Skip to content
From Potential to Performance:

Canada’s Investment in AI Compute Infrastructure

October 2024

Foreword: The Dais’ AI Policy Approach

The Dais at Toronto Metropolitan University is an independent non-partisan policy and leadership think tank dedicated to building the bold ideas and better leaders needed to ensure more shared prosperity and shared citizenship for Canada. We do this in part by engaging in deep public-policy research, advising and advocating on innovation issues in Canada, and providing a depth of qualitative and economic policy analysis.

If it is governed well, artificial intelligence (AI) technology has the potential to bring more wealth and prosperity to Canada, into the hands of Canadian workers, firms, and researchers.

Our organization has a long track record of conducting policy-relevant economic research on the impact of AI on Canadians and firms. We do this through surveys, roundtables, and analysis of public and enterprise data for emerging public policy needs. The depth of our work at the Dais bridges the economic opportunity of building AI Compute infrastructure1 and promoting business adoption2 with proposals on how to better govern Canada’s AI economy so that it works well for workers,3 4 including through participation in the Standards Council of Canada’s AI and Data Governance Standardization Collaborative. In other words, our work in AI spans four relevant policy domains: adoption and commercialization, skills and talent, infrastructure, and governance.

At the request of the Dais’ advisory board, we researched and published Canada’s first public independent report on the country’s AI Computing capacity in March 2024. The report found that Canada has allowed a gap in AI Computing capacity that could threaten our talent and innovation advantage. Our report helped to inform the creation of the Government of Canada’s $2 billion, five-year fund to increase AI Compute capacity.

This report informs the federal government’s consultations around the design and implementation of the AI Compute Access Fund and the Canadian AI Sovereign Compute Strategy. These programs seek to catalyze further infrastructure expansion among institutional investors, start-ups, scale-ups, researchers, and government entities.5 These programs need to be designed with expertise, since they are currently not large enough to meet the demands of all stakeholders.

Our "AI Compute Infrastructure Roundtable" highlighted the need for more multi-stakeholder dialogue. Much of the information about AI computing capacity is held in private hands, and is commercially sensitive and market moving. And yet this is a burning public-policy issue.

When we convene at the Dais, we don’t ask parties to leave their interests at the door, but we do create a setting for a more intentional conversation, and to force to the surface the underlying issues.

On Wednesday, June 19th, with the support of IREN, an owner and operator of next-generation data centres and AI cloud services, the Dais hosted a multi-stakeholder roundtable on AI Compute infrastructure. We shared a framing paper, based on our March 2024 “Can Canada Compute?” paper, as an important factual context. Critical issues related to Canada’s AI Compute capabilities were addressed through structured discussions and expert insights.

Our process, proceeded from a factual basis, helped stakeholders better hear each other’s perspectives. It created a venue for better public policy around the procurement and uses of AI Compute and downstream AI innovation and adoption. At the end of this report, all participants who agreed to be associated with this final report are identified.”

Karim Bardeesy, Executive Director,
The Dais at Toronto Metropolitan University

“IREN provides AI cloud and colocation services, powered by 100 percent renewable energy. Our next-generation data centres in British Columbia are home to one of Canada’s largest deployments of AI infrastructure, and today service a diverse range of AI customers.

Canada’s combination of affordable and sustainable energy, plus AI research talent and innovative start-ups present a significant opportunity for Canada to lead globally in AI infrastructure.

IREN’s 160MW of operating data centres in British Columbia is sufficient to host >120,000 NVIDIA H100 GPUs. This capacity is immediately available to help solve the AI Compute infrastructure challenge.

IREN stands ready to support the growth and development of Canadian AI, while also supporting the local communities in which we operate through employment opportunities, skills training, and economic development.”

Kent Draper, Chief Commercial Officer,
IREN

Executive Summary

Canada was the first G7 country to develop a national AI strategy, and we continue to be a global leader in AI innovation. However, we are now lagging in the adoption of AI in our economy and need access to more computing power to take full advantage of our assets. Rapid action is required to ensure local, sustained, broadly shared economic benefit from AI innovation.

In June 2024, the Dais at TMU brought together start-up, scale-up, and established private sector players, along with not-for-profit, research, and government leaders, to discuss approaches to ensuring greater availability of AI computing infrastructure for Canada. A key focus was to propose options and recommendations for the deployment of the recent five-year, $2 billion Government of Canada investment in Canada’s AI Compute infrastructure, the central element of its new $2.4 billion investment in the sector.

The core of the roundtable consisted of three lightning talks and discussions on pivotal issues. The first issue concerns the hardware requirements for AI computing resources in Canada. The second is Canada’s lack of data centres, and the fact that this dearth is in spite of Canada’s affordable, sustainable energy advantages. The third issue is the need to establish a marketplace to form stronger negotiating AI Compute procurement coalitions and provide subsidized access to research and innovation ecosystems.

The roundtable’s open and unattributed format allowed participants to speak freely about the upcoming support that the program needs, and the ecosystem challenges of AI Compute infrastructure access, expansion, and sustainability.

The global demand for AI Compute resources continues to grow, and domestic public-private investment strategies are required to deliver sustainable long-term solutions for economic growth and effective program delivery.

Components of AI computing infrastructure

The roundtable focused on three critical components of improving Canada’s domestic AI Compute infrastructure:

  • High-Performance Computing (HPC) infrastructure: The hardware and software necessary to ensure effective and scalable computing and data centre resources across the economy.
  • Competition and innovation ecosystem modernization: The underlying research, innovation and early-adopter firm ecosystems driving the demand for AI infrastructure investment and support programs.
  • Energy: The capacity needed to build and operate AI infrastructure and data centre expansion.

Policy Considerations

Alongside identifying the barriers and challenges of AI computing infrastructure procurement, costs, and long-term sustainability, three considerations emerged from the discussions of Canada’s AI computing landscape:

  1. AI computing infrastructure investments must act urgently on the AI Compute consultation and needs of its AI ecosystem: The Government of Canada’s $2 billion investment commitment is significant, but will occur in an environment in which hyper-scalers, sovereign wealth funds, and the most well-capitalized AI companies in the world have projects to procure new computing power, talent, and intellectual property (IP) that run in the tens of billions of dollars. In that context, AI Compute investments in Canada must consider:
    1. Demand-side approaches that incentivize value-oriented procurement and business adoption by developing frameworks for cost-effective procurement of AI Compute chips and infrastructure based on firm maturity, industry, and potential productivity benefits.
    2. Supply-side approaches that leverage existing Canadian data centre capacity, as well as building out the required longer-term energy and AI Compute capacity, while nurturing intellectual property and talent retention to bolster the longer-term economic value of domestic supply chains and innovation ecosystems.
    3. That the combination of these policy approaches will require the oversight of a public-interest intermediary that can assess the technical computing needs of AI infrastructure and the economic value of sovereign AI Compute and access programming.
    4. That there is a need to facilitate the responsible adoption of AI for Canadian workers, firms, and society, while continuing to serve the needs of researchers and other innovators whose talent has kept Canada’s AI ecosystem strong.
  2. AI Compute infrastructure expansion must be coordinated as part of Canada’s larger energy supply needs and growth: Demand for energy to power Canadian data-centre expansion was high even before the Government of Canada’s AI Compute commitment. Tapping into additional domestic data centre capacity sources, including in regional areas, could help solve this supply challenge. In addition, further AI Compute energy demand requirements must be incorporated into Canada’s long-term energy planning, requiring provincial and local governance coordination. With Canada’s growing energy needs, AI Compute infrastructure expansions face economic and social barriers. Public and private stakeholders need to secure the significant social buy-in of communities affected by high-performance computing infrastructure and data centre expansion.
  3. Existing innovation and commercialization programming must be modernized to retain top AI talent and intellectual property: Canada’s AI infrastructure and developer start-ups and scale-ups need globally competitive public support for IP creation, talent retention, and commercialization efforts. Existing innovation support programs such as the Scientific Research and Experimental Development (SR&ED) tax incentive, the Industrial Research Assistance Program (IRAP) through the National Research Council, the Canada Infrastructure Bank, the Strategic Innovation Fund, and Innovative Solutions Canada need more tailored and responsive processes for Canadian AI enterprises.

Our roundtable surfaced many actionable ideas, based on the considerations above, to inform the design the AI Sovereign Compute and Access Programs. These are listed below:

Demand-side policy approaches

Offer procurement subsidies for AI Compute hardware and systems

  • Subsidy programs for the AI Compute infrastructure needs of firms, scale-ups, and small and medium enterprises (SMEs) are allocated by a public intermediary with domain expertise in economic analysis and computing infrastructure operations, with careful consideration of the cost-to-computing performance needed.
  • Procurement programming subsidy rates and eligibility should be established through extensive economic analyses of the market demand for physical AI Compute infrastructure, the price-to-performance of computing infrastructure for specific industry needs, and the potential productivity implications on domestic firms, workforces, and innovation ecosystem investments.

Expand strategic industry co-investment for AI Compute through public-private partnerships: 

  • Allocate a portion of AI Compute funding to multi-stakeholder partnerships, incentivizing adoption-driven AI Compute infrastructure expansion.
  • Provide expedited funding for partnerships aligning with Canada’s national interests through innovation ecosystem partnerships, net-zero goals, and sectors that serve to benefit from AI-augmented productivity gains. Public-private partnerships, similar to co-investment strategies adopted by Canada’s Scale AI and DIGITAL investment portfolio, can serve as an approach for growing the domestic AI value chain.

Supply-side policy approaches

Coordinate short-term AI Compute access and long-term sovereign infrastructure investments for research and innovation ecosystems:

  • The expansion of AI Compute infrastructure will require time-intensive planning and execution for the expansion of high-performance computing sites. AI researchers’ and start-ups’ immediate computing needs must be addressed by providing expedient access to currently available, in-country cloud computing service providers (examples of firms offering capacity from Canada include IREN, Qscale, and others).
  • A portion of available funds should be allocated to subsidize cloud computing access, with a preference given to service providers operating from within Canada for short-term needs while simultaneously planning domestic HPC site expansion for long-term sovereign computing capacity.
  • Establish a clear asset map of existing data centre capacity within Canada, identifying which firms have the immediate capacity to host AI Compute at scale. Domestic data-centre AI Compute capacity sources exist, but are not being used at full capacity for AI Compute.

Modernize intellectual property and commercialization processes:

  • Modernize IP-related programming and expedite patent processes for AI-related innovations, particularly in strategic sectors for the AI supply chain such as semiconductor design, infrastructure development, and energy-saving AI technologies.
  • Support the development and retention of IP and skilled labour in Canada for sectors critical to the AI Compute infrastructure supply chain, enabling the opportunities for AI supply-chain firms to attract the necessary capital to scale globally. 

Align domestic AI Compute expansion with energy expansion requirements and environmental sustainability goals:

  • Integrate infrastructure expansion project proposals with energy generation requirements that align with Canada’s pre-existing clean energy expansion goals.
  • Provide regulatory oversight of AI Compute infrastructure expansion by aligning national energy demands and environmental sustainability of infrastructure construction, operation, and maintenance to ensure data centre and high-performance computing projects forecast and budget for future energy requirements.

As Canada moves forward with its $2-billion investment in AI computing, the Government of Canada will need to coordinate a multi-sectoral approach to energy, innovation, and industry policy. This will help to fully understand current levels of capacity and address the gaps in infrastructure for a potential return on investment. By carefully considering these policy areas, Canada can maximize the economic impact of its AI Compute infrastructure investment, foster a thriving AI ecosystem, and address broader long-standing societal and economic objectives through a multi-faceted investment strategy.

Introduction

Artificial intelligence (AI) is touted as a general-purpose technology innovation, showing value for productivity-boosting and cost/time-reducing applications. Recent OECD reporting, summarizing a growing international body of research on AI’s economic implications, finds that the emergence of generative AI shows productivity benefits for both workers and nations. The authors posit that improving digital infrastructure and skills development will be critical for AI adoption and economic growth.

AI Compute infrastructure can enable economic benefits through several pathways:6

  • Productivity gains of workers through task automation or time reduction
  • Cost and energy reductions for firms through production-related efficiencies
  • Net new economic growth through commercialization of AI innovations and related job creation

Physical computing infrastructure is a necessary
tangible input to operationalize emerging AI innovations for broader productivity growth. AI Compute infrastructure is currently a critical input gap in Canada, despite our wealth of talented AI innovators and entrepreneurs. Investing in AI Compute infrastructure expansion has the potential to catalyze AI adoption, security, and innovation, and is a critical component of sovereign economic policy in the digital age. Despite the large infrastructure deficit, the IMF ranks Canada’s workforce preparedness for AI as among the top internationally but among the lowest for G7 nations.7 8 The need for policy measures to advance AI infrastructure expansion, particularly concerning energy needs, human capital gains, and inclusive industry adoption, will be critical.

What is Artificial Intelligence Compute?

From our previous report we noted AI systems require computing power to generate inferences from their inputs, which are generally conducted by parallel computing infrastructure such as high-performance computing or cloud-based computing systems and data centres. 9

We use the OECD’s working group definition of Artificial Intelligence Compute (AIC):

AI Compute is a specialized stack of hardware and software involving processors or chips, servers, storage, software, and networking, all designed to support AI-specific workloads and applications. It thus covers a range of different technologies, from AI chips to data servers to cloud computing.

AI development has two phases: training and inference. In the first phase, an AI model is “trained” on data. In the second phase, a trained AI model is deployed and then “infers” (i.e., makes decisions or takes actions) in the field based on new data. Training and inference phases run on AI Compute infrastructure in a private data center or in the public cloud.

Current State of AI Compute infrastructure in Canada

Accessing AI Compute in Canada presents significant challenges for research and innovation ecosystems. Initially, only a small fraction of the 2017 Pan-Canadian AI Strategy (PCAIS) funding was directed towards improving infrastructure. This limited allocation has hindered the ability to effectively scale AI research and development. The renewed interest and recent announcement of a $2 billion federal investment in Canada’s AI Compute infrastructure raises critical questions about the allocation and program delivery method of said funds.

Domestic high-performance computing environment for AI innovation ecosystems

Canada’s current AI Compute capacity is only a fraction of that available in leading countries like the US, Japan, and China. For instance, Canada’s computing power is approximately 41 petaFLOPS, just one percent of the 3,700 petaFLOPS available in the US. Canadian AI innovators and researchers often rely on US-based cloud providers, which can be both cost and operationally restrictive. Furthermore, Canada’s domestic capabilities for procuring HPC and data centre infrastructure, namely tensor processing units (TPUs) and accelerators, are constrained in supply by the immense international demand from nations and multinational technology firms.

Access to enterprise AI Compute infrastructure

Canadian AI start-ups and research institutions face high costs and too few resources to permit access to advanced computing resources. These often represent the most significant barriers to scale and commercialization. Canadian policymakers have promised targeted investments in AI computing infrastructure supporting research and enterprise needs.

In June 2024, with the support of IREN, an owner and operator of next-generation data centres and AI cloud services, the Dais hosted a multi-stakeholder roundtable on AI Compute Infrastructure. A framing paper, based on our March 2024 “Can Canada Compute?” paper, set factual context, and the format consisted of structured discussions and expert insights.

The following section summarizes the key challenges, discussion highlights, and recommendations that emerged during the Dais roundtable, grouped into three sections: high-performance computing infrastructure, competition and innovation ecosystem modernization, and energy.

Roundtable Challenges, Discussions, and Recommendations

High-Performance Computing (HPC) infrastructure 

Roundtable Question:

What hardware should Canada be building related to AIC? Who? Where? When?

Issue background

The growing computational needs of AI and efficiency innovations

Recent reporting shows AI computing power requirements have doubled every nine months, whereas AI Compute efficiency has only doubled every two to three years. 10 11Researchers at Stanford University estimate that the computational needs for training frontier AI models have quadrupled since 2020.12 Computational efficiency innovations for frontier AI have shown promise, but are yet to be widely adopted or implemented in the majority of model training processes.13 Fortunately, this large growth of computational resources for frontier AI training will necessitate more efficient chip designs in the very near future. 

Canada’s domestic semiconductor ecosystem

Canada is home to over 100 companies focused on semiconductor R&D, autonomous vehicles, and generative AI. AI infrastructure and compute hardware firms in Canada, like Ranovus, Tenstorrent, Waabi, Untether AI, Tartan AI, Cerebras, and EPIC Semiconductors, manufacture cloud infrastructure, server compute cores, and AI-specialized inference hardware.14 Investing in a broad portfolio of industry applications will give Canada’s economy the potential to commercialize and adopt specific use cases and chip designs that are in short supply globally.

The growing demand for semiconductors globally has driven increased interest in application-specific integrated circuits (ASICs) for industry-specific uses. Alongside the $2 billion in AI infrastructure expansion investment, the federal government has already invested close to half a billion dollars in semiconductor start-ups and manufacturing plants domestically through Canada’s Strategic Innovation Fund (SIF), with the bulk of funding invested in programming to accelerate semiconductor prototyping and production.15 16

Benefits of expanding AI Compute infrastructure for domestic semiconductor innovation

Capturing growing market demand for more efficient and competitively priced AI Compute hardware is an opportunity for Canada’s AI innovation ecosystem, with increased public investment for HPC and data centre access programming and subsidies. Canada’s support of the efforts of domestic semiconductor design firms will also be instrumental in improving AI Compute energy efficiency and emerging computing capabilities.17

An example of this is found in the current semiconductor design landscape. Nvidia, the incumbent market leader in AI semiconductors, began developing applications for its data processing library (called CUDA) in 2006, almost 20 years after it started as a graphics processing firm. The investment in hardware implementation for AI computing was not realized until the inception of AI pioneer Geoffrey Hinton and his team’s computer vision innovations that leveraged Nvidia’s proprietary CUDA processing cores in 2012.18 Hinton’s team demonstrated CUDA’s computational efficiency in meeting its AI training requirement for the research ecosystem. The demonstration provided the catalyst for the immense demand and expansion of Nvidia’s enterprise compute offerings outside consumer and commercial graphics processing, leading to its eventual top position among publicly traded firms in 2024.

Roundtable discussion

Varied needs of industry and research ecosystems

The industry’s AI Compute needs are widely varied, which poses a challenge for policymakers to optimally allocate resources among AI innovators, scale-ups, and series-funded AI enterprises in Canada. Iterative and multi-faceted investments in infrastructure are necessary to reap the benefits of adoption and innovation opportunities.

Allocating AI Compute funds will be challenging because of the need to balance large, centralized AI Compute facilities for model building and geographically dispersed AI Compute facilities for enterprise needs. This will be a consideration for cloud service providers, innovation ecosystems, and AI-adjacent industries looking to expand HPC systems domestically.

There are three main needs for AI Compute: model training, fine-tuning, and inference. For training, centralized and parallelized AI Compute infrastructure is critical to domestic innovation and research ecosystems. For inference and tuning efforts, the localization of HPC infrastructure for domestic-enterprise AI inference presents a less computationally intensive need, as physical infrastructure is critical for firm-specific tuning of pre-trained neural networks and generative AI business adoption. In the case of market-based infrastructure expansion, dispersing domestic AI Compute infrastructure regionally reduces latency (or response times) and capacity-load balancing for products and services that leverage cloud computing infrastructure.

Allocation frameworks for public AI Compute infrastructure subsidies

Roundtable stakeholders identified examples of AI Compute allocation frameworks to address cost and access concerns. To identify the firm-specific infrastructure needs and market demand of physical AI Compute infrastructure, a call for proposals is necessary to better understand the ability to effectively subsidize purchases at the national level. Scaling to a national AI Compute allocation framework would also require a public intermediary (or private-public coalition) with the technical expertise to assess value-for-money infrastructure procurement, as well as the net impact of potential firm and worker productivity implications for the supply-side policy approach.

Semiconductor innovation and industry support

As innovations in AI applications increase exponentially, supporting domestic semiconductor innovation ecosystems can generate substantial downstream impacts for nascent firms developing more efficient computing infrastructure hardware. The need to support domestic infrastructure innovation development is as critical as supporting AI Compute infrastructure expansion.

Recommendations

A multi-faceted approach to targeting semiconductor design, infrastructure procurement, and enterprise-adoption challenges suggests:

  • Investing and subsidizing capital for domestic fabrication-less semiconductor firms, which Canada has a history of with semiconductor company ATI Technologies (acquired by AMD in 2006) and many more, is an important, long-term strategy that holds downstream economic growth opportunities for job and firm creation.
  • Subsidizing the AI Compute hardware procurement needs of established AI scale-ups and innovation ecosystems, earmarking investments for AI Compute hardware expansion to retain Canada’s lead in AI research and talent-development ecosystems.
  • Developing international trade procurement and research access to compute agreements with like-minded countries, including buttressing supply chains and potentially building trade agreements to bolster purchasing power through scale.

Canada’s investment in AI Compute infrastructure must consider the shorter-term computing access needs in domestic AI enterprise and innovation ecosystems. In the medium term, Canada’s investment will need to support start-ups and researchers who desire to design and build AI solutions and computing infrastructure. Longer-term planning is needed in providing strategic support for AI Compute infrastructure firms in Canada in order to ensure the domestic AI Compute supply chains are robust and competitive in global markets.

Post-roundtable quote: CentML’s mission to improve the efficiency of AI training and inference

“CentML is a leading Canadian startup in software for optimized AI infrastructure delivery, operations, and management. We are the only company in Canada tackling the compute availability problem from a utilization-optimization angle.

Our mission is to make transformational GenAI technology accessible to everyone. We achieve this with the CentML platform, which pulls compute from different cloud vendors and applies advanced optimizations to reduce deployment costs. We offer our customers — Canadian enterprises and AI startups alike — the best performance-per-dollar option to deploy GenAI workloads.

Access to low-cost compute will allow us to deliver a greater competitive advantage for Canadian companies. The CentML platform can increase Canada’s collective compute capacity beyond hardware investment alone, enabling further innovation and creating opportunities to foster a new generation of Canadian AI startups.

- CentML

Competition and Innovation Ecosystem Modernization

Roundtable Question:

How can we strategically establish a program in Canada to provide subsidized access to not-for-profit adopters, frontier innovators, and early-adopter firms?

Issue background

Canada’s AI Compute capabilities lag international comparators

Adjusting for international market and population differences, Canada lags significantly behind in AI Compute capabilities—by a factor of 8 to 11 times compared to the US, approximately eight times compared to Japan, and two-to-three times compared to France and Germany. This gap signifies that the Canadian AI ecosystem has substantially less access to computing power domestically than its international peers.19

While Canada’s $2-billion investment is in line with similarly sized advanced economies such as the UK and France, AI Compute investments will need to provide mechanisms that boost adoption and crowd-in private infrastructure investment.20 21 22

The need for strategic and cost-competitive domestic AI Compute hardware continues to dampen the scale of commercialization efforts. Private infrastructure investments in Canada remain relatively low, with industry-owned computing accounting for less than 10 percent of computing capacity—significantly lower than the United States, where industry-owned computing comprises under a third of domestic performance capacity (31 percent). While enterprise cloud service providers have invested over $4 billion alone among major metropolitans in Canada, Amazon plans to spend $35 billion on new data centres in Virginia by 2040.23

Despite gaps in infrastructure, Canada holds a vital research and talent pool for AI

Canada leads in AI research publications per capita among G7 nations as of May 2023.24 Canada’s domestic AI workforce also represents a relatively larger proportion of its employed labour than the US in 2022, adding to the substantial opportunity for capital investment for AI enterprise.25 26 Canada also boasts a vibrant AI start-up ecosystem, including over 1,500 AI companies, venture capital investment, total funding raised, and patent filing per capita compared to all other G7 nations from 2022 to 2023.27 28 29 However, the demand for AI services and trust within Canada remains low relative to other OECD countries, indicating a significant growth opportunity: Dais research found only four percent of domestic enterprises have adopted AI, while 14 percent are experimenting with GenAI.30

Roundtable discussions

Stakeholders highlighted the need for support for innovation-oriented, domestically grown enterprises through expedited IP processes and subsidies through existing programs. The proposed benefits would be job creation, international competitiveness, and the long-term sustainability of domestic AI enterprise expansion.

At the roundtable, consensus emerged regarding Canada’s AI workforce in enterprise, infrastructure, and semiconductor design. This workforce is abundant, but they often seek employment internationally for career advancement and higher remuneration. Participants also reached the consensus that the expansion of computing infrastructure in Canada has the potential to support talent and IP retention, providing opportunities for skilled labour to stay here and contribute to the domestic economy.

Research stakeholders highlighted the need for policy to identify the national outcomes related to AI proliferation, whether industry-focused or a concern of national sovereignty. Renewable energy, critical minerals, and climate innovations are areas of industry focus that were highlighted in the 2024 federal budget that could align with AI-driven solutions to drive national prosperity.

Enterprise stakeholders mentioned the need to modernize innovation support programs for highly skilled talent retention and IP protections of AI-related services to foster early business ventures and global market entry. Providing protections and incentives for IP protection in critical sectors can attract and retain skilled labour, boost global competitiveness, and ensure sustainability for long-term economic growth in alignment with AI innovation.31

Recommendations

The synergies between public interest and private enterprise in AI can be seen as significant gaps in Canada’s domestic AI ecosystem or as substantial value propositions for Canada’s industrial and economic future. Canada’s robust AI ecosystem is a source of untapped economic value in Canada’s lagging macroeconomic performance and a critical driver for Canada’s productivity growth.32 33

  • Leverage existing innovation supports such as IRAP, SR&ED, the Strategic Innovation Fund, the Canada Infrastructure Bank, Innovative Solutions Canada and Regional Development Programs that aid transition to commercialization, to provide tailored and responsive processes for Canadian AI enterprises to remain competitive with international innovation support programs and competitors.
  • Expedite patent processes for firms in AI to foster international competitiveness for domestic firms, similar to existing processes for green technology.34
  • Address the immediate AI Compute needs of firms in critical industries and adoption-ready SMEs with subsidized access to domestically-based HPC cloud infrastructure and TPU procurement based on firm characteristics and compute resourcing needs.

Post-Roundtable Quote: Cerebras’ quest to keep up with semiconductor giants

…Cerebras’ mission is to address the AI Compute challenge by developing a full-stack AI accelerator solution, encompassing everything from chip design to system integration, software, and ML/AI. Over the past year, we’ve developed and open-sourced leading foundation models, with many of the key innovations emerging from our talented team in Canada (we have over 100 employees in Toronto). Our team excels in chip design, compiler development, and ML/AI research. We completely understand the challenge of leveraging compute for AI – it’s not easy. We firmly believe that Cerebras’ Wafer-Scale technology is ideally suited to meet Canada’s future AI compute needs. [We] advocate for Canada to invest in building AI Compute clusters that will have a lasting impact on the nation. By leveraging cutting-edge technologies, we can empower Canadian researchers and companies to pioneer new and innovative approaches to AI model development and deployment. Let’s harness the technologies of today and tomorrow to drive innovation in both compute and AI.

- Cerebras Systems Inc.

Energy and Environmental Sustainability of AI Compute Infrastructure

Roundtable Question:

Why has Canada’s lead on affordable, sustainable energy and the suitable climate for data centres not translated into more HPC centres? What can we do about it?

Issue background

The energy of AI Compute operations poses a significant barrier to expanding computing infrastructure and AI adoption in Canada. Infrastructure expansion needs to be considered a necessary but singular input to the complex relationship of innovation-driven economic growth. Investment decisions for infrastructure should consider the direct and indirect impacts of AI Compute on local communities.

Balancing the direct costs and uncertain benefits of AI Compute expansion

The OECD’s report on AI energy highlights considerations for the direct and indirect impacts of AI Compute infrastructure. The report underscores that the indirect outcomes of infrastructure, such as efficiency gains in industry and labour productivity growth, must outweigh the total cost of energy and resources required to build, operate, and maintain infrastructure. The tradeoff of infrastructure expansion emphasizes the need for:

  • Climate impact standards for AI Compute: To assess the environmental cost through transparent and measurable policy frameworks.
  • Incentives for sustainability: To encourage and promote research into using renewable energy sources and more energy-efficient AI technologies.
  • Proactive international governance: To set AI climate impact standards and sustainable pathways for AI Compute infrastructure expansion.

These recommendations have significant implications for Canada’s policy challenges with AI Compute expansion. Policymakers must balance the cost of infrastructure’s direct impacts with the uncertain economic benefits for citizens and industry.35 Careful consideration should be given to which infrastructure providers, industries, and firms globally are seeing net positive productivity gains from AI adoption relative to the energy and computing costs of the goods and services produced. In light of these considerations, we highlight some key challenges that stakeholders face in enterprise growth, infrastructure expansion, and energy regulations.

Domestic public infrastructure is energy-inefficient

Public infrastructure is available, yet Canada could use its energy more efficiently relative to the vast computing capacity in the US. Measured by performance as a function of kilowatts needed to power public-use AI Compute infrastructure, Canada’s effective HPC infrastructure is less than half as efficient as more powerful and abundant AI Compute in the US (0.007 versus 0.016 PFLOPS per kilowatt).36

The energy demands of AI Compute concern both energy regulators and AI Compute stakeholders

Canada (USD 0.125/kWh) offers lower commercial electricity costs to the US (USD 0.162/kWh), as measured by dollar-per-kilowatt-hour and almost a quarter of the UK’s electricity costs (USD 0.4/kWh). Canada’s colder climate also provides a comparative advantage for sustaining optimal AI Computing Infrastructure operating temperatures. This lower temperature delta (over 10 degrees Celsius compared to its southern neighbour) presents an opportunity to reduce operating and cooling costs.37 38 The price advantage supports Canada’s hospitality to AI Compute and data centre expansion, but will need to consider other cost factors.

Currently, data centres account for 1 to 1.5 percent of global electricity use and one percent of total carbon emissions.39 The International Energy Agency estimates that energy demands from AI Compute and data centres alone will double between 2022 and 2026.40

For perspective, Microsoft’s energy consumption has doubled between 2020 and 2023. Energy consumption growth is presumably tied to its increasing growth from cloud, enterprise, and software services, which account for over 75 percent of its total revenue growth. Google’s energy consumption saw similar growth from 2019 to 2023, and the company recently retracted its commitment to carbon neutrality.41 Estimates suggest that generative AI search engine integration is four to five times more energy-intensive than traditional search. The daily energy consumption from queries to OpenAI’s ChatGPT is roughly equivalent to the yearly consumption of 33,000 US households.42

For the first time in 8 years, Canada’s current electricity demands exceeded its generation capacity, as referenced by the net deficit of electricity exports in the first quarter of 2024.43 44 While the roundtable revealed significant sources of AI Compute data centre capacity available in Canada, it also raised considerations about the build-out of further AI Compute infrastructure in Canada.45

Roundtable discussion

Stakeholders raised concerns about domestic opposition to new AI Compute infrastructure projects. Proposals for expansion can raise regional concerns about energy outages and capacity. More traditional economic development projects, such as advanced manufacturing, are also perceived to create more direct economic growth and favourable employment outcomes for local communities than AI Compute projects. Securing the social licence to expand enterprise and public AI Compute could face opposition. Barriers from regional economic and energy regulations can also slow the execution of computing infrastructure projects. Data centre and cloud service providers often need to meet complex regulatory requirements.  Energy contract disputes often burden negotiations and execution of high-performance computing expansion in Canada relative to its international peers. That said, the roundtable also heard the perspective that Canada’s rural and regional communities have significant and underutilized electrical infrastructure that supports material amounts of existing data centre capacity.46

Stakeholders noted the demand for computational power will expand significantly even after initial investments, highlighting the need for persistent and interoperable infrastructure for domestic AI enterprise development and adoption. Consequently, the energy consumption of high-performance computing is significant and continues to increase as frontier AI integrates into enterprise services.

Similarly, stakeholders identified that the use of AI in non-tech industries has the potential to reduce energy consumption in the long run, but replacing existing products and services will take time to realize the energy and cost-saving advantages. Energy regulators and HPC providers will need to consider the regional energy needs of local economies alongside data centres and AI Compute infrastructure expansion.

Recommendations

Canada’s energy-generating capacity must grow with the expansion of the AI industry, a significant balancing act for energy and industrial policy. The need for public-private coordination in designing, managing, and executing the expansion of AI Compute infrastructure projects will require several considerations:

  • Conscientious regulatory design and programming of renewable energy expansions, incentives, and subsidies for renewable energy use alongside AI Compute infrastructure to attract private enterprise investment domestically.
  • Inclusion of and consultation with local and Indigenous communities concerning the social implications of infrastructure investment and job creation involve civil convening for both public and private interests that benefit local communities.
  • Responsible leveraging of legacy electrical infrastructure in rural communities that have seen a decline in jobs from traditional industries (e.g. closure of forestry-linked businesses in British Columbia).
  • Proper balance for the regional allocation of AI Compute expansion projects and mapping out energy use needs is necessary. For example, developing regional and strategic HPC centres provides optimal AI Compute locales for innovation ecosystems. In contrast, more minor, dispersed, and regionalized enterprise data centres and HPC infrastructure can improve the cost, energy, and latency concerns of domestic enterprise AI services and adoption.
  • Standardizing AI Compute infrastructure efficiency frameworks by employing leading practices in the transparency of energy and computational consumption of AI Compute used and the intended purposes, such as those employed by AI research institutes and cloud-computing industries.

Final Considerations

Through proactive policy and investment mechanisms, informed by multi-stakeholder discussions and consultation, Canada can catalyze its robust AI innovation ecosystem. However, several barriers must be considered alongside the shortage of AI Compute access. Talent attrition in Canada’s AI ecosystem is a consequence of the growing global demand for AI talent and services.47 48 49Despite growing demand internationally, Canada also lags in business adoption, which will be a critical factor for the value of expanding data and computing infrastructure domestically. 50 51Taking advantage of Canada’s AI talent and research strengths by actioning short- and long-term solutions through AI Compute infrastructure investments is necessary to address the challenges its firms, researchers, and public institutions face. The success of Canada’s AI infrastructure strategy will depend on new investments and the efficient use of existing resources. Recognizing and optimizing current capacities within the country is essential to meet both immediate and long-term AI infrastructure demands.

The Canadian AI Sovereign Compute Strategy and the AI Compute Access Fund represent critical opportunities to chart a successful path for Canada’s prosperity in the AI economy. June’s hosted roundtable and this summary report are part of the Dais’ ongoing efforts to inform the fast-moving policy and investment processes underway, part of a commitment to advancing AI adoption and innovations in Canada. The need to act urgently will have significant implications for the opportunities that lie ahead for the Canadian economy.

1

Graham Dobbs and Jake Hirsch-Allen, Can Canada Compute? Policy Options to Close Canada’s AI Compute Gap, The Dais, March 2024, https://dais.ca/reports/can-canada-compute/.

2

Angus Lockhart, Automation Nation? AI Adoption in Canadian Businesses, The Dais, 2023, https://dais.ca/reports/automation-nation-ai-adoption-in-canadian-businesses/. 

3

Viet Vu and Angus Lockhart, Submission to HUMA Committee on AI and Work, The Dais, 2023, https://dais.ca/reports/submission-to-huma-committee-on-ai-and-work/.

4

The Dais at Toronto Metropolitan University and the Centre for Media, Technology and Democracy at McGill University, Submission on the Proposed Artificial Intelligence and Data Act, November 2023, https://dais.ca/reports/submission-on-the-proposed-artificial-intelligence-and-data-act/. 

5

“Consultation on Artificial Intelligence (AI) Compute,” Innovation, Science and Economic Development Canada, June 26, 2024, https://ised-isde.canada.ca/site/ised/en/public-consultations/consultation-artificial-intelligence-ai-compute.

6

Francesco Filippucci, Peter Gal, Cecilia Jona-Lasinio, Alvaro Leandro, and Giuseppe Nicoletti, “The Impact of Artificial Intelligence on Productivity, Distribution and Growth: Key Mechanisms, Initial Evidence and Policy Challenges,” OECD, April 16, 2024, https://doi.org/10.1787/8d900037-en.

7

Giovanni Melina, “Mapping the World’s Readiness for Artificial Intelligence Shows Prospects Diverge,” IMF, June 25, 2024, https://www.imf.org/en/Blogs/Articles/2024/06/25/mapping-the-worlds-readiness-for-artificial-intelligence-shows-prospects-diverge.

8

 Mauro Cazzaniga, Florence Jaumotte, Longji Li, Giovanni Melina, Augustus J. Panton,Carlo Pizzinelli, Emma J. Rockall, Marina Mendes Tavares, “Gen-AI: Artificial Intelligence and the Future of Work,” IMF, January 14, 2024, https://www.imf.org/en/Publications/Staff-Discussion-Notes/Issues/2024/01/14/Gen-AI-Artificial-Intelligence-and-the-Future-of-Work-542379.

9

Excerpt from Graham Dobbs and Jake Hirsch-Allen, Can Canada Compute? Policy Options to Close Canada’s AI Compute Gap, The Dais, March 2024, https://dais.ca/reports/can-canada-compute/.

10

Jaime Sevilla, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn and Pablo Villalobos, “Compute Trends Across Three Eras of Machine Learning,” Epoch AI, last updated May 2, 2022, https://epochai.org/blog/compute-trends.

11

Marius Hobbhahn, Lennart Heim and Gokce Aydos, “Trends in Machine Learning Hardware,” Epoch AI, November 9, 2023, https://epochai.org/blog/trends-in-machine-learning-hardware.

12

Jaime Sevilla and Edu Roldan, “Training Compute of Frontier AI Models Grows by 4-5x per Year,” Epoch AI, May 28, 2024, https://epochai.org/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year.

13

Rui-Jie Zhu,  Yu Zhang, Ethan Sifferman, Tyler Sheaves, Yiqiao Wang, Dustin Richmond, Peng Zhou, and Jason K. Eshraghian, “Scalable MatMul-Free Language Modeling,” ArXiv, last revised June 18, 2024, http://arxiv.org/abs/2406.02528. 

14

Anita Balakrishnan,  “Here’s How Canada’s Semiconductor Industry Stacks up,” The Logic, July 5, 2023, https://thelogic.co/news/special-report/heres-how-canadas-semiconductor-industry-stacks-up/. 

15

Murad Hemmadi, “Canada Puts Up $120M for Program to Help Semiconductor Startups Get Going,” The Logic, July 4, 2024, https://thelogic.co/news/exclusive/sif-funding-semiconductor-startups/. 

16

“FABrIC – CMC Microsystems,” Accessed July 4, 2024, https://www.cmc.ca/fabric/. 

17

Daniel Munro and Creig Lamb , “Chip Shot: A Semiconductor Strategy for Canada,” Canadian Standards Association, 2024, https://www.csagroup.org/article/public-policy/chip-shot-a-semiconductor-strategy-for-canada/. 

18

Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” Communications of the ACM 60 (6) (May 24, 2017): 84–90, https://doi.org/10.1145/3065386.

19

Graham Dobbs and Jake Hirsch-Allen, Can Canada Compute? Policy Options to Close Canada’s AI Compute Gap, The Dais, March 2024, https://dais.ca/reports/can-canada-compute/. 

20

“Host Sites for the Next Wave of UK Government AI Infrastructure,” UK Research and Innovation, January 24, 2024, https://www.ukri.org/opportunity/host-sites-for-the-next-wave-of-uk-government-ai-infrastructure/. 

21

National Cloud Strategy: Launch of the Industrial Plan to Support the Sector | Enterprises.Gouv.Fr,” Direction générale des Entreprises, September 11, 2021, https://www.entreprises.gouv.fr/fr/actualites/numerique/strategie-nationale-cloud-lancement-du-plan-industriel-de-soutien-la-filiere. 

22

“United Kingdom Artificial Intelligence Market 2023,” International Trade Administration, March 29, 2023, https://www.trade.gov/market-intelligence/united-kingdom-artificial-intelligence-market-2023. 

23

Roger Wehner, “Learn about AWS’s Long-Term Commitment to Virginia,” Amazon, June 7, 2023, https://www.aboutamazon.com/news/aws/aws-commitment-to-virginia. 

24

“AICan: The Impact of the Pan-Canadian AI Strategy,” CIFAR, October 2023, https://cifar.ca/ai/impact/.

25

Graham Dobbs, Ibrahim Abullail and Angus Lockhart, Built to Scale? Microcredentials Use Among Digital Professionals, The Dais, October 21, 2023, https://dais.ca/reports/built-to-scale-microcredentials-use-among-digital-professionals/.

26

Diana Gehlhaus and Ilya Rahkovsky, “U.S. AI Workforce: Labor Market Dynamics,” Center for Security and Emerging Technology (blog), April 2021, https://cset.georgetown.edu/publication/u-s-ai-workforce.

27

“Canada Artificial Intelligence (AI) Companies,” Crunchbase Hub Profile, Accessed February 21, 2024, https://www.crunchbase.com/hub/canada-artificial-intelligence-companies. 

28

“Live Data,” OECD.AI Policy Observatory,” Accessed February 21, 2024, https://oecd.ai/en/data. 

29

“Impact and Opportunities: Canada’s AI Ecosystem – 2023,” Deloitte Canada, Accessed February 21, 2024, https://www2.deloitte.com/ca/en/pages/deloitte-analytics/articles/impact-and-opportunities-canadas-ai-ecosystem-2023.html. 

30

“Prompting Productivity: Generative AI Adoption by Canadian Businesses,” Canadian Chamber of Commerce Business Data Lab, May 16, 2024, https://bdl-lde.ca/publications/prompting-productivity-generative-ai-adoption-by-canadian-businesse/.

31

Daniel Munro and Creig Lamb, “Chip Shot: A Semiconductor Strategy for Canada,” Canadian Standards Association, 2024, https://www.csagroup.org/article/public-policy/chip-shot-a-semiconductor-strategy-for-canada/.

32

 “New Report Highlights How Generative AI Can Transform Canada’s Future with a Potential to Add $187B to the Canadian Economy by 2030,” Microsoft News Centre Canada, June 4, 2024, https://news.microsoft.com/en-ca/2024/06/04/new-report-highlights-how-generative-ai-can-transform-canadas-future-with-a-potential-to-add-187b-to-the-canadian-economy-by-2030/.

33

Rannella Billy-Ochieng, Anusha Arif and Daniella Garcia, “Artificial Intelligence Technologies Can Help Address Canada’s Productivity Slump,” TD, May 28, 2024, https://economics.td.com/ca-AI-tech-can-help-productivity-slump.

34

“Advanced Examination for Green Technologies,” Canadian Intellectual Property Office, Government of Canada, https://ised-isde.canada.ca/site/canadian-intellectual-property-office/en/patents/patent-application-and-examination/advanced-examination-green-technologies. 

35

“Measuring the Environmental Impacts of Artificial Intelligence Compute and Applications,” OECD, 2022, https://doi.org/10.1787/7babf571-en.

36

Data acquired from https://top500.org/statistics/sublist/, excluding HPC sites with no energy consumption in November 2023.

37

“Climate Comparison: Canada / United States,” n.d., WorldData.info, Accessed July 11, 2024, https://www.worlddata.info/climate-comparison.php?r1=canada&r2=usa.

38

“Electricity Prices Around the World,: December 2023,” GlobalPetrolPrices.com, Accessed July 11, 2024, https://www.globalpetrolprices.com/electricity_prices/.

39

“Data Centres and Data Transmission Networks,” July 11, 2023. IEA, https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks. 

40

“Electricity 2024: Analysis and Forecast to 2026,” IEA, January 2024, https://www.iea.org/reports/electricity-2024.

41

Ben Gomes and Kate Brandt, “Our 2024 Environmental Report,” Google, July 2, 2024, https://blog.google/outreach-initiatives/sustainability/2024-environmental-report/.

42

Sarah McQuate, “Q&A: UW Researcher Discusses Just How Much Energy ChatGPT Uses,” UW News (blog), updated August 2, 2023, https://www.washington.edu/news/2023/07/27/how-much-energy-does-chatgpt-use/.

43

The Canadian Press, “Electricity Imports from U.S. Topped Exports in February Amid Drought: StatsCan,” CTV News, April 29, 2024, https://www.ctvnews.ca/business/electricity-imports-from-u-s-topped-exports-in-february-amid-drought-statcan-1.6866326. 

44

Government of Canada, Statistics Canada, “Electric Power Generation, Monthly Receipts, Deliveries and Availability,” July 31, 2024, https://www150.statcan.gc.ca/t1/tbl1/en/tv.action?pid=2510001601. 

45

David Reevely, “Canada Wants AI but Doesn’t Have the Juice to Power It,” The Logic, June 10, 2024, https://thelogic.co/news/canada-wants-ai-but-doesnt-have-the-juice-to-power-it/. 

46

Martin Fairbank, “An ABC of Closed Canadian Pulp and Paper Mills,” Paper Advance, October 10, 2023, https://www.paperadvance.com/blogs/martin-fairbank/an-abc-of-closed-canadian-pulp-and-paper-mills.html. 

47

“2024 Edelman Trust Barometer,” Edelman Canada, 2024, https://www.edelman.ca/trust-barometer/2024-edelman-trust-barometer.

48

David Zarifa  and David Walters, “Revisiting Canada’s Brain Drain: Evidence from the 2000 Cohort of Canadian University Graduates,” Canadian Public Policy / Analyse de Politiques 34 (3): September 2008: 305–19, https://doi.org/10.3138/cpp.34.3.305. 

49

Vivian Li, Mind the Gap: Compensation Disparity Between Canadian and American Technology Workers, The Dais, October 2023, https://dais.ca/reports/mind-the-gap-compensation-disparity-between-canadian-and-american-technology-workers/.

50

Laura McDonough, “Employer-Sponsored Skills Training: A Picture Of Skills Training Opportunities Provided By Canadian Employers,” Future Skills Centre (blog), November 2023,  https://fsc-ccf.ca/research/employer-sponsored-skills-training-shift-lmic.

51

Angus Lockhart, Automation Nation? AI Adoption in Canadian Businesses, The Dais, August 14, 2023, https://dais.ca/reports/automation-nation-ai-adoption-in-canadian-businesses/.