystem consultation. Allocation mechanisms need to consider potential market disruptions of private AIC infrastructure investment.  A separate analysis would be necessary to identify the magnitude and dynamic response to public AIC infrastructure and subsidy programs. This analysis would specifically address the scope, timelines, and costs that would dynamically arise from differing levels of public intervention in the domestic market for AIC infrastructure and cloud service provision. 

Bridging the AI Compute gap should align with Canada’s positioning as a leader in responsible AI innovation and adoption, including the Pan-Canadian AI Strategy, the Artificial Intelligence and Data Act (AIDA) legislative and regulatory process, and other initiatives.75 In line with the thinking of some of our AI leaders, investment in AIC can be directed towards positive use cases including health and safety.76

Bridging the AI Compute gap must factor implications for energy use and climate objectives. Canadian AIC infrastructure is currently half as energy efficient as AI in the United States.77 Yet, Canada’s electricity is predominantly “green”,78 the aggregate cost of power is cheaper than in many other nations including the US and the UK,79  and the colder climate80 offers the potential for AIC to be more efficient in its cooling needs. These factors create the conditions for cost-competitiveness in attracting AIC infrastructure investments to Canada, while also aligning with Canada’s climate goals. 

These options remain preliminary and are aimed at informing fast-moving discussions across policy and AI ecosystem actors. Further research, analysis, and industry engagement are required to assess the feasibility and cost-effectiveness of these various options. 

Policy design considerations and risks

As the government deliberates investing in private-sector AIC, considerations include:

• Insufficient demand: One major risk of large government purchases of AIC is capacity being underutilized, for the purchase rise far exceeding current real demand for AIC. As the procurement mechanism of Scenario 1 relies on non-market mechanisms, a clear understanding of demand must be established before any policy implementation.
• Private investment crowd-out: Given the very small current domestic AIC capacity accessible to the private sector, any large government procurement/subsidization programs will have second-order pricing impacts that affects private companies accessing AIC resources outside of these programs. These dynamic crowding-out effects should be carefully modelled and understood in all scenarios.
• Innovative market mechanisms: In the ideal, medium-to-long-term future for Scenarios 2 and 3, the organization(s) deploying domestic AIC would have sufficient capacity to employ markets to either recoup the costs of this investment through margins on the savings generated by scale, or even through taking equity in private AIC buyers.
• Leverage past tech market models: While very concentrated, the market for cloud storage has also become very dynamic with the ability to move computing and storage functions across vendors in real time. As the AIC market matures, Canada will need to ensure its AIC can similarly compete on dynamism.
• Implementation inefficiencies: It is of particular concern that if these options are to be undertaken, the way a particular program is delivered impacts program use and uptake. Decisions around eligibility, processing timelines, and direct subsidy thresholds are more critical than the resources being offered by public intervention. This can be particularly important in deciding on the correct vehicle of distribution and eligibility for Scenario 1, as expediency and access are most important for addressing immediate cost and access barriers to AIC in Canada.