Canada's Sovereign AI Compute Strategy: Necessary But Not Sufficient

Introduction

In December 2024, the Government of Canada announced one of its most ambitious technology initiatives in recent memory: the Canadian Sovereign AI Compute Strategy. Backed by $2 billion in federal funding over five years, the strategy represents Canada's response to a growing global consensus that artificial intelligence infrastructure is a matter of national sovereignty and economic security.

The initiative emerged from public consultations conducted throughout the summer of 2024, during which more than 1,000 stakeholders contributed their perspectives on Canada's AI compute needs (Government of Canada, 2024a). What became clear was a shared concern: Canadian researchers and businesses increasingly relied on foreign-owned cloud infrastructure, creating dependencies that posed risks to data security and protection of intellectual property. This concern is real. Unfortunately, the strategy does not adequately address it.

Sovereignty, at its most basic, means freedom from external control. By that standard, the Canadian Sovereign AI Compute Strategy falls short. It builds infrastructure on Canadian soil but does not address who controls the hardware, where it comes from, or who can access the data processed on it. The strategy treats location as a proxy for sovereignty, a simplification that overpromises what the investment can deliver.

This paper raises four concerns. First, the term “sovereignty” is misleading in this context. It implies the achievement of a level of independence from foreign control that Canada cannot attain given the realities of global supply chains. Second, infrastructure alone will not drive adoption without mandates, talent investment, and regulatory frameworks that compel its use. Third, stakeholder consultations raised concerns that the strategy does not address, and the government has not explained why it narrowed its scope to infrastructure alone. Fourth, the strategy establishes no measurable success criteria, making accountability impossible.

The thesis of this paper is simple: Canada’s Sovereign AI Compute Strategy is a domestic infrastructure investment, not a sovereignty strategy. It should be named and evaluated accordingly.

The Architecture of the Strategy

Canada’s Sovereign AI Compute Strategy is built on three interconnected pillars, each designed to address different aspects of the compute capacity challenge.

The first pillar, the AI Compute Challenge, allocates up to $700 million to mobilize private-sector investment in AI data centres (ISED, 2024c). This program operates on a competitive basis, inviting projects from industry, academia, and public-private partnerships to build or expand commercial AI-specific data centres across Canada. The program's objectives include safeguarding Canadian data security and sovereignty, providing flexible and affordable compute offerings for Canadian innovators, and advancing sustainable compute solutions.

The second pillar is the AI Sovereign Compute Infrastructure Program, commonly known as SCIP. With funding of up to $705 million (ISED, 2024b), SCIP aims to build a state-of-the-art, Canadian-owned high-performance supercomputing system that will serve academic researchers and private-sector research and development. The program's objectives extend beyond capacity building to include developing and retaining AI talent in Canada, protecting Canadian data, and fostering innovation through cross-sector collaboration.

The third pillar, the AI Compute Access Fund, addresses the prohibitive cost of compute resources for small and medium-sized enterprises. With up to $300 million allocated, this fund subsidizes the purchase of AI compute resources by Canadian innovators and businesses, with particular emphasis on sectors that require high-performance computing and show high potential for AI adoption, such as life sciences, energy, and advanced manufacturing (ISED, 2024a).

Complementing these three pillars are additional investments: up to $200 million for near-term augmentation of existing public compute infrastructure and a smaller secure computing facility led by Shared Services Canada and the National Research Council for government and industry research and development, including work related to national security (ISED, 2024a).

The strategy implicitly defines sovereignty in terms of four criteria: Canadian ownership, Canadian legal jurisdiction, physical location in Canada, and connectivity through Canadian networks. Notably absent from this definition is any consideration of hardware supply chains, workforce development, or policy reform. There are no mandatory adoption mechanisms, no data classification system, and no regulatory framework compelling use of domestic infrastructure. These omissions become significant when evaluating what “sovereignty” means in practice.

The Sovereignty Question

The strategy's central promise is sovereignty. But what does sovereignty require in the context of AI infrastructure?

True sovereignty means control over the entire technology stack: the hardware that performs computations, the software that runs on it, the supply chains that produce it, and the workforce that operates it (IEEE Communications Society, 2025). No nation fully achieves this standard. The global AI supply chain is deeply interconnected, with critical dependencies crossing borders regardless of where data centres are physically located. The question is not whether a nation has achieved complete independence, but whether it controls enough of the stack to have meaningful autonomy where it is deemed necessary.

The United States comes closest to this standard. American companies, led by Nvidia, control an estimated 80 to 95 percent of the data centre GPU market (Leswing, 2025). The U.S. also controls access to these chips through export restrictions that limit what can be sold to adversaries (Bureau of Industry and Security, 2022). But even the United States depends on Taiwan's TSMC for advanced chip fabrication. China, despite being subject to American export controls, has made massive investments in domestic semiconductor production and controls approximately 76 percent of global lithium-ion battery production (Shanghai Metal Market, 2025), a critical component of AI data centre infrastructure. Taiwan's TSMC manufactures over 90 percent of the world's most advanced semiconductors (MLQ AI, 2025), making every nation dependent on a single island's fabrication capacity.

What distinguishes the United States and China from other nations is not complete independence but bargaining power. American control over chip design and export policy gives it leverage in negotiations. Chinese control over battery production and rare earth processing gives it similar leverage. When supply chain disputes arise, these nations can negotiate from positions of strength. Canada cannot. Canadian data centres, including those funded under this strategy, will rely on chips from the US, batteries from China and semiconductors from Taiwan. If any major players face supply constraints or companies pivot their business strategies, or become subject to new export restrictions from competing nations, Canadian AI infrastructure would likely be significantly impacted, regardless of where the data centres are located.

The Bell Canada example illustrates how fast the ground can shift. In May 2025, Bell announced a “sovereign” AI network with Groq as its exclusive inference provider (Groq, 2025). At the time, Groq represented a genuine alternative to Nvidia, having developed specialized chips for inference workloads (Tom’s Hardware, 2025). Seven months later, Nvidia acquired Groq for approximately $20 billion, consolidating control over both training and inference markets (Faber, 2025). In less than a year, the competitive landscape that existed when Bell made its announcement had fundamentally changed. Canada’s sovereign AI compute strategy includes no mechanisms to anticipate, monitor, or respond to this kind of supply chain disruption. There is no contingency planning, no requirement to diversify hardware providers, and no framework for assessing how market consolidation affects Canadian infrastructure. The strategy assumes a static world in an industry defined by rapid change.

What the strategy actually delivers is more modest but still valuable: intellectual property protection and jurisdictional insulation. The U.S. Clarifying Lawful Overseas Use of Data Act (CLOUD Act), enacted in 2018, allows American authorities to compel U.S.-based technology companies to produce data in their possession regardless of whether the data is stored in the United States or abroad (U.S. Department of Justice, 2019). Data processed on Canadian-owned infrastructure, operated by Canadian companies, and subject to Canadian law is protected from these foreign legal demands. For organizations handling sensitive research, health records, or government data, this protection matters.

The problem is that "sovereignty" implies something more. It suggests freedom from external control, self-sufficiency, and independence. The strategy cannot deliver these things because they depend on factors outside Canada's control. Using the term sets expectations that the investment cannot meet.

Recommendation: Rename the strategy to reflect what it actually delivers. "Canadian AI Infrastructure Strategy" or "Domestic AI Compute Initiative" would be more accurate. Be honest with Canadians about what $2 billion buys: domestic capacity with intellectual property and jurisdictional protections, not freedom from external control.

Why Infrastructure Alone Won't Work

The strategy assumes that building infrastructure is sufficient. It is not. Seven interconnected problems explain why.

No mandates. The strategy builds capacity but does not compel its use. PIPEDA does not require data residency in Canada, and Protected B government data can be processed on foreign clouds with appropriate safeguards. There is no "licence to operate" framework requiring organizations that handle sensitive Canadian data to use domestic infrastructure. Such frameworks exist elsewhere in the Canadian regulatory landscape: financial institutions must comply with OSFI guidelines on data residency for certain operations, and healthcare providers in some provinces face restrictions on where patient data can be stored. The AI compute strategy creates no equivalent requirement. Without mandates, adoption is voluntary, and voluntary adoption faces significant headwinds.

Cannot compete with hyperscalers. Amazon Web Services, Microsoft Azure, and Google Cloud offer integrated platforms with hundreds of services, decades of optimization, and global scale that spreads development costs across millions of customers. A new Canadian sovereign compute facility cannot compete on these terms. It would need to build service offerings from scratch, establish customer relationships, and match the reliability of providers with decades of experience. More fundamentally, it would be outspent: Microsoft alone committed $19 billion in Canadian AI investments over four years, from 2023 through 2027 (Microsoft, 2025), and TELUS announced $70 billion in Canadian infrastructure investments through 2029 (TELUS, 2025). The federal government's $2 billion cannot compete at this scale. But it doesn't need to. The $2 billion could solve a specific problem that hyperscalers cannot: CLOUD Act insulation for sectors requiring Canadian ownership. Defence, sensitive government data, healthcare records, and research intellectual property all have legitimate reasons to require domestic infrastructure. A focused investment would identify these sectors, quantify their compute needs, build infrastructure sized to meet that demand, and mandate its use. Instead, the strategy builds general-purpose infrastructure with an undefined scope. This vagueness creates the risk of attempting to compete where Canada has no advantage, while leaving the gap hyperscalers cannot fill unaddressed.

The funding conditions question. Relying on American companies to build or operate Canadian infrastructure is not inherently wrong when domestic expertise falls short. The problem is that public funding should come with conditions that close these gaps over time. The first major award under the AI Compute Challenge went to Cohere, which partnered with CoreWeave, an American company, to build its data centre because no Canadian firm could deliver what Cohere needed (Sali, 2025). Requirements to upskill Canadian workers, transfer knowledge, or develop domestic expertise would help close the gap so that future companies have different options. But the $240 million awarded to Cohere included no published requirements for building Canadian capacity. The gap that made CoreWeave necessary will persist.

Skills gap not addressed. The strategy acknowledges that talent is essential but provides no dedicated funding for education, curriculum reform, or hiring incentives. SCIP asks applicants to "demonstrate a strategy to develop, attract and retain AI talent in Canada" (ISED, 2024b), but this outsources the problem to funding recipients rather than addressing it directly. The "What We Heard Report" from consultations noted that "the gap between academic training and industry needs" was a significant concern, and stakeholders called for investment in "educational programs to develop a strong pipeline of AI talent." The strategy does not respond to this feedback. If a well-funded Canadian AI company cannot find adequate domestic compute options, as happened with Cohere, smaller organizations certainly will not, and the skills gap that prevents Canadian providers from competing will persist.

The subsidy trap. The government subsidizes compute access: free for researchers through the Digital Research Alliance of Canada, and up to two-thirds of costs for small and medium-sized enterprises (SMEs) through the AI Compute Access Fund. But these subsidies are temporary. The Digital Research Alliance received $85 million of funding for 2025 to 2027 (Digital Research Alliance of Canada, 2025b). What happens when the funding ends? If sovereign infrastructure is more expensive than hyperscaler alternatives, or costs the same but offers fewer features and less reliability, organizations will migrate back to AWS and Azure. The success of the strategy actually increases the subsidy burden by attracting more users who will later need to be weaned off. Without a path to cost competitiveness, the strategy creates dependency on ongoing public support with no exit plan.

No long-term commitment. AI hardware has a short lifespan. The most advanced GPUs today will be obsolete within two to three years as new generations offer dramatically improved performance. The $2 billion strategy covers five years. What happens in year six? The private sector makes sustained, long-term investments. The public sector has not matched this commitment. Furthermore, government funding priorities shift with elections. A program launched by one administration may be defunded or restructured by the next. Without legislative protection or cross-party consensus, long-term infrastructure investments remain vulnerable to political cycles. Without a plan for ongoing hardware refresh and stable funding across electoral transitions, the initial infrastructure will become increasingly uncompetitive, potentially recreating the adoption problems the strategy was designed to solve.

Recommendations: Define which sectors and data types require Canadian-owned infrastructure. Mandate its use for those categories through a "licence to operate" framework similar to those in other regulated industries. Attach conditions to public funding that close capability gaps, including requirements for knowledge transfer and workforce development. Invest directly in AI training programs and curriculum reform. Commit to ongoing funding with a ten-year or longer horizon, not a single five-year allocation, and seek mechanisms to insulate this commitment from electoral cycles.

The Gap Between Consultation and Strategy

The timeline of the strategy raises questions about the role consultations played in shaping it.

Budget 2024, announced in April, committed $2 billion to AI compute infrastructure. Consultations ran from June to September 2024. The strategy was announced in December 2024. The scope and funding level were determined before consultations began. This sequencing is not inherently problematic. Governments often consult on implementation details after making high-level commitments. But it does raise the question: what was the purpose of the consultations?

The “What We Heard Report,” published in November 2024, documents stakeholder concerns that extend well beyond infrastructure. Participants called for “establishing a regulatory framework for data management” tailored to Canada’s needs. They highlighted the skills gap and called for investment in educational programs, curriculum alignment with industry needs, and financial incentives to retain talent. They noted concerns about supply chain stability and the risks of depending on foreign hardware. They asked for data classification frameworks that would identify which sectors and data types require domestic infrastructure. Most significantly, stakeholders called for the mandatory adoption mechanisms discussed earlier in this paper, frameworks that would require organizations handling sensitive Canadian data to use domestic infrastructure, similar to requirements in other regulated industries (ISED, 2024d).

The government has not published its rationale for prioritizing infrastructure over the other concerns stakeholders raised. Some of these concerns may lie outside federal jurisdiction. Regulatory frameworks for health data, for example, involve provincial governments. Some may require longer timelines than infrastructure spending. Some may involve political risks that the government chose not to take. But these are inferences. The "What We Heard Report" itself states explicitly that it "does not attempt to interpret respondents' feedback or translate it into policy solutions." This is unusual. Consultation reports typically explain how feedback informed decisions. This one does not. Canadians are left to wonder: Why was infrastructure prioritized? How was the $2 billion figure determined? What role, if any, did stakeholder feedback play in shaping the final strategy?

Recommendations: Publish the analysis that informed the strategy’s scope. Explain why infrastructure was prioritized over regulatory frameworks, talent investment, and data classification. Clarify what role consultation feedback played in shaping decisions. Transparency about the rationale would build confidence that the strategy reflects evidence and deliberation, not simply the path of least resistance.

Aspirations Are Not Outcomes

A well-designed government program establishes measurable success criteria upfront. The Canadian Sovereign AI Compute Strategy does not.

The strategy's vagueness begins with its language. The industry has a term for specialized, GPU-focused cloud providers built specifically for AI workloads: neoclouds (Tairych & Delp, 2025). CoreWeave, the American company building Cohere's data centre, is the world's largest neocloud (Weinberg, 2025). Yet the government's strategy documents never use this terminology, instead relying on phrases like "AI compute" and "AI-specific data centres." This vagueness may reflect a lack of technical fluency, or it may be intentional flexibility. Neither interpretation is reassuring.

The distinction matters because AI infrastructure is not monolithic. Training, the computationally intensive process of teaching models using massive datasets, is dominated by Nvidia. Inference, running trained models to generate outputs, is a different engineering challenge, and companies like Groq built specialized chips for that market before Nvidia's acquisition consolidated the two segments. Power requirements, cooling systems, and network architecture all vary by use case. Saying "AI compute" is like saying "transportation infrastructure" without distinguishing between highways, railways, and airports: it sounds comprehensive while committing to nothing.

If policymakers are blurring these distinctions because they do not understand them, that is a competence problem. If they are blurring them deliberately to avoid choosing priorities, that is a strategy problem. Trying to serve academic researchers, commercial AI companies, national security agencies, and small businesses with the same infrastructure investment risks serving none of them well. Real strategy requires deciding what you will and will not do. The term "AI compute" lets Canada avoid making that choice.

The strategy's objectives are qualitative. SCIP aims to "enhance AI research capacity," "promote innovation and collaboration," and "develop, attract and retain AI talent" (ISED, 2024b). These are aspirations, not metrics. There are no adoption targets, no benchmarks for cost competitiveness, no measures of reduced dependence on foreign infrastructure. How will we know if the investment succeeded? The Digital Research Alliance of Canada currently provides advanced research computing access to over 24,000 researchers (Digital Research Alliance of Canada, 2024a) at more than 100 Canadian universities, colleges, and research institutions (Digital Research Alliance of Canada, 2024b). Should sovereign infrastructure aim to support all of them, a subset, or expand that number? The strategy does not say.

The governance process itself raises questions. The AI Compute Challenge opened for applications on December 5, 2024. Cohere was awarded $240 million on December 6, 2024, one day later (Government of Canada, 2024b). The criteria for evaluating applications have not been published. The alternative proposals considered have not been disclosed. Cohere's partner in building the data centre is CoreWeave, an American company. What criteria determined that $240 million to a company partnered with an American firm was the best use of sovereignty-focused funding? The government has not said.

Without upfront metrics, success will be declared based on outputs rather than outcomes. Dollars spent, facilities opened, and press releases issued will substitute for meaningful evaluation of whether Canada actually reduced its dependence on foreign infrastructure, whether adoption targets were met, or whether the investment delivered value for taxpayers.

Recommendations: Define specific use cases requiring sovereign compute before building more infrastructure. Establish measurable success criteria upfront: target adoption rates, percentage of government and academic workloads on sovereign infrastructure, research output metrics, cost competitiveness benchmarks. Create transparent governance with published evaluation criteria for funding decisions. Report publicly on outcomes, not just expenditures.

Conclusion

The Canadian Sovereign AI Compute Strategy responds to a legitimate concern. Canadian researchers and businesses do depend on foreign-owned cloud infrastructure. That dependence does create risks for data security and intellectual property protection. The instinct to address these risks is sound.

But the response has significant gaps. The strategy’s promise of “sovereignty” overstates what domestic infrastructure can deliver. Building facilities without mandates or regulatory frameworks risks underutilization. Stakeholder feedback was collected but not visibly incorporated. The absence of success criteria prevents meaningful accountability. These are not isolated problems. They reflect a common root cause: the strategy was designed around what the government could easily do (build infrastructure) rather than around what the problem really requires (a comprehensive approach to data security that includes regulatory reform, talent development, and mandatory adoption for sensitive sectors).

The core problem is one of misdirection. The $2 billion could fill a gap that hyperscalers cannot fill: protecting truly sensitive Canadian data from foreign legal jurisdiction. A focused investment would identify which sectors require this protection, mandate use of domestic infrastructure for those categories, and build capacity sized to that demand. Instead, the strategy is at risk of attempting to compete with hyperscalers on general-purpose capabilities, a competition that Canada cannot win given the scale of private-sector investment.

The strategy is not without value. Domestic infrastructure with Canadian ownership and legal jurisdiction does provide intellectual property protection. It does insulate sensitive data from foreign legal demands. For organizations with genuine sovereignty requirements, these protections matter. But the strategy should be evaluated for what it actually delivers, not what its name implies.

The recommendations in this paper are as follows:

Transparency

1. Rename the strategy to reflect its true scope.

2. Publish the rationale for prioritizing infrastructure over regulatory frameworks, talent investment, and data classification.

Scope and Mandates

3. Define which sectors and data types require Canadian-owned infrastructure.

4. Mandate use of domestic infrastructure for those categories through a “licence to operate” framework.

Funding and Sustainability

5. Attach conditions to public funding that close capability gaps, including requirements for knowledge transfer and workforce development.

6. Invest directly in AI talent development.

7. Commit to sustained funding that extends beyond electoral cycles.

Accountability

8. Establish measurable success criteria upfront.

9. Create transparent governance with published evaluation criteria for funding decisions.

10. Report publicly on outcomes, not just expenditures.

References

Bureau of Industry and Security. (2022, October 7). Commerce implements new export controls on advanced computing and semiconductor manufacturing items to the People's Republic of China (PRC) [Press release]. U.S. Department of Commerce. https://www.bis.gov/press-release/commerce-implements-new-export-controls-advanced-computing-semiconductor-manufacturing-items-peoples

Digital Research Alliance of Canada. (2025b). Alliance receives $85M to advance Canada's digital research infrastructure. https://www.alliancecan.ca/en/latest/news/alliance-receives-85m-advance-canadas-digital-research-infrastructure

Faber, D. (2025, December 24). Nvidia buying AI chip startup Groq's assets for about $20 billion in its largest deal on record. CNBC. https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html

Government of Canada. (2024a, December 5). Canada to drive billions in investments to build domestic AI compute capacity at home. Innovation, Science and Economic Development Canada. https://www.canada.ca/en/innovation-science-economic-development/news/2024/12/canada-to-drive-billions-in-investments-to-build-domestic-ai-compute-capacity-at-home.html

Government of Canada. (2024b, December 6). Deputy Prime Minister announces $240 million for Cohere to scale up AI compute capacity [Press release]. Department of Finance Canada. https://www.canada.ca/en/department-finance/news/2024/12/deputy-prime-minister-announces-240-million-for-cohere-to-scale-up-ai-compute-capacity.html

Groq. (2025, May 28). Groq becomes exclusive inference provider for Bell Canada's sovereign AI network. https://groq.com/newsroom/groq-becomes-exclusive-inference-provider-for-bell-canadas-sovereign-ai-network

IEEE Communications Society. (2025, December 17). Sovereign AI infrastructure for telecom companies: implementation and challenges. IEEE ComSoc Technology Blog. https://techblog.comsoc.org/2025/12/17/sovereign-ai-infrastructure-for-telecom-companies-implementation-and-challenges/

ISED. (2024a). Innovation, Science and Economic Development Canada. Canadian Sovereign AI Compute Strategy. https://ised-isde.canada.ca/site/ised/en/canadian-sovereign-ai-compute-strategy

ISED. (2024b). Innovation, Science and Economic Development Canada. AI Sovereign Compute Infrastructure Program (SCIP). https://ised-isde.canada.ca/site/ised/en/ai-sovereign-compute-infrastructure-program

ISED. (2024c). Innovation, Science and Economic Development Canada. AI Compute Challenge. https://ised-isde.canada.ca/site/ised/en/canadian-sovereign-ai-compute-strategy/ai-compute-challenge

ISED. (2024d). Innovation, Science and Economic Development Canada. What We Heard Report: Consultations on AI Compute. https://ised-isde.canada.ca/site/ised/en/what-we-heard-report-consultations-ai-compute

Leswing, K. (2025, November 25). Nvidia says its GPUs are a generation ahead of Google's AI chips. CNBC. https://www.cnbc.com/2025/11/25/nvidia-says-its-gpus-are-a-generation-ahead-of-googles-ai-chips.html

Microsoft. (2025, December 9). Microsoft deepens its commitment to Canada with landmark $19B AI investment. Microsoft On the Issues. https://blogs.microsoft.com/on-the-issues/2025/12/09/microsoft-deepens-its-commitment-to-canada-with-landmark-19b-ai-investment/

MLQ AI. (2025). AI chips and accelerators. https://mlq.ai/research/ai-chips/

Sali, D. (2025, March 20). Ottawa is putting $240M into Cohere's $725M Canadian AI compute plan. The Logic. https://thelogic.co/news/exclusive/cohere-ai-data-center-coreweave/

Shanghai Metal Market. (2025). Chinese Li-ion battery market: Ways to survive under fierce competition. https://www.metal.com/en/newscontent/103286919

TELUS. (2025, May 27). TELUS investing $70 billion in Canada through 2029 to enhance connectivity, support Canadian AI leadership and fuel economic growth [Press release]. https://www.telus.com/en/about/news-and-events/media-releases/telus-investing-7-0-billion-in-canada-through-2029-to-enhance-connectivity-support-canadian-ai-leadership-and-fuel-economic-growth

Tom's Hardware. (2025, December 28). Nvidia's $20 billion Groq IP deal bolsters AI market domination. https://www.tomshardware.com/tech-industry/semiconductors/nvidia-confirms-20-billion-groq-deal-to-bolster-ai-inference-dominance

U.S. Department of Justice. (2019, April). Promoting public safety, privacy, and the rule of law around the world: The purpose and impact of the CLOUD Act [White paper]. https://www.justice.gov/dag/page/file/1152896/download

Tairych, D., & Delp, A. (2025, October 14). What is a neocloud? Equinix Blog. https://blog.equinix.com/blog/2025/10/14/what-is-a-neocloud/

Weinberg, N. (2025, June 25). Neoclouds roll in, challenge hyperscalers for AI workloads. https://www.networkworld.com/article/4011187/neoclouds-roll-in-challenge-hyperscalers-for-ai-workloads.html