Canada's Sovereign AI Intelligence Amplification Blueprint: 2%+ GDP Growth

Key finding: With even modest penetration — roughly 10% of hours meaningfully augmented at a ~20% productivity gain — the compounding impact on GDP to 2030 is equivalent to adding an entire finance sector to the economy, without new federal outlays. The formula: %ΔGDP ≈ share affected × productivity gain = 10% × 20% = 2%.

A Nation-Building Mission for Economic Sovereignty

Foreword

This document is written from a sense of urgency — and of possibility. Canada is facing challenges that are both new and recently imposed upon us. We cannot abandon the long-term perspective, but emerging pressures require results that alter our growth trajectory before existing programs reach their limits. We continue to prioritize research, even though the core dynamics now demand applied solutions at scale. A perfect plan that produces results too late is essentially a poor plan. Therefore, the approach outlined in the following pages does not dismiss ongoing initiatives. However, it is designed to be executable, scalable, and swift — a practical, on-the-ground approach. In this context, AI technologies — from hyperscaler infrastructures and networking to the higher end of the value chain of sovereign data management — represent the most valuable capital allocation Canada can foster, for us and others. What follows is not an ode to technology; it is a roadmap for capital efficiency. For decades, Canada's productivity gap has widened, despite our talent, capital markets, and scientific leadership. Over the past 50 years, our output per hour worked has grown more slowly than every other G7 country. Most recently, we have fallen behind not just relative to others, but in absolute terms as well, with GDP per capita down 2.5% from its peak in the second quarter of 2022. The trade crisis further worsens the outlook. This report argues for a pivot: treat productivity as a nation-building mission and use today's convergence — AI, clean energy, and smarter procurement — to deliver measurable gains in the near term. AI, in particular, should not be framed as "just another tool." It is a workforce amplifier. With even modest penetration (e.g., ~10% of hours meaningfully augmented), the compounding impact on GDP to 2030 could be equivalent to adding an entire finance sector to the economy — without new federal outlays. Why? Because the biggest near-term lift comes from AI inference (deployment and usage), which is far less capital-intensive than training foundational models and is immediately accessible, at scale, across industries and public services. %ΔGDP ≈ share affected × productivity gain = 10%×20% = 2%. Canada holds a structural advantage the world now prizes: clean, reliable energy. Compute capacity tracks electricity. If policy certainty aligns with provincial execution, Canada can become a top-tier AI computing hub — exporting "intelligence amplification" as a service, just as we historically exported resources and power. These deployments can be entirely privately financed when governments create credible pathways — land-use, power allocation frameworks, and fast, outcome-based approvals that reinforce interoperability and competition. There is one precedent we must change: procurement. Our vendor-selection logic was designed for a last-century reality of fixed borders and slow cycles. Today's problems are networked, dynamic, and cross-jurisdictional. We continue to seek a single vendor when the situation calls for an ecosystem. "Buying Canadian" matters — but not if it locks in inferior technology, inflates costs, and postpones impact. Policy should champion platforms and interoperability, not long-term lock-ins that take decades to unwind. The most valuable technology firms on earth succeed by orchestrating ecosystems; Canada's public sector can do the same, setting open standards and purchasing outcomes rather than parts lists.

Ultimately, technology is a vehicle, not the destination. It requires the fuel of capital allocation and velocity — two areas where Canada underperforms. We can fix this by pairing ecosystem procurement with market instruments that accelerate private deployment (capacity auctions, "energy-for-compute" call options, and sovereign co-investment that crowds in pension capital without permanent public balance-sheet exposure).

What you will find in the attached report

• A concise narrative for why productivity is now a security and sovereignty imperative — and why incrementalism will not close the gap.
• A practical focus on AI inference at scale — the fastest, broadest path to economy-wide impact — and how to convert Canada's energy advantage into exportable compute.
• A blueprint for ecosystem-first procurement (outcome-based, multi-vendor, open-standards, anti-lock-in by design) that accelerates adoption across government and industry.
• A short list of policy frameworks the federal government can champion and provinces can operationalize — aligned with Canadian values and designed to be copied by allies.
• An executive grid mapping current problems to proposed solution paths, to guide immediate action and cabinet-level coordination.

Strategic policy recommendations

1. Launch an Energy-to-Compute Pathway with provinces: pre-permitted sites, transparent power blocks, and standardized "revenue-per-watt" contracts that unlock fully private financing for AI inference campuses. Create AI Free Trade Zones with expedited zoning/permitting for purpose-built AI data centers. Create an Energy-Call-Options market aligned with AI growth projections.

2. Align Policy for Optimal Capital Allocation: a Scale-Up Co-Investment Window for strategic late-stage rounds; introduce immediate expensing/accelerated depreciation for AI, software, and data infrastructure; launch an AI Adoption Tax Credit for SMEs; provide financing guarantees, not subsidies, to remove financing hurdles of scaling technology to privilege commercialization in Canada and interoperable IP; and publish a quarterly Capital Velocity Scorecard to track how quickly public policy is converting projects into deployed, productive assets.

3. Announce an Ecosystem Procurement Pilot in two priority domains (e.g., digital service delivery and government optimisation): outcomes-based tenders, mandatory interoperability, multi-phase "pilot → scale" awards.

4. Sovereign Computing Architecture by building a five-province, Tier IV, post-quantum-secure architecture with fragmented storage, high-availability pairing, and lights-out operations. This keeps sensitive data resident and services online while enabling cost-efficient inference at a national scale.

5. Create a transitory Clerk-led Productivity & Sovereignty Taskforce to align federal levers (procurement, skills, data governance) and measure quarterly progress toward adoption and deployment targets.

This is a moment to change the chessboard, not chase someone else's strategy. With the right policy frameworks — focused on ecosystems, velocity, and exportable compute — Canada can convert a decades-long productivity challenge into a national advantage felt in every region and sector.

Introduction

Thesis: Deploy now, measure fast, grow durably. The attached essay is a decision manual, not a think piece. Its core claim is direct: Canada can lift living standards and strategic autonomy by treating productivity as a national-security imperative and by moving AI from pilots to production at scale — while exporting clean, sovereign compute as a new services surplus. Read this document as a blueprint for immediate deployment, capital allocation, and institution-level execution, not as a request for further study.

How to read this document. The essay is structured to take you from problem definition to investable actions. Context Definition sections establish the constraints — productivity gaps, power and compute bottlenecks, procurement friction — and fix the lens: GDP impact above all. They tell you what's broken, how we know, and how to measure improvement. Nation Building Initiatives then translate that context into examples you can fund and govern: concrete deployments, procurement patterns, and infrastructure moves that move output per hour, shorten cycle times, and create an exportable compute industry.

Capital allocation north star: AI is the most efficient asset. Across the plan, treat AI as the highest-leverage destination for scarce dollars because it compounds on three fronts: (1) rapid payback from inference-first deployments embedded in daily work; (2) software efficiency that makes the same silicon do more over time; and (3) durable cash flows from serving models at high utilization. This is not a bet on bleeding-edge research; it is a program to convert existing, mature capabilities into measurable throughput in government and industry — and to monetize Canada's clean-power advantage as a tradable compute export.

What makes Nation Building Initiatives fundable (and safe to scale). Every Nation Building Initiative in this essay is gated by five foundations so leaders can approve with confidence and hold teams accountable:

• Mature technologies only — production-grade components, not lab concepts.
• Deployable with balanced risk/impact — outcome-tied funding, staged rollouts, and clear guardrails.
• Scalable across public and private sectors — a common stack, open interfaces, and sovereignty-by-design.
• High, measurable impact within 24 months — tracked as output/hour, cycle time, and defect rates.
• Net-export potential — a path to sell services (compute, software, integration) abroad to strengthen.

List of recommendations. A summary of all the recommendations in this essay is in Appendix A.

25 Canadian Laws of the AI Economy

The Physics of AI

1. "Compute is the new oil, but electricity is the refinery" — No power, no AI.
2. "GPUs eat CPUs for breakfast" — Yesterday's infrastructure is today's scrap metal.
3. "AI compounds at the speed of software, not hardware" — Models double in capability every 6 months.
4. "Inference costs 10x less than training over lifecycle" — The real money is in deployment of AI, not the development of training models.
5. "Every watt not used is intelligence lost" — Idle compute is economic suicide.

The Economics of Intelligence

6. "AI doesn't sleep, doesn't strike, doesn't retire" — 24/7 productivity at marginal cost.
7. "10% AI adoption = 2% GDP" — Small adoption, massive impact.
8. "The productivity gap is now an intelligence gap" — Fall behind on AI, fall behind forever.
9. "AI is deflationary for tasks, inflationary for capabilities" — Everything gets cheaper AND better.
10. "Winner takes most, second place takes some, third place takes orders" — AI markets are power laws.

The Geopolitics of Compute

11. "Sovereign compute or digital colony — pick one" — Control your AI or someone else will.
12. "Data knows no borders, but servers have addresses" — Physical location still matters.
13. "Trust is the new tariff" — Democratic values are competitive advantages.
14. "The cloud has a flag" — Every bit stored abroad is cash flow surrendered.
15. "Allies share APIs, not just armies" — Digital alliances matter more than military ones.

The Strategy of Speed

16. "The best AI strategy in 2025 beats the perfect one in 2027" — Speed trumps perfection.
17. "Large IT projects are where good ideas go to die" — Deploy or decay.
18. "Infrastructure decisions are 20-year bets made in 20 months" — The window is now.
19. "Every 18-month RFP is a gift to your competitors" — Bureaucracy is the enemy of innovation.
20. "Ecosystems eat vendors for lunch" — Orchestrate, don't control.

The Reality Check

21. "AI isn't coming - it's here and you're already behind" — The future is unevenly distributed.
22. "Your competition isn't local anymore" — A teenager in Estonia can disrupt Toronto.
23. "AI doesn't replace jobs, it replaces tasks" — The work changes, not disappears.
24. "Legacy systems are anchors, not assets" — Technical debt compounds faster than financial debt.
25. "Countries don't fail at AI, they fail at deciding" — Paralysis is the only fatal error.

Five Decisions to Restart Growth — Now

Executive Summary

Canada's productivity problem is no longer an academic curiosity — it is the central constraint on living standards, fiscal room, and strategic autonomy. Labour productivity per hour worked sits far below the United States (about US$74.7 vs. US$97.0 in 2023), and real GDP per capita was lower in 2023 than a decade earlier. Left unaddressed, this trajectory implies stagnant wages, weaker public services, and less national leverage in a more adversarial world.

The solution is not to work harder but to work smarter — accelerating capital deepening and, above all, broad-based adoption of general-purpose technologies like AI. The strategic case is decisive: productivity growth is the only path that reconciles higher incomes with balance-sheet stability, allowing countries to "grow their way" to health rather than inflating or borrowing their way out of trouble.

Canada's window is open but narrowing. Our research strength, clean power, and stable institutions are advantages; our diffusion gap — slow scaling of new technology across traditional sectors — is the liability. The playbook is clear: treat productivity as a national security imperative, align policy to unlock adoption at scale, and build the enabling infrastructure (compute, data, skills) to compound gains. This thesis aligns with the attached briefing's argument that productivity and technological sovereignty are now inseparable, and that Canada must act with the urgency usually reserved for defence.

Context definition

The problem is structural, measurable, and compounding. OECD-based comparisons show a persistent productivity gap with the United States; by 2023 Canada produced roughly one quarter less GDP per hour worked. This is showing up in lived experience: Statistics Canada and major bank research document multi-year declines in real GDP per capita, a stark departure from post-war norms where growth in output per worker paid for rising incomes and public services.

Productivity is now geopolitics. In a world of weaponized supply chains and contested standards, productivity buys resilience. That weak productivity is evolving into a sovereignty risk: countries that cannot efficiently produce strategic goods, services, and digital infrastructure will be rule-takers, not rule-makers. Europe's "tech sovereignty" push is one response; Canada needs its own, adapted to our strengths and constraints.

AI is the near-term lever — if we scale it. Historical productivity waves followed not discovery but diffusion (electrification, computing, the internet). Early evidence suggests AI can unlock double-digit task-level gains today: a large, real-world deployment in customer support increased output per hour by 14%, especially for less-experienced workers; randomized trials and field studies in software development show up to 55% faster task completion. These are not theoretical benefits — they are production-grade deltas that, when compounded across functions, move firm- and sector-level productivity.

But Canada's adoption engine is misfiring. Our economy excels at research and pilots but struggles to scale — especially in public services, health, natural resources, and SMEs where the productivity payoff is largest. Procurement rarely rewards outcomes; capital incentives still favour structures over software and intangibles; and many organizations lack the compute, data governance, and skills to operationalize AI safely. The result: we import scaled solutions and export talent and IP, locking in the productivity gap.

Time matters because the technology curve is steepening. The AI trendline is not linear. Algorithmic efficiency, compute scale, and "unhobbling" (long-context, tools, agents) will convert chatbots into high-autonomy digital coworkers this decade — magnifying the advantage for jurisdictions that have already built the socio-technical plumbing to deploy them. Canada should plan for that world, not yesterday's.

Conclusion and recommendations

Thesis: Canada's most leveraged national project is a productivity acceleration driven by scaled, safe AI adoption and modern capital deepening. Success requires demand signals, enabling infrastructure, and institutional capacity that reward measured outcomes over activity.

Five decision-oriented moves (start now):

1. Launch a "5x500" adoption mission. In 24 months, put five production-grade AI systems (copilot, forecasting, quality, workflow, and knowledge) into 500 large institutions — federal/provincial agencies, hospitals, post-secondary institutions, crown corporations, and anchor firms. Tie funding to measured output/hour improvements and service levels, not PoCs.

2. Procure for productivity. Rewrite major public-sector RFPs so 20–30% of evaluation is realized productivity gain verified in controlled pilots; set aside a portion for Canadian-owned solutions that meet security and outcome thresholds. Use multi-phase procurements (sandbox → scale-up) to collapse the adoption cycle from years to quarters.

3. Build sovereign-grade enablers. Stand up a federated Compute & Data Trust using clean power and allied cloud; provide secure runtime, red-team/assurance services, and standardized data access for health, justice, and natural-resource use-cases. Sovereignty is control, not autarky — design for portability across trusted providers while keeping legal and operational control in Canada.

4. Tilt capital toward intangibles. Introduce accelerated expensing and an investment tax credit for AI/automation software, data engineering, and worker augmentation tools (not just hardware). Condition credits on demonstrated workflow redesign and training hours to prevent "shelfware" and ensure complementarity with labour.

5. Institutionalize "deployment science." Fund cross-sector productivity labs that publish playbooks on redesigning work with AI, drawing on empirical results (e.g., 14% gains in service operations, 20–55% in coding). Mandate transparent measurement (output per hour, defect rates, cycle time) and share pre-approved legal, safety, and change-management patterns to reduce compliance friction.

What success looks like in three years: real GDP per capita growing again; measurable output/hour improvements in health and public administration; a visible pipeline of AI-enabled exports from Canadian firms; and a policy regime that treats productivity as core statecraft, not a side-project. That is the shortest path to higher incomes, fiscal sustainability, and strategic autonomy.

Making Intelligence the Most Capital-Efficient Asset

Executive Summary

Thesis. AI is, right now, the most capital-efficient place to deploy scarce public and private dollars. Three reinforcing facts make this true. First, effective compute — the combination of hardware scale, algorithmic efficiency, and "unhobbling" techniques that turn models into agents — keeps compounding by orders of magnitude (OOMs), compressing years of capability growth into single investment cycles. Aschenbrenner's Situational Awareness shows a repeat of the GPT-2→GPT-4 jump by 2027 based on 3–6 OOMs of additional effective compute, taking us from "smart high-schooler" to PhD-level agents and drop-in remote workers.

Second, the CPU→GPU transition has ended the general-purpose computing era; accelerated computing is now the default for everything that matters (search, recommenders, generative apps), and the installed base is being refreshed accordingly. This is not a marginal upgrade but a full "AI factory" build-out with annual hardware leaps and whole-system co-design. The core economic implication is simple: the same dollar of capex buys far more compute-hours each year, and the platform's usefulness rises even faster because models improve while the hardware sits in your racks. For a contemporaneous public statement of this shift, see NVIDIA's GTC keynote: "general-purpose computing has run out of steam; accelerated computing has reached the tipping point."

Third, inference eclipses training. Once models ship, serving them — often through multi-step, tool-calling, "reasoning" pipelines — dominates lifetime compute. That makes AI a utility-like, high-utilization asset class with recurring cash flows. Even big buyers admit the scale: Meta told investors it would have ~350,000 H100s by year-end 2024, largely to serve AI across billions of user interactions. Internally, industry leaders expect inference to grow by orders of magnitude as agentic systems proliferate — an "inference era" that favors investors who own efficient GPUs, power, and networks.

Bottom line. Money invested in AI compute converts into capability quickly (months, not years), is redeployable across workloads, appreciates in effective output as software improves, and generates recurring revenue via inference. Few other asset classes offer that blend of short payback, compounding utility, and strategic optionality.

Context definition

1) What "capital efficiency" means in AI — payback, reuse, and compounding

Capital efficiency is the ratio of durable capability you gain per dollar and how fast it starts producing cash. On both counts, AI is outrunning alternatives:

• Short payback. OpenAI's run-rate doubled from ~$1B (Aug '23) to ~$2B (Feb '24), with Microsoft reporting billions in incremental AI revenue. A GPT-4-class cluster believed to cost in the low billions can be repaid within the early ramp of product revenue. Capex–revenue curves frame why boardrooms keep green-lighting 10× larger clusters.

• Asset reuse. A GPU hour tomorrow is more valuable than today's, because models, compilers, kernels, and quantization tricks make the same silicon do more work. Aschenbrenner's "counting the OOMs" shows ~0.5 OOMs/year of algorithmic efficiency historically, on top of hardware gains. The net effect: effective compute doubles roughly every 6–8 months even without adding racks.

• Demand durability. Serving models is not a one-off; it's an annuity. Agentic systems call models many times per task and will increasingly "think" longer at inference time (as opposed to training time), driving sustained utilization.

2) The CPU→GPU super-cycle: why general-purpose computing is over

The economic story behind GPUs is straightforward: parallel workloads dominate modern computing (search, ranking, genAI, vision), and GPUs deliver orders-of-magnitude better performance-per-watt-per-dollar for them. Jensen Huang, NVIDIA's CEO, flags a trillion-dollar infrastructure refresh as hyperscalers rebuild around accelerated computing and fast fabrics; co-designed networking and software stacks drove 100,000× performance in a decade and even ~30× jumps generation-to-generation in select workloads. Conversely "general-purpose computing has run out of steam."

For operators, that means new economics of data centers: denser racks, liquid cooling, NVLink/InfiniBand/Spectrum-X, high-throughput storage, substation-scale power — and a faster refresh cadence than the old 3–5-year CPU rhythm. This is why AI factories are being built at GW scale; Microsoft and OpenAI's reported "Stargate" concept is a $100B, multi-GW campus aimed at the next generation of frontier models.

3) Why 2027 looks like a step-change, not a step-function

Progress can be decomposed in the period 2023→2027 to deliver 20x–30x in raw compute scale, and ~10x–30x algorithmic efficiency.

"Agents as coworkers" are not hand-waving. Measured uplifts from scaffolding and tools turn a base model with 5–20% task success into 2–4× higher performance without retraining; that's pure software leverage atop a fixed capex base.

4) Inference eats the world — recurring revenue on top of sunk capex

The industry is already shifting from training models to inference models as the main load. Internal analyses anticipate orders-of-magnitude growth as agentic patterns (multi-call pipelines, tool use, retrieval) become the default UX, and they point to Meta's fleet composition — hundreds of thousands of H100s, with the vast majority earmarked for serving — to illustrate the shape of demand. Meta confirmed the magnitude publicly on its earnings call.

Why it matters for capital efficiency: high utilization shortens payback; predictable load underpins project finance; and every compiler/runtime improvement that reduces tokens-per-task or raises throughput expands effective capacity without fresh capex.

5) National strategies — capital efficiency at the level of states

Peers are racing to lock in compute as a strategic national asset:

• UK. The AI Research Resource (AIRR) and Isambard-AI/Dawn clusters are being scaled; in 2025, the government committed an extra £1B to expand capacity ~20× by 2030. This is an explicit industrial policy to secure national computing.

• Singapore. Budget 2024 allocates >S$1B over five years across compute, talent, and industry — crucially including local access to the latest hardware to shorten domestic innovation cycles. This mirrors the "close the loop between capex and deployment" playbook summarized in our capital-efficiency brief.

• France. The Jean Zay supercomputer quadrupled to ~126 PF; French agencies emphasize growing AI-oriented capacity as a backbone for national research and industry.

• Netherlands. A co-funded "AI factory" in Groningen (EU + national + regional) is pooling capital to deliver shared access for startups, SMEs, and government — lowering per-country cost while raising utilization.

These moves are capital-efficient because pooled investment in shared AI capacity yields immediate productivity gains across sectors and strengthens sovereignty. The Capital Efficiency of AI brief distills the lesson: countries that invest early report more high-skilled jobs, more startups, better public services, and stronger FDI pipelines.

6) The "clean compute exporter" thesis — why power and GPUs are the new LNG

Serving AI is energy-intensive. Jurisdictions with abundant low-carbon power and rule-of-law can monetize clean compute as a tradable service. Canada is a textbook case: >80% of electricity comes from non-emitting sources (hydro + nuclear) in many provinces, with policy aimed at expanding clean generation. The AI opportunity is to export inference and training services from GW-scale, low-carbon campuses — exactly the "AI inference hub / clean compute exporter" strategy described in the CPU→GPU brief. Hyperscalers are already booking long-dated capex in such regions; AWS, for example, announced ~US$17.9B (≈C$24.8B) of cumulative investment in Canadian regions through 2037.

7) Capital formation — why AI infra crowds in private money

Because the cash flows are utility-like and demand visibility is high, AI campuses are financeable with institutional and sovereign capital if governments de-risk interconnects, land, and permitting. The Capital Efficiency of AI paper recommends alliance-based procurement, SPVs with junior public tranches to crowd in pensions/SWFs, and "energy-for-compute" contracts to link new clean generation to AI campuses — mechanisms that maximize private dollars per public dollar.

8) Security reality check — don't leak your edge

With capital concentrating into a few labs and campuses, the security threat surface is enormous. Some nation-states exfiltrate source code and weights with mundane tooling and data harvesting, underscoring why "lock down the labs" is not optional. Capital efficiency collapses if your model weights walk out the door.

Conclusion and recommendations

If capital efficiency is your north star, prioritize AI. The unit economics of accelerated compute, the compounding of software efficiency, and the annuity-like nature of inference make AI infra and adoption the highest-leverage use of dollars in the 2025–2035 window. The question is no longer "whether to invest," but "how fast can we convert capital into durable AI capacity and adoption?"

1. Make compute a first-class balance-sheet asset. Treat GPUs, power contracts, and networking as strategic inventory. Aim for payback <18 months via high-utilization inference services and internal productivity. Tie business cases to tokens served and latency/quality SLAs, not vanity "AI pilots."

2. Unhobble aggressively. Most ROI now hides in agentic scaffolding, long-context workflows, tool use, and post-training. Budget for engineering that converts base models into reliable task-doers; this is where 2–10× productivity pops show up.

3. Concentrate on inference excellence (not training models). Optimize serving stacks first; training is episodic, inference is forever. Build observability down to per-token costs; continuously re-compile and quantize to harvest the software efficiency dividend on sunk silicon.

4. Finance like infrastructure. Use SPVs and project finance against contracted inference demand; pair with long-dated power PPAs. Expect software-driven capacity gains that lower effective depreciation per unit of work over time.

5. Overweight regions with clean, reliable power. They will win the siting race and command premium utilization. Clean-grid jurisdictions are de-risked on policy and ESG; the effect on cost of capital is real.

6. Procure and share national compute. Copy the UK's AIRR: buy clusters for research and SMEs, publish access rules, and plan for 20× scale by 2030.

7. Pool scale with allies. Form a compute-procurement alliance among mid-sized democracies to secure supply, share governance, and interconnect national AI clouds. Use co-funded "AI factory" models to lower per-member cost.

8. Link power policy to AI siting. Create "energy-for-compute" programs that reserve clean GW blocks for AI campuses with open-access obligations. Fast-track interconnects and cooling permits in exchange for transparency and safety commitments.

9. Harden security. Mandate SCIF-like controls for frontier weights and require incident reporting; fund red-team capacity. The cheapest way to lose $10B of capex is a laptop with export-controlled weights.

What to watch

• Hyperscaler siting and power deals. The pace of GW-scale campus announcements (e.g., Microsoft/OpenAI) is a real-time indicator of where value will pool.
• Public capacity programs. Watch the AIRR, EU EuroHPC-adjacent "AI factories," and Singapore's NAIS 2.0; these signal where talent and startups will cluster, which in turn derisks private buildouts.

Decision. Move AI to the top of the capital stack. Reallocate 3–5% of multi-year capex plans to AI infrastructure and adoption immediately; pair with power procurement and security. In a world of constrained balance sheets, this is the rare investment that compounds while you sleep — because the models, they just keep getting better.

Two Engines for a 2% GDP Gain: AI Productivity at Home + AI Compute Exports

What we mean by GDP — and how AI moves it

Thesis. Canada can lift real GDP over the next several years through two reinforcing AI channels:

GDP = C + I + G + (X − M)

(1) domestic productivity — organizations doing more per hour with AI copilots (as per calculation details below), and (2) net exports of compute — selling GPU capacity from Canadian data centres to foreign buyers while global supply remains tight (not accounted for in GDP contribution calculations to provide a more conservative forecast).

1. Domestic productivity

A pragmatic way to size the upside GDP effect as a first-order, mechanical level shift driven by adoption × utilization × task-level productivity for the domestic channel, and by exportable GPU-hours × (price − levelized import content) for the trade channel.

When a subset of workers becomes more productive, GDP moves in proportion to a simple rule of thumb: impact ≈ adoption share × productivity lift × utilization. That is the engine behind the scenario bands: we combine org-level adoption in the private and public sectors (not personal use counts), weight them by their employment shares, multiply by (m−1) where m is the task-level productivity multiple (e.g., 1.14, 1.26, 1.67), and apply u (the fraction of hours actually touched by AI). This yields a first-order, mechanical-level effect on GDP that you can scale up or down as assumptions change.

Canada AI Adoption Grid — a 2.04% ΔGDP lever effect

Using ΔGDP ≈ [ s_priv·a_priv + s_pub·a_pub ] × u × (m − 1):

Adoption band ↓ \ Multiplier →	Low ×1.14	Base ×1.26	High ×1.67
A — "implemented footprint" (a_priv 12.2%, a_pub 22%)	0.80%	1.49%	3.84%
B — "near-term ramp" (a_priv 14.5%, a_pub 38%)	1.10%	2.04%	5.25%
C — "implemented + pilots convert" (a_priv 24.9%, a_pub 54%)	1.75%	3.25%	8.36%

Using organization-level adoption (not personal use), realistic task multipliers (typical ×1.14–×1.26; ×1.67 for well-scoped tasks), and a utilization factor for the share of hours actually augmented, the implied near-term level effect on GDP for Canada spans roughly +0.8% to +8.4%, depending on how quickly implementation footprints and pilots convert and where average task effects settle. These are not one-year growth rates; they phase in over multiple years as tools, data, workflows, skills, and governance mature.

Guardrails for annual paths. The OECD's micro-to-macro work suggests AI could raise aggregate productivity on the order of +0.25–0.6 percentage points per year over a decade as adoption, complements, and utilization ramp. That provides a sensible envelope for converting level effects into growth paths.

Facts on the ground. Recent Canadian data show the private sector is moving, but still early: 12.2% of firms reported using AI to produce/deliver goods or services in Q2-2025, 17.9% planned to adopt AI software within 12 months (Q2), and 14.5% reported plans in Q3-2025. In the public sector, 22% of organizations have implemented AI and 32% are piloting; 48% of public servants use AI at work. Weighting public vs. private employment (≈ 21.6% public) aligns the top-down grid to Canada's structure.

Important caveat on adoption metrics: The adoption metrics vary widely as the measurement standards have yet to be clearly established. For example, some statistics combine personal and professional AI use. Equally impactful is the merging of generative AI (such as chatbots acting as support agents) with industrial Supply Chain Optimization AI systems. These two factors alone tend to overstate the impact of AI deployments on GDP. Therefore, we use a lower figure of 10% to account only for deep-integration AI process conversion.

2. Net Exports of Compute

Net-export upside (X−M). If Canada scales GPU infrastructure faster than peers and sells a slice of that capacity cross-border, those sales count as exports of computer services under EBOPS 2010 ("Telecommunications, computer and information services → Computer services (SI2)"), directly boosting (X−M). With global compute still supply-constrained and major cloud list prices implying single-digit US$ per H100-hour on hyperscaler instances (e.g., AWS p5.48xlarge from about US$55.04/hour for 8×H100; Azure ND H100 v5 around US$98.32/hour; GCP A3 H100 listed regionally), a tranche of exportable GPU-hours could make a visible addition to services exports. The binding constraint is increasingly power and interconnection, not demand.

Why this approach is conservative and decision-useful

• It stays mechanical: no heroics about full general-equilibrium spillovers, or reallocation; those could add upside but are uncertain and slow.
• It uses organization-level adoption metrics (not employee self-reports), which align to firm-level capital/labour decisions that move GDP.
• It separates domestic productivity and trade channels, avoiding double-counting — a common modeling failure.

Conclusion and recommendations

1) Set organization-level adoption targets and measure hours-touched

• What to do. Establish "AI-in-work" KPIs by sector: share of organizations with implemented copilots, share of hours augmented, and average task multiplier by function.
• Why. The rule of thumb depends on a, u, m; manage those inputs deliberately.
• Evidence anchors. Statistics Canada Q2/Q3 2025 adoption baselines; OECD guardrails for plausible growth paths.

2) Make the public sector the throughput proof-point

• What to do. Prioritize copilots in high-document, high-queue domains (benefits, permits, casework), then re-engineer workflows and authority ladders to realize throughput, not just time-saved. Track cases/agent/day and backlog-days.
• Why. Public administration is ~21.6% of employment; visible wins cascade to private complements (vendors, standards).
• Evidence anchors. KPMG (22% implemented; 32% piloting; 48% use), NBER 14% productivity in service agents.

3) Build a national compute-export lane — fast, clean, connected

• What to do. Create a "Compute-to-Export" program that bundles: (i) expedited zoning/permitting for purpose-built AI data centres, (ii) long-term clean-power PPAs (hydro, wind, nuclear uprates), (iii) interconnection prioritization, and (iv) a trade facilitation track to pre-clear cross-border service delivery.
• Why. Global demand is strong; the bottleneck is power + interconnect.
• Evidence anchors. IEA projections of data-centre electricity demand doubling; hyperscaler price bands indicating monetizable GPU-hours; EBOPS Mode-1 treatment ensures exports hit X correctly.

4) Publish a compute-exports satellite account under EBOPS SI2

• What to do. In national accounts releases, break out Computer services (SI2) with a compute-exports memorandum item; adopt a levelized import content (LIC) framework for imported CAPEX/licences so policy can track net exports.
• Why. Clear measurement avoids policy whiplash and double-counting; it clarifies the (X−M) contribution.

5) Use the Digital STRI to prioritize destination markets

• What to do. Aim compute-export commercialization first at markets with low digital-trade restrictiveness (cross-border data flows, localization rules), then expand as agreements improve.
• Why. Market access determines the export share parameter in the GPU-hours equation.

6) Power-aware siting and procurement

• What to do. Tie data-centre siting to grid nodes with surplus low-carbon power; enable zonal pricing and behind-the-meter options (industrial wind/solar + storage; small modular reactors where viable).
• Why. Power is the cap; clean power protects both margins and climate goals.
• Evidence anchors. IEA electricity outlook and AI & energy analysis.

Bottom line. A practical, decision-oriented program that (i) drives organization-level AI adoption/use in high-hours roles, (ii) unlocks public-sector throughput with targeted copilots and workflow redesign, and (iii) accelerates data-centre build-out tied to clean power and cross-border digital market access can shift Canada onto a higher GDP path — without double-counting between productivity and trade channels.

Canada's Next Wave of "AI Factories"

Executive Summary

Canada's AI competitiveness will be determined by a single hard constraint: how much GPU-grade compute can we power, interconnect, and make usable for Canadian researchers, firms, and government. Conventional metrics — square footage, building "full-build" megawatts, or CPU server counts — are now misleading. Measured correctly, only about 11% of Canada's currently installed IT load is GPU-accelerated; roughly 89% remains CPU-only and thus not relevant to modern AI capacity.

Our base today is modest but real. Canada hosts ~239 operating data centres with an estimated ~750 MW installed IT load (not "full-build"), concentrated in Ontario and Quebec. Yet the portion genuinely "AI-ready" (dense power, liquid cooling, high-speed interconnects, and accelerator inventory) is small — again underscoring that floor space ≠ AI capacity. Announced projects (e.g., Bell's AI Fabric targeting up to 500 MW across a BC super-cluster) signal a directional pivot to GPU-first sites, but most of this is still in planning and depends on power access, permitting, and supply chains. At the same time, federal commitments to anchor sovereign capacity through a Cohere-led buildout are a welcome step that must be accelerated and tied to on-shore deployment and access guarantees.

The global context is unforgiving. AI demand, particularly for inference at scale, is pushing data-centre electricity use to ~945 TWh by 2030 (roughly double today), while rack densities and chip thermal design power continue to surge. State-of-the-art AI GPUs moved from ~400 W a few years ago to ~700 W in 2023, with >1 kW parts arriving — driving a rapid shift to liquid cooling and 50–100+ kW per rack designs. Capital is mobilizing accordingly: analysts now project $1.2T in global data-centre capex by 2029, with accelerators the single largest growth driver; individual sovereign and hyperscaler programs are moving from $10B clusters to $100B+ campuses. This is the industrial race that determines who owns the "AI factories" of the 2020s. Canada has the talent and clean power — but we must align on a GPU-first stack to capture value.

The thesis of this paper is simple: treat compute like strategic infrastructure, and manage the entire stack — energy → data centres → accelerators → models → applications — as one decision surface.

Context definition

1) Energy as the first input. AI is a power problem before it is a software problem. The IEA projects that data-centre electricity use will approximately double by 2030 to ~945 TWh, with AI a primary driver; meeting this growth will require rapid grid connections, 24/7 carbon-free supply, and large-block power contracting. In practice, AI-ready campuses must support 50–100+ kW/rack densities and liquid cooling from day one; this elevates site selection (proximity to firm, low-carbon power) from a nice-to-have to a gating factor. Canada's advantage — abundant hydro, cool climates, and emerging nuclear — can translate into durable cost/performance leadership if we prioritize power allocations for GPU campuses with binding service-level timelines.

2) Data centres as "AI factories," not real estate. A CPU-oriented colocation hall is not an AI facility. Purpose-built AI plants combine high-density electrical and mechanical systems, liquid cooling, ultra-low-latency fabrics (InfiniBand/Ethernet with RDMA), and physical layouts optimized for large GPU pods. Global operators are already re-tooling: densification, liquid cooling standardization, and "sovereign AI zones" are now mainstream, not edge cases. Canada's installed base — ~239 sites / ~750 MW — is therefore only a starting point; the binding constraint is AI-fit megawatts and accelerator slots, not shells or "full-build" marketing numbers. Policy should explicitly target energized, AI-fit MW rather than headline capacity.

3) Accelerators (GPUs) as the engine. Modern AI capability is bounded by access to top-tier accelerators and high-bandwidth interconnect. NVIDIA's stack (chips, systems, CUDA and networking) remains dominant; alternatives exist but are not yet substitutable at scale. As Canada's policy community recognized, "compute is the oil of modern AI," and today we host <2% of global AI-grade GPU clusters — far below our talent share. Meanwhile, chip power is rising: 2023-era flagships operate around 700 W, with 2024/25 designs approaching ~1.2 kW — reinforcing the need for liquid cooling, higher-amp distribution, and re-designed white space. The takeaway: count GPUs and interconnect, not buildings.

4) Models and algorithmic velocity. Canada's edge in fundamental AI research is genuine, and companies like Cohere provide a domestic champion at the model layer. The federal investment to anchor a multi-billion-dollar AI data centre and bring sovereign compute online is strategically correct and should be expanded. But research is now capital-intensive: the field advances through orders-of-magnitude (OOM) gains in compute, algorithmic efficiency, and "unhobbling" (e.g., tools, scaffolding). Canada's model layer must sit on top of Canadian accelerators to keep IP, talent, and data governance at home.

5) Applications and adoption — the value capture layer. The stack only pays off when Canadian firms and the public sector absorb AI into workflows. Our SME adoption has lagged; where government leads as a first buyer, diffusion follows. A national playbook — privacy-safe retrieval, domain-tuned models, and opinionated guardrails — should be packaged with compute so organizations can move from pilot to production in weeks, not quarters. Supply-side buildouts must therefore be matched with demand-side enablement (reference solutions, procurement corridors, and hands-on integration support).

Where Canada stands today.

• Installed base vs. AI-fit: ~750 MW installed IT; ~11% GPU-accelerated today — hence the "capacity illusion" if we quote gross MW or floor space.
• Pipeline: AI-oriented projects (e.g., Bell AI Fabric's 500 MW target across six facilities) and allied private investment are accelerating; we should prioritize these projects for energization and grid interconnect.
• Global race: capex will keep compounding — $1.2T by 2029 — with accelerators already one-third of spend; energy demand is tracking to ~945 TWh by 2030, shifting siting to power-rich regions.

Conclusion and recommendations

Decision principle: Treat AI as an industrial stack. Manage constraints from the bottom up (power → facility → accelerator → model → application), with money and policy aimed at the tightest bottleneck: AI-fit GPU megawatts on Canadian soil.

1) Publish the right metric. Adopt a national indicator — AI-Fit MW and GPU Slots — reported quarterly by region. Stop quoting floor space or CPU racks; measure energized, cooled, and networked GPU capacity ready for training and inference.

2) Create a fast-lane for power and interconnect. Establish an "AI-Grid" permitting lane with guaranteed timelines for large blocks (100–500 MW) tied to AI-fit technical specs (liquid cooling, high-density distribution). Prioritize provinces with surplus clean power and rapid build-ability.

3) Stand up a Sovereign Compute Facility (SCF). Back a multi-tenant GPU cloud (public-private) with service-level access for universities, SMEs, and critical public workloads. Tie federal capital to on-shore accelerator placement and open, published access policies.

4) Close the accelerator gap with procurement scale. Pool demand (federal + provinces + anchor enterprises) to pre-buy accelerator cohorts in tranches, hedging supply risk. Require systems to ship with liquid cooling, 100+ kW/rack readiness, and fabric bandwidth consistent with frontier training.

5) Co-site models with compute. Condition federal model-layer grants (foundation and sector-tuned) on training or serving on Canadian GPUs, with residency and privacy guarantees. This keeps IP and data governance at home and reduces egress costs.

6) Accelerate AI Factory projects and comparable campuses. Designate "AI Factory" projects of national significance (e.g., Bell AI Fabric) for expedited approvals and standardized interconnect offers. Target 500–1,000 AI-fit MW energized by 2027.

7) Make government the reference adopter. Create a Crown-wide Gen-AI procurement corridor with pre-approved patterns (RAG, document automation, case triage) and bundled compute credits on the SCF — so agencies can go live in <90 days with compliance baked in.

8) Track energy impact transparently. Publish an AI-energy dashboard that maps AI-fit MW to contracted clean supply (hydro, wind, nuclear, 24/7 CFE PPAs). Align grid planning with IEA-consistent demand paths to avoid local bottlenecks.

9) Build talent where the clusters are. Co-locate workforce programs (liquid-cooling technicians, HPC SREs, AI systems engineers) with AI campuses; fund short micro-credentials tied to employer demand.

Canada has the ingredients — clean power, credible operators, world-class scientists. What we need now is execution on a GPU-first national AI stack: power it, build it, fill it with accelerators, and put Canadian models and workloads on top. Do that, and we will capture — not rent — the next decade of AI value.

GigaWatts Powering Canada's Next Wave of "AI Factories"

Executive Summary

AI will scale as fast as electricity allows. Over the next decade, Canada's growth opportunity is to convert clean, reliable power into "compute" — and then into GDP. The binding constraint is not land or buildings; it is AI-ready megawatts that can cool, interconnect, and run accelerator-dense racks at 25–100 kW per rack and beyond. Provinces that can quickly marshal clean power, transmission, and permitting around GPU-class campuses will capture disproportionate economic value. Provinces that cannot will see their data and AI spending flow elsewhere.

Three facts define the decision space. First, global data-centre electricity demand is set to surge toward ~1,000 TWh by mid-decade, with AI as a primary driver. Jurisdictions with surplus clean power will be magnets for hyperscale investments. Second, within Canada, Québec, Manitoba, British Columbia, and Newfoundland & Labrador retain structural clean-power advantages; Alberta and Ontario can add firm capacity quickly if they align siting, interconnection, and (where needed) gas-with-abatement. Third, leaders in AI are already planning gigawatt-class clusters; national readiness is measured in GPU slots tied to liquid-cooling-capable data halls, not square footage.

Thesis: Canada should treat energy policy as AI industrial policy. The play is to (1) quantify and publish AI-ready MW by province, (2) stand up a rules-based Energy Call Options market to bankabilize clean supply and transmission, (3) concentrate permitting velocity in AI Free-Trade Zones, and (4) anchor trusted, sovereign-grade workloads via AI Data Embassies. These moves, executed now, turn provincial power assets into durable digital exports.

Context Definition

The global hinge: Electricity availability and AI capacity are now tightly coupled. Frontier labs and cloud providers are racing from hundred-megawatt campuses toward gigawatt clusters; power, cooling, and interconnect are the bottlenecks. Expect a decade of industrial mobilization centred on GPUs and electrons — electricity becomes the acutely binding constraint on compute.

Canada's position: Canada already produces a predominantly clean electricity mix and can expand materially by 2035. A national strategy that dedicates a portion of incremental generation to AI workloads could support multi-GW of new data-centre load with high utilization factors and strong local tax bases. StatCan confirms hydro remains the backbone (≈58% of national generation in 2023), underscoring a comparative advantage in low-carbon power that hyperscalers prize.

A province-by-province grid (selected parameters)

What matters: approximate current generation, dominant sources/clean share, surplus/deficit posture, committed additions by ~2030–35, and "AI readiness" signals (policy/projects). Values below synthesize the attached dossiers and authoritative public sources; figures are rounded to the nearest whole number where appropriate.

Province	Approx. annual generation (TWh)	Dominant sources / clean share	Surplus or constraint	Additions/commitments (2030–35)	AI readiness notes
Québec (QC)	~196	≈94% hydro	Historic exporter (30–35 TWh/yr), though 2024 exports were constrained by low inflows; domestic demand rising	+4 GW wind by 2030; exploring new large hydro	Hydro-Québec flags data centres as major new load; abundant legacy hydro + cool climate make QC a prime GPU hub.
Ontario (ON)	~145	~54% nuclear, ~24% hydro, ~8% gas, ~14% wind/solar	Tight in peaks; heavy refurbishments underway	Nuclear refurb life-extension; first SMR (300 MW) targeted ~2029; multi-GW storage/renewables procurements	IESO projects DCs ≈13% of new demand and ~4% of total by 2035; retrofit focus on liquid cooling in GTA.
British Columbia (BC)	~66	≈95% hydro	Traditionally modest surplus	Site C 1,100 MW adding ~5.1 TWh/yr (~2024–25)	Low-carbon tariffing + Site C improve flexibility; off-peak hydro well-matched to steady inference loads.
Alberta (AB)	~52	Gas-dominant; coal now retired	Capacity additions paced by market signals/transmission	Coal retired in 2024; rapid wind/solar build; efficient CCGTs with CCS potential	Open market + gas and wind resources suit purpose-built "bring-your-own-power" AI parks; interconnection queues already swell with DC proposals.
Manitoba (MB)	~38	≈97% hydro	Surplus exports (new hydro online)	Keeyask 695 MW, ~4.4 TWh/yr; enhanced U.S./ON ties	Low-cost hydro + export ties: ideal for clean AI campuses at moderate scale.
Saskatchewan (SK)	~24	Coal & gas heavy (coal retiring by 2030)	Replacement capacity needed	New CCGT, wind/solar, exploring SMR collaboration ~2035	Limited surplus; near-term focus on self-sufficiency; targeted AI loads feasible near Saskatoon with matched new build.
Nova Scotia (NS)	~9	2022 mix: ~44% coal, ~19% gas, ~30% renewables	Must replace coal; imports supplement	80% clean electricity by 2030; wind build + Atlantic interties	For AI parks, anchor on 24/7 CFE contracting + storage; policy is moving quickly toward clean power.
New Brunswick (NB)	~12	Nuclear (Point Lepreau), hydro, gas, wind	Tight; coal exit path required	ARC-100 SMR (~≥100 MW) targeted around 2030; further SMR/firming options assessed	Potential AI/DC clusters near Point Lepreau contingent on SMR schedule and 24/7 CFE portfolios.
Newfoundland & Labrador (NL)	~40	Large hydro (Churchill Falls)	Significant exports to QC; new Labrador options	Churchill Falls MOU with QC; Gull Island concept under joint development	East-coast AI potential hinged to Labrador hydro and future offshore wind/hydrogen ecosystems.

National growth envelope: Canada can plausibly add on the order of 150 TWh by 2035 across hydro uprates/new builds, nuclear (refurbs + SMR), wind, solar, storage, and gas-with-abatement — lifting total generation toward ~700–800 TWh. Dedicating even 20% of new supply (~30–35 TWh/yr) to AI data centres would sustain ~3–4 GW of continuous load — roughly 30 campuses at 100 MW each — without starving household or industrial electrification. This is squarely in line with the speed at which global AI fleets are scaling.

Potential Solutions

1) Measure what matters: publish an "AI-ready MW" inventory

Why: CPU colocation megawatts are not AI capacity. GPU-class clusters saturate power and cooling before floor space; direct-to-chip liquid cooling (DLC) and 400/800 G fabrics are now table stakes. Today, only a subset of Canadian facilities can host 25–100 kW/rack AI halls. Count those megawatts and GPU slots.

Action set:

• National metric: Quarterly inventory of GPU slots and AI-ready MW (liquid-cooling-capable rooms with requisite interconnect) by province and by campus. Tie grid allocations and incentives to growth in AI-ready capacity, not aggregate IT MW.
• Transparency: Require major campuses to report average/peak rack densities and DLC adoption to an independent registry (confidential where needed). Use this to coordinate provincial permitting and utility planning.
• Anchor benchmarks: Highlight existing GPU-ready anchors (e.g., QScale Q01 in Québec) as reference designs for liquid-ready buildouts and heat-reuse integration.

2) Make clean power bankable: Energy Call Options (5–20-year tenors)

Why: Subsidy races are fragile and expensive. An options market converts future clean, deliverable energy blocks (with carbon-intensity floors) into bankable rights that derisk new generation and transmission — without writing cheques. Buyers pay the premium; exercise requires verified deliverables (e.g., DLC, demand-response, heat reuse, local grid upgrades). Provinces retain jurisdiction; a federal clearinghouse standardizes contracts.

Design sketch (decision-oriented):

• Underlying: X MW firm capacity with Y% availability + transmission deliverability + hourly carbon-intensity caps (24/7 CFE tiers).
• Tenors: 5/10/20 years matched to DC depreciation schedules.
• Strike price: Indexed to wholesale benchmarks + regulated T&D; CPI-linked escalators.
• Exercise conditions: DLC readiness, curtailment automation, heat-reuse tie-ins, telemetry. Non-performance → premium forfeiture; options reauctioned.

Outcome: Lower cost of capital for new clean builds; faster intertie upgrades; clear allocation of scarce, clean, deliverable MW to AI and advanced manufacturing without public-funded arms races.

3) Build fast where it's safe to build fast: AI Free-Trade Zones (AIFTZs)

Why: Hyperscalers and model labs optimize for time-to-compute. Canada can outcompete by compressing the timeline from "site selection" to "first compute" — while raising environmental and safety performance.

AIFTZ toolset:

• Single-window permitting with shot-clocks (e.g., 90/180/365 days by project class), pre-zoned industrial land, and modular substation kits for rapid interconnection.
• Hard environmental floors: binding 24/7 CFE shares, dynamic water budgets with recycling/air-side economization, and mandatory heat-reuse feasibility tied to district energy.
• Safety governance: compute-accounting thresholds, pre-deployment model evaluation for high-risk systems, and content provenance.
• Talent & partnerships: fast-track visas for power/datacentre roles; early and meaningful Indigenous partnership, including equity in generation and zone infrastructure.

4) Attract and retain the highest-trust workloads: AI Data Embassies

Why: Sensitive public-sector and regulated-industry AI workloads need trusted compute with clear jurisdiction and lawful access rules. Canada's brand for privacy and rule of law is an export.

What it is: Sovereign-grade compute enclaves on Canadian soil, created by enabling legislation and bespoke intergovernmental agreements. Technical controls include hardware-rooted attestation, confidential computing, strict egress, and continuous compliance monitoring. Tenancy requires adherence to Canadian and international human-rights norms; environmental standards equal or exceed external thresholds.

5) Align power with workloads: Match AI's demand profile to provincial portfolios

Québec / Manitoba / BC / NL: Use hydro flexibility and seasonal surpluses to anchor 24/7 CFE for inference-heavy campuses; prioritize heat-reuse into district systems where climate/urban form allows. (BC's Site C adds ~5.1 TWh/yr; Manitoba's Keeyask ~4.4 TWh/yr.)

Ontario: Protect nuclear refurb windows and de-risk Darlington SMR schedules; steer AI campuses toward zones with transmission headroom and fast-track liquid-ready retrofits in the GTA. Data centres are already 13% of new demand by 2035 — plan distribution stations accordingly.

Alberta: Lean into market speed with purpose-built AI parks that pair high-efficiency gas (with CCS and strict methane performance) to variable renewables and storage; the coal exit is done — enable low-emissions firming and transmission expansions to accommodate ≥100 MW nodes.

Atlantic (NS/NB): Tie AI loads to the coal phase-out trajectory and to new firm resources (e.g., NB's ARC-100) with contractual hourly-matching; leverage the Atlantic Loop and Labrador hydro as those interties mature.

6) Get the accounting right: 24/7 CFE and trusted compute attestation

Adopt hourly guarantees-of-origin so that "clean load growth" is real, not netted on annual averages. Pair with trusted compute attestation standards so operators can verifiably report compute used for training and inference — now prerequisites for safety governance, export controls, and high-assurance customers.

Conclusion and Recommendations

The decision: Canada has the electricity, climate, and institutions to become a first-tier AI compute hub. But capacity must be AI-ready, and policy must move at the speed of build. The cost of hesitation is straightforward: hyperscale campuses — and their construction dollars, jobs, and downstream digital services — route to faster jurisdictions. The cost of action is measured and bankable: standardized options markets, transparent capacity metrics, and zone-based permitting that rewards environmental excellence.

Near-term (0–6 months):

1. Publish the baseline. Direct NRCan and provincial utilities to produce a public inventory of AI-ready MW and GPU slots by campus, including rack-density limits and DLC readiness. Use this as the capacity ledger for allocations and incentives.

2. Stand up pilot Energy Call Option auctions in provinces with headroom (QC/MB/BC/NL), with 5- and 10-year tranches and 24/7 CFE floors. Reserve a portion for AI campuses that meet heat-reuse and response-time standards.

3. Designate two AIFTZ pilots: one brownfield (retrofit with liquid loops) and one greenfield megasite (pre-zoned + modular substation). Set binding permit shot-clocks and publish interconnection queues and tariffs.

4. Secure anchor loads and sovereign use-cases. Conclude MOUs for the first AI Data Embassy tenants (allied governments/multilaterals). In parallel, use challenge-based procurement to deploy AI across public services — an essential domestic-demand signal that keeps value onshore.

5. De-risk firm supply schedules. Where SMRs or hydro uprates are on the path-critical path (ON/NB/NL), align transmission and campus timelines; in AB/SK, codify gas-with-abatement as eligible firming under 24/7 CFE accounting.

6. Scale the option market to 20-year tenors that anchor long-lead firm assets (hydro uprates, nuclear refurb, long-duration storage). Use premium revenue and interconnection fees to finance local grid upgrades.

7. Institutionalize measurement. Make AI-ready MW and 24/7 CFE shares part of federal-provincial scorecards; tie tax treatment to verified hourly matching and heat-reuse performance.

Bottom line: Count accelerators, not buildings. Allocate clean, deliverable power with discipline. Permit liquid-ready campuses at speed in the right places. Do this, and Canada turns electrons into intelligence — and intelligence into growth.

Out-Deploy, Not Out-Spend: Allocating Capital for Sovereign AI Productivity

Executive Summary

Canada's growth problem is not a shortage of ideas; it is a capital allocation problem. We fund research generously, but we fail to route money, talent, and infrastructure quickly enough into products, revenues, and scale. The fix is not another grant program. It is a re-wiring of how Canada deploys public dollars to crowd in private investment and compress the cycle from appropriation to impact — focused on three national assets we can scale now:

(1) inference-first AI deployment to lift productivity in service-heavy sectors;

(2) a sovereign, resilient compute fabric — multi–data-center, post-quantum secure — to run sensitive workloads at home; and

(3) capital velocity instruments that turn policy intent into booked revenue, sooner.

The diagnosis is established. Canada's innovation system skews toward upstream research (HEI dominance, generous SR&ED) but underperforms in business R&D, late-stage financing, anchor firm creation, and innovation-friendly procurement. Fragmented programs slow founders and diffuse accountability; early exits and foreign-led rounds export IP and management scale. The outcome is the familiar paradox: strong science, weak productivity.

The strategy proposed is deliberately decision-oriented:

• Deploy AI where GDP lives. Focus scarce public risk capital on inference-first tools at the point of work — claims, permits, case prep, coding, compliance — measured by hours saved, cycle-time reductions, and defect rates, not by research papers. A 12-month national program can produce tangible service-level gains and a base-case ~2% lift in real GDP if only 5% of workers realize a 40% productivity improvement on their task mix.
• Build a sovereign compute backbone. A five-province, Tier IV, post-quantum-secure architecture — fragmented storage, HA pairing, lights-out operations — keeps sensitive data resident and services online, while enabling cost-efficient inference at national scale.
• Increase capital velocity. Institutionalize a 90-day innovation procurement fast lane; stand up an automatic 40% public co-investment for qualified growth rounds; convert SR&ED into real-time cashflow; launch a frontier tech visa and friendshored R&D grants; and operationalize a senior-lender guarantee pool to unlock credit at 5–10x leverage.
• Secure the network with PQC at scale. Standardize and certify post-quantum cryptography (PQC) across ministries and critical infrastructure; use CANARIE and QEYSSat-linked pilots to validate throughput, latency, and failover; and convene allies on interoperable "sovereign PQC stacks." (See also NIST's finalized FIPS 203/204/205 and Canada's federal PQC migration roadmap.)

The next sections define context, present concrete instruments, and end with sequenced recommendations. The tone is assertive because the window is narrow: the world is adding zeros to AI infrastructure each year, and sovereignty is increasingly an engineering and capital allocation problem. The "counting the OOMs" mindset — more compute, more algorithmic efficiency, more un-hobbling of agentic systems — explains the speed and scale at which this is unfolding and why Canada must choose where to lead and where to buy the falling cost curve.

Context definition

Canada's innovation paradox is now a growth constraint. We excel at knowledge creation but trail peers in converting invention into industrial impact. Business R&D intensity (~1.1% of GDP) lags while public support (SR&ED, grants, HEI spending) is generous and fragmented. The result: valuation and talent leakage via early exits, foreign-led growth rounds, and limited anchor firms to acquire, mentor, and buy from domestic startups. Public procurement rarely plays "first customer." The empirical through-line in the attached analysis is clear: we underweight commercialization infrastructure and scale-up capital relative to upstream research support.

AI shifts the payoff frontier — but only if we allocate to deployment. Canada does not need to win the race to train the largest frontier models to win the GDP race. The fastest route to measurable output is inference-first deployment: copilots and agents embedded in workflows (writing, analysis, code, case work, adjudication). Training is a one-off compute spike; at scale, inference dominates the lifecycle compute bill (often 40–70%), especially for test-time reasoning. This is where Canada's service economy can harvest productivity in 12–24 months — if procurement pays for outcomes and if a sovereign control plane makes adoption safe by default.

Sovereignty requires a national compute fabric. Sensitive workloads — health, justice, defence, citizen data — need Canadian residency, Canadian keys, and quantum-resilient cryptography. A five-region, Tier IV architecture with high-availability pairing, asynchronous geo-recovery, encrypted fragmentation, and lights-out operations is technically available and aligns with Government of Canada security baselines. Post-quantum cryptography (PQC) and data sharding eliminate single points of compromise and deny "harvest now, decrypt later" adversaries. The architecture marries private data centers with hyperscaler technology under Canadian control.

Quantum security is a near-term deployment opportunity. The PQC standards are finalized (FIPS 203 ML-KEM, FIPS 204 ML-DSA, FIPS 205 SLH-DSA), and the Canadian Centre for Cyber Security (CCCS) has published a federal migration roadmap. Canada's research and education backbone (CANARIE) is already upgrading to 400 Gbps capacity and working with allied R&E networks to provision 400-Gbps transoceanic circuits — ideal testbeds to validate sovereign, certified PQC stacks before production rollouts across ministries and critical infrastructure.

Policy alignment is emerging but incomplete. The Canada Innovation Corporation (CIC) blueprint signals a shift to outcome-driven commercialization, with NRC-IRAP transitioning into CIC over time. That creates a vehicle to coordinate procurement, co-investment, and SR&ED cash-flow reform around commercialization milestones and IP retention. But instruments and governance must be designed to accelerate capital, not just announce it.

Potential solutions

1) Deploy an Inference-First National Productivity Stack (12 months to impact)

Thesis. Allocate capital where it turns into GDP fastest: inference at the point of work. Standardize a sovereign stack; pay for cycle-time reductions, defects avoided, and hours returned to service. Avoid vendor lock-in with a brokered model layer; insist on Canadian data residency, customer-managed keys, and policy-as-code.

Why it works. Inference-time reasoning dominates lifecycle compute as usage scales; deployment is capital-light relative to frontier training clusters; and service-sector GDP is where Canada can move the needle quickly. Buy the falling cost curve, don't finance it.

2) Build a Sovereign Multi–Data-Center Architecture (defence-grade, quantum-secure)

Thesis. Sensitive public and critical-infrastructure workloads require a home-field advantage — tier-IV resiliency, Canadian jurisdiction, and PQC by default — without sacrificing throughput or latency.

Design (from the architecture paper).

• Five-province footprint (BC, AB, ON, QC, NS) for geo-diversity and sovereignty; sites paired for synchronous HA within region and backed up asynchronously cross-region for disaster recovery.
• Tier IV facilities with independent power/cooling, dual power feeds, fault-tolerant network links; "lights-out" automation to reduce insider risk.
• Fragmented storage (erasure coding/secret sharing) across sites so no single location holds reconstructable data; N-of-M retrieval ensures continuity under outage.
• Zero-trust networking over dedicated, encrypted, diverse fiber paths; SDN for instant rerouting.
• Hyperscaler tech under Canadian control, e.g., air-gapped cloud platforms deployed on Canadian soil with Canadian administrators and keys.

Security baseline (PQC now). Adopt NIST's PQC suite (FIPS 203/204/205) for KEM and signatures; align with CCCS's federal PQC migration roadmap. Validate performance and crypto-agility on CANARIE (400-Gbps backbone) and allied R&E circuits before full production.

Why it works. The architecture eliminates single points of compromise and assures continuity under component, site, or regional failure. It is purpose-built for sovereign inference at scale and quantum-resilient communications, and it positions Canadian vendors in export markets for certified PQC modules.

3) Install Capital Velocity Instruments (from policy to purchase orders)

Thesis. Canada's issue is not just how much we spend; it's how fast dollars turn into deployed solutions, revenue, and reinvestable returns. Five instruments — implemented together — move the whole pipeline.

1. 90-Day Innovation Procurement Fast Lane (IPFL). A standardized, challenge-based federal procurement track that moves from problem to pilot in ≤90 days, with milestone payments, pre-negotiated IP/data terms, and cloud/AI safety guardrails embedded. KPIs: median award cycle ≤90 days, time to first payment ≤30 days post-milestone, ≥25% pilots to scaled adoption, ≥1.5× private dollars per public dollar. Why it matters: converts appropriations into contracts, produces reference customers fast, and shortens sales cycles that typically starve startups.

2. Automatic 40% Public Co-Investment ("Scale-Up Speed Fund"). A rules-based facility, managed by CIC/BDC, that automatically matches up to 40% of qualified Series A–C rounds at the lead investor's terms (pari passu), with caps and anti-flip provisions tied to Canadian HQ and IP. Why it matters: accelerates closings, reduces time fundraising, crowds in larger private rounds, and strengthens Canadian lead investors. KPIs: round close time, private/public leverage, % with Canadian leads, % reaching next stage in 12–18 months.

3. Growth Equity Tax Accelerator (GETA). Time-bound tax incentives to direct domestic capital (individuals and corporates) into Canadian growth funds and rounds: capital gains deferral/partial exclusion, loss pass-through, reinvestment windows, and transparency via a public registry of qualifying funds. Why it matters: pulls pensions/corporates/household savings into tech, shortens fund raises, and thickens local syndicates. KPIs: domestic share of LP commitments, time to fund close, Canadian vs. foreign leads, number/size of later-stage rounds.

4. Real-Time SR&ED ("File-and-Flow") + AI/Compute Credits. Convert SR&ED from an annual tax event into monthly advances against verified R&D payroll and eligible cloud/compute spend, reconciled at year-end with telemetry-based audits; layer a capped, tradable compute sub-credit with safety guardrails. Why it matters: collapses working-capital gaps, reduces reliance on SR&ED factoring, and shifts spend into product, not interest. KPIs: days from R&D spend to reimbursement (≤30), R&D headcount growth, product release cadence.

5. Frontier Talent Visa + Friendshored Allied R&D Grants. A 10-day visa for AI/quantum/cyber/semiconductor talent and founders, plus matching grants when allied firms co-site R&D with Canadian startups. Why it matters: compresses team-formation and lab-stand-up timelines; imports know-how and allied customers. KPIs: time-to-hire, number of allied labs, private co-investment per grant dollar.

Why it works together. IPFL pulls prototypes into pilots; File-and-Flow stabilizes cash; the Speed Fund closes rounds; GETA deepens domestic capital; and the visa/friendshoring instrument supplies people and equipment. Sequenced inside regional commercialization hubs, these instruments reduce friction at every hand-off.

4) Crowd-in private credit via a Senior-Lender Government-Backed Funding Pool

Thesis. Many of Canada's best SMEs are "near-bankable": strong demand and IP, thin collateral. A government guarantee to the senior lender, priced for expected loss, unlocks private credit at 5–10× the public reserve and recycles the same dollar across multiple loans.

Mechanics

• Partial coverage (50–80%) of net loss after recoveries; step-downs over time (e.g., 80% year 1, 60% year 2) to encourage performance monitoring and refinancing.
• Per-loan and portfolio caps; sector, regional, and vintage diversification thresholds; lender accreditation and common underwriting/servicing standards.
• Risk-based fees paid by lenders; light-touch covenants aligned to milestones, cash burn, and revenue — not fixed collateral tests.
• Transparent dashboards on leverage achieved, time-to-cash, survival/growth vs. matched controls, default and recovery, and administrative cost per $ mobilized.

Why it matters. Unlike grants, guarantees are contingent; most loans perform, so the reserve recycles. The product is fast (30–45 days from application to funding), non-dilutive, and complementary with equity and procurement-led growth — bridging receivables and regulatory milestones that make or break early commercialization.

5) Prioritize Quantum-Secure Communications over speculative hardware bets

Thesis. Public risk capital should target deployable quantum-secure communications — PQC everywhere, QKD/fragmentation for selected high-assurance links — not speculative quantum computing hardware. Standards, guidance, and networks are ready; deployments cut risk now and position Canadian vendors for export.

Policy moves.

• Mandate PQC (FIPS 203/204/205) for federal systems with crypto-agility and FIPS 140-3 module validation pathways.
• Publish departmental PQC roadmaps aligned to CCCS milestones; train "crypto-change agents" via the Cyber Centre's learning hub.
• Use CANARIE and provincial R&E networks to validate throughput, latency, failover, and mutual recognition profiles with Five Eyes/EU/NATO partners; bridge to early space-to-ground pilots (e.g., QEYSSat) for extended reach.
• Codify "sovereign and certified PQC stacks." Agree suites, key lifetimes, certificate policies, and validation processes that partners can mutually recognize, unlocking cross-border digital trade and defence interoperability.

Why it works. The threat is "harvest now, decrypt later." PQC reduces current exposure, integrates into existing TLS/VPN/PKI, and avoids forklift upgrades. DOE's Quantum Internet Blueprint signals allied momentum that Canada can help operationalize through certification leadership.

Conclusion and recommendations

Canada's lever is allocation velocity. We cannot outspend superpowers on trillion-dollar training clusters, but we can out-deploy them where it matters for our economy: inference-first productivity, sovereign compute, and quantum-secure communications — underwritten by procurement, co-investment, and credit guarantees that move capital now.

Twelve-month plan (sequenced for impact)

1. Stand up the sovereign control plane + IPFL.

   • Approve a catalogue of AI services/templates; embed privacy, security, and PQC guardrails into the framework.
   • Launch ten national workflows across health, benefits, inspections, case preparation, grant adjudication, and RFP drafting; baseline cycle-times and error rates.

2. Launch the Scale-Up Speed Fund and File-and-Flow SR&ED.

   • Publish eligibility rules; synchronize legal close with lead investors; report leverage and time-to-close.
   • Start monthly SR&ED advances against verified payroll/compute; instrument telemetry-based audits.

3. Operationalize the Senior-Lender Guarantee Pool.

   • Capitalize the reserve; accredit an initial cohort of lenders; pilot three streams (IP-rich software, advanced manufacturing, clean tech); target 30–45 days application-to-funding.

4. Harden the backbone with PQC and begin staged migrations.

   • Issue departmental PQC migration plans aligned to CCCS; prioritize high-value inter-ministerial links; test on CANARIE 400-Gbps segments before production.

Twenty-four months and beyond

• Expand the workflow catalogue to 100+ deployments across provinces and priority industries; publish quarterly dashboards of hours saved, dollars avoided, and service-level gains.
• Commission two HA-paired Tier IV sites and start encrypted fragmentation of priority datasets; move critical services to the sovereign fabric.
• Institutionalize GETA and the frontier visa with reciprocity pilots (UK/Australia/EU); measure domestic LP shares and time-to-hire as primary KPIs.

Governance. Task CIC to own commercialization outcomes — not just disbursements — by co-chairing (with PSPC/TBS/CRA/IRCC) a delivery board for IPFL, Speed Fund, File-and-Flow, visa/friendshoring, and the guarantee pool. Standardize metrics: award cycle time, time-to-cash, private leverage, domestic LP share, scale-up survival, and GDP-adjacent productivity indicators by workflow.

Bottom line. Canada's comparative advantage is governed deployment at scale. Allocate to inference-first productivity, anchor it on a sovereign, quantum-secure compute fabric, and accelerate dollars through fast procurement, automatic co-investment, real-time SR&ED, and risk-sharing credit. Do this, and we convert ideas into domestic productivity, revenues, and compounding returns — fast enough to matter.

Orchestrate, Don't Megabuy: A Control-First Playbook for Government Procurement

Executive Summary

Thesis. Canada should pivot federal procurement from compliance theatre to a control-first, capability-maximizing model that buys properties (verifiable control, portability, attestable compliance) rather than brands or postal codes. That pivot reconciles two imperatives:

(1) enforceable sovereignty over data, operations, and exit; and

(2) access to frontier capability in cloud, AI, and security.

The path is practical: define sovereignty as evidence of control (keys, policy, attestation, exit), reform procurement to orchestrate a modular AI and data ecosystem, and use tiered guardrails and challenge clauses to keep vendors contestable without sacrificing performance.

Why now. Cybersecurity and sovereignty are not synonyms. Cybersecurity mitigates adversarial risk; sovereignty prevents external control and legal compulsion from trumping Canadian decision rights. That distinction matters as AI, extraterritorial law, and supply-chain concentration compound dependencies faster than governance has adjusted. Procurement is the lever we control.

What good looks like. Use a multi-level sovereignty continuum (from data residency to full sovereign clouds) and a multi-axis blueprint (legal/operational/technological/economic + data sensitivity + governance maturity) to right-size controls per workload. Then buy to those properties with open interfaces, attestation-bound key release, non-bypassable admin controls, and credible exit.

The cost of inaction. Two-thirds of ~7,500 federal applications are in poor health; every year of monolithic, lock-in-prone buying deepens negative compounding and widens the capability gap. AI's trajectory compresses timelines further; failing to reform procurement now bakes in a capability tax across services, security, and the economy.

Context definition

Procurement is Canada's sovereignty lever. Federal debates often overweight vendor nationality and where data sits; they underweight who controls decryption keys, what policies are enforced in real time, whether integrity is attestable, and whether we can walk away. Geography can complement control, but it does not define it. A control-first posture operationalizes sovereignty through four enforceable pillars: keys, policy, attestation, and exit.

Cybersecurity ≠ sovereignty. Classic cyber programs harden systems against threats. Sovereignty asks who can compel access and under which laws — and whether providers can act without Canada's explicit approval. The operating model changes: customer-owned root keys; non-bypassable split-key or external key management; challenge-and-notify obligations against foreign legal demands; and an independent kill-switch to render data cryptographically inert.

Canada already has policy rails. The Policy on Service and Digital sets digital authority and can anchor platform-level orchestration; the Directive on Automated Decision-Making (ADM) and its Algorithmic Impact Assessment (AIA) provide risk-tiering and disclosure for AI-enabled systems; GC API Standards and guidance embed openness and portability; and the GC Cloud Guardrails and Cyber Centre advice provide a multicloud security baseline. These instruments should be tied directly to procurement clauses and acceptance tests.

The estate we must integrate. The Auditor General reported ~7,500 applications, with about two-thirds in poor health, including hundreds essential to Canadians' safety and economic well-being. AI must work with this legacy, not after a perfect modernization program that will not arrive. That requires adapters, open APIs, and modular contracting, not megaproject RFPs that lock choices for a decade.

A risk-based sovereignty continuum. Not every workload needs maximum control. A 5-level model ranges from simple residency (L1–L2) to legal/operational sovereignty (L3), technological sovereignty (L4), and full sovereign cloud (L5). Overlay data sensitivity (public → personal → critical infrastructure → defence/AI models) and governance maturity (manual audit → continuous, AI-assisted conformance). Result: right-size controls and specify them in contracts.

Urgency is real. Frontier compute, models, and tooling are compounding; organizations with access to state-of-the-art capability will separate quickly from those without. Policy needs to assume rapid step-changes, not gradualism — and anchor in contestable platforms rather than long-lived bespoke stacks.

Potential solutions

1) Redefine sovereignty in operational terms — and require Evidence-of-Control

Codify in policy and procurement that sovereignty = enforceable control over access, processing, and movement of government data, evidenced continuously by cryptographic, operational, and legal mechanisms under Canadian jurisdiction. Make evidence — not assurances — the basis of authority to operate: (a) Crown or Canadian-custodian key custody; (b) policy-as-code gating all privileged actions; (c) remote attestation binding key release to platform integrity; and (d) tested exit paths. Score bids and pay milestones only when evidence is produced.

2) Stand up a Government of Canada Key Custody Authority

Operate certified HSMs in Canada; generate, rotate, and destroy master keys under split control; expose external key management interfaces to cloud/SaaS so providers cannot decrypt without Canadian approval. Log all cryptographic operations immutably; require attestation-bound key release (no valid attestation, no keys). Build PQC-readiness (inventory, hybrid modes, migration plans) into the authority to future-proof against quantum threats.

3) Mandate confidential computing and attestation for sensitive workloads

Process designated data classes only in Trusted Execution Environments (TEEs) with remote attestation; allowlist acceptable measurements; bind decryption key release to verified hardware/firmware state and workload identity. Where TEEs are not yet viable, enforce compensating controls (threshold crypto, minimization, narrow time-bounded plaintext exposure).

4) Control-plane separation and Canadian oversight of privileged operations

Keep identity, keys, policy engines, and session brokering outside the provider domain you aim to constrain. Require just-in-time elevation, multi-party approval for any action that could expose data or alter control planes, and recorded admin sessions brokered through Canadian-controlled systems. Contractually prohibit out-of-band access paths; require disclosure of all administrative tooling.

5) Treat residency as complementary, not defining

Continue to favour in-Canada residency and domestic routing for latency and jurisdictional benefits, but make clear in procurement that residency is one layer of a multilayer control stack — secondary to provable control of keys, policy, and attestation.

6) Procure a GC AI & Data Platform Layer; separate platform from apps

Act as ecosystem orchestrator, not solution builder. Fund and operate a common platform — identity and policy enforcement; model gateway; retrieval and data contracts; evaluation/red-team harness; observability — behind open interfaces. Then buy applications as plugins that must meet interface, security, and portability tests. Tie all AI procurements above a risk threshold to the ADM Directive and deliverables, published by default.

7) Move to modular contracting with objective acceptance tests

Replace big-bang, design-prescriptive RFPs with short, iterative task orders. Compete capabilities (e.g., retrieval layer, evaluation tools, adapters) against common test suites aligned to NIST AI RMF and ISO/IEC 42001 practices. Use the U.S. FAR 39.103 modular-contracting regime as a policy reference for outcome-based, low-lock-in buys; adapt to Canadian instruments.

8) Mandate open interfaces and portability in every award

Insert standard "no-lock-in" clauses:

• Data/metadata export (including embeddings, vector indexes, evaluation traces) in open formats;
• Documented API conformance to GC standards and change-log obligations;
• Model portability where feasible (e.g., ONNX/MLflow export, documented prompts/templates, weight escrow for fine-tunes). Anchor these to GC API Standards/Guidance and, for cross-government alignment, to the European Interoperability Framework principles on openness and reusability.

9) Institute spend controls for platform consistency

Borrow the UK model: require central approval for digital/AI spend that duplicates platform capabilities, and publish an annual "risk and importance" portfolio to steer reuse and contestability. Grant earned autonomy to mature teams that demonstrably meet policy-as-code, security, and portability baselines.

10) Launch innovation sandboxes with dual-use pilots and a production on-ramp

Stand up cross-sector sandboxes (e.g., predictive maintenance, workflow automation, smart infrastructure) under a fixed governance playbook. Require portability to the GC platform as the graduation criterion; publish reusable adapters and lessons learned. Use these sandboxes to pay for integration, not demos, and to build domestic capacity across defence, finance, healthcare, and municipalities.

11) Adopt a tiered guardrail regime mapped to sovereignty levels and data sensitivity

Publish a classification + control mapping: from public info (L1–L2) to sensitive personal data (L3–L4) to national-security and strategic-model workloads (L5). Tie bidding eligibility to level-appropriate controls (e.g., L3 requires in-country operations, domestic key custody; L4 adds tech/IP control; L5 mandates full sovereign cloud). Build continuous conformance monitoring into the platform.

12) Make exit and reversibility real

Require termination assistance, export of data/configurations/logs/model artifacts in open formats, and annual exit drills for designated systems. Price switching explicitly in contracts and score vendors on time-to-exit in evaluations.

Conclusion and recommendations

Bottom line. Canada does not need to choose between shrinking the supplier base in the name of sovereignty and outsourcing control in the name of capability. With evidence-of-control and procurement-led orchestration, we can have both: capability today, sovereignty always.

90-day actions

1. Issue a control-first mandate. Treasury Board and PSPC define sovereignty as Evidence-of-Control and publish the minimum proof objects for keys, policy, attestation, and exit; make them gating criteria for awards.

2. Publish v1.0 of the GC AI/Data Platform specs. Interfaces, data-contract templates, evaluation protocols, logging schemas, and Zero Trust & Guardrails alignment checklists.

3. Stand up the Key Custody Authority blueprint with PQC-readiness; begin vendor integrations via external key management.

12-month milestones

4. Run five modular competitions for platform capabilities (retrieval layer, evaluation, adapters) with objective acceptance tests aligned to NIST AI RMF/ISO 42001; award on performance and portability.

5. Bind key release to attestation for all new sensitive workloads; publish quarterly conformance reports.

6. Launch two cross-sector sandboxes (e.g., workflow automation and predictive maintenance) with a production on-ramp contingent on meeting platform interfaces.

24-month outcomes

7. Majority of Tier-1 data under customer-owned keys with attestation-gated use; all AI procurements above risk threshold publish AIA/model cards and evaluation traces.

8. Spend controls prevent duplication of platform capabilities; departments gain earned autonomy by meeting portability/security baselines.

9. Workload-to-level mapping completed; L5 sovereign cloud instantiated for designated national-security and strategic-model use cases.

KPIs to track

• % of high-sensitivity workloads with attestation-gated key release;
• % of privileged actions executed via Canadian-controlled brokers with multi-party approval;
• Time-to-exit (tested);
• % of AI awards meeting API/portability clauses;
• % of Tier-1 applications integrated to the GC platform;
• Reduction in duplicated stacks under spend controls.

Delivering these measures turns procurement from a paperwork exercise into the operating system of sovereignty — one that keeps Canada in control while compounding capability across the state and economy.

Sovereignty: Secure by Design, Federated by Treaties

Executive Summary — We live in a connected world

The nation-state's power once hinged on customs posts and geography. Today, it hinges on who controls compute, data, cryptographic keys, and the operational runbooks that keep systems alive under duress. Digital sovereignty is not a slogan; it is a multi-axis program of legal, technological, operational, and economic control over digital assets — and it must be tiered by risk. That means treating critical infrastructure and defense as Level 4–5 sovereign workloads (legal, operational and technological control end-to-end), while allowing lower-risk public services to run at Level 1–2 with residency and basic controls.

Urgency is growing. AI, data, and cloud capacity are compounding at breathtaking pace; every six months, budgets and clusters add another zero, and national security stakes escalate accordingly. In short: do not expect the curve to flatten. A sovereignty posture designed for yesterday's internet will be overrun by tomorrow's AI.

The strategic answer is cooperative sovereignty: build national "sovereignty stacks" (law + keys + cloud + operations) at home, and federate them abroad through multilateral agreements that preserve control while enabling joint operations, cyber defense, and cross-border commerce. This paper frames the context and closes with decision-oriented recommendations — covering critical-infrastructure security, defense, multilateral sovereignty stacks, post-quantum communications across borders, and federated surveillance/monitoring and data-sharing.

Context — We need multilateral agreements and data sharing without surrendering control

Sovereignty is layered, not binary. A credible strategy spans four dimensions — jurisdictional (whose law applies), technological (who owns & can modify the stack), operational (who runs it, 24/7, with root), and economic (who benefits) — and advances along a five-level continuum from mere data residency to fully sovereign cloud for classified missions. Map workloads to levels by sensitivity, then enforce with governance that continuously audits compliance.

Critical-infrastructure security sets the floor. Power, water, logistics, healthcare and telecom now depend on IT/OT convergence. Regulators are raising the baseline: the EU's NIS2 obliges risk management and incident reporting across 18 critical sectors and mandates cross-border cooperation, strengthening national strategies while enabling joint response. In parallel, CISA's Cross-Sector Cybersecurity Performance Goals (CPGs) provide a pragmatic, high-impact control set for owners/operators to reduce risk fast. Technical playbooks for ICS/OT remain anchored in NIST SP 800-82 (Rev. 3), tailored to real-time constraints of control systems. Bottom line: CNI workloads generally warrant Level 3–4 sovereignty (domestic legal/operational control, locally governed keys, minimal foreign dependency), with segmentation to prevent "sovereignty leaks" from lower-tier systems.

Sovereignty and defence demand the top tier. Defense and intelligence are the ultimate use case for Level 5: national ownership of hardware, software, staff, and key material; air-gapped or severely gated connectivity; and resilience under wartime isolation. Yet allied operations still require interoperability. NATO's Federated Mission Networking (FMN) demonstrates how sovereign nodes can federate via agreed roadmaps, specifications, and data-exchange policies — "sovereignty in concert." The lesson: architect for federation by consent — you retain control at home, but expose standardized, audited interfaces for coalition missions.

Guard against extraterritorial law and supply-chain leverage. The U.S. CLOUD Act and similar statutes clarify when providers can be compelled to disclose data — even if hosted abroad — creating jurisdictional exposure if foreign vendors hold keys or administrative control. France's ANSSI moved to counter this with its SecNumCloud "trusted cloud," embedding organizational and technical safeguards specifically to resist extraterritorial access. Germany's BSI C5 provides a rigorous, regularly updated control catalogue for cloud assurance and attestation. Practically, this means externalizing key management and eliminating foreign "root" in sensitive systems: whoever holds the keys holds the kingdom.

Multilateral agreements for "sovereignty stacks." The way out of the efficiency-vs-sovereignty trap is to standardize the interfaces of sovereignty, not to outsource it. Two proven building blocks exist:

• Federated trust frameworks (e.g., Gaia-X): define attestation, identity, policy metadata, and service conformance so that independently controlled clouds can exchange data and verify properties without ceding control.
• Operational coalitions (e.g., NATO FMN): codify how sovereign networks federate for missions through agreed specs and baselines.

A bilateral or allied "Sovereignty Stack Mutual Recognition" should combine both: legal reciprocity (no extraterritorial claims), cryptographic and operational reciprocity (mutual attestation, joint audit hooks), and emergency break-glass procedures governed by treaty rather than vendor policy.

Post-quantum communications across borders. Cross-border links should migrate now to NIST-approved post-quantum primitives: ML-KEM (FIPS 203) for key establishment, ML-DSA (FIPS 204) and SLH-DSA (FIPS 205) for signatures, with staged hybrid deployments during transition. This is the fastest path to quantum-resilient diplomatic, defense, and trade traffic while maintaining interop. Publish national profiles (cipher suites, key lifetimes, audit requirements) and require domestic control of PQC key material at each sovereign endpoint.

Surveillance, monitoring, and data sharing with partners. A modern cyber defense posture is data-led and collaborative. Allied cybersecurity agencies are already issuing joint OT guidance and building telemetry-sharing pipelines to raise the baseline; emulate and extend that model with federated SOCs that exchange high-value indicators, models, and playbooks while keeping raw citizen data at home. NATO's 2025 Data Strategy points the same direction: treat data as an alliance asset with governance that enables — but does not compel — sharing.

Conclusion & Recommendations — Decide, design, and drill

1. Adopt a national sovereignty map — then fund to it. Classify workloads by sensitivity and assign minimum levels: public digital services (L1–2), internal administration (L2–3), critical infrastructure & regulated industries (L3–4), defense and intelligence (L5). Publish deadlines and procurement rules accordingly.

2. Keys at home, always. Mandate externalized, domestically governed key management for all sensitive workloads; providers may host encrypted data, never sovereign keys. Bake this into contracts and certifications (e.g., align with C5 and SecNumCloud-style controls).

3. Harden critical infrastructure to a common bar. Require CPG-aligned controls as a national baseline and require operators to evidence OT-specific mitigations per NIST SP 800-82 Rev.3; couple with sectoral tabletop exercises and sovereign incident-response drills.

4. Stand up a national orchestrator — with an independent auditor. Separate duties: a sovereign cloud/AI orchestrator integrates and operates the national stack; a statutory, technically capable auditor runs continuous compliance and anomaly detection with real-time access and enforcement powers.

5. Codify "Sovereignty Stack Mutual Recognition" with allies. Borrow from Gaia-X (trust framework) and NATO FMN (mission federation): mutual legal non-assert (no extraterritorial reach), shared attestation, cross-audit APIs, and pre-agreed data corridors for crisis response.

6. Move cross-border links to PQC now. Issue a national profile that standardizes ML-KEM for key exchange and ML-DSA/SLH-DSA for signatures on government and CNI interconnects; prioritize border gateways, diplomatic links, and defense peering; require dual-stack (hybrid) during transition.

7. Build federated surveillance and joint response. Establish a treaty-anchored, privacy-preserving telemetry exchange (IOCs, TTPs, models), aligned with NIS2 cooperation obligations and Five-Eyes-style joint advisories for OT. Measure success in dwell-time reduced and mean-time-to-patch across the federation.

The choice is not isolation vs dependence. It is cooperative sovereignty: absolute control where it matters, interfaces that federate when it counts, and cryptography that survives the next decade. The countries that make these decisions now will write the operational terms of freedom in a connected age.

Conclusion

Urgency: The window for action is narrowing

Canada's competitive position deteriorates quarterly. Q2 2025 showed labor productivity decline of 1.0% after modest 2024 recovery (+0.6% annual). OECD projects Canada dead last among advanced economies through 2060 for GDP per capita growth — a verdict on current trajectory absent intervention. Business investment declined 6 times in 7 quarters (2022-2024). Canada dropped from 5th to 8th in Tortoise Global AI Index in single year (2023-2024) with only 0.7% global compute capacity. Of 3,162 STEM graduates sampled, 25% work outside Canada rising to 66% for software engineers — 84% of Waterloo 2020 software engineering class planned US work. This talent exodus is accelerating not stabilizing.

US private AI investment reached $109 billion in 2024 — 12 times China, 24 times UK, and roughly 15 times total Canadian venture capital across all sectors ($7.1B in 2023). US produced 40 notable AI models in 2024 versus Europe's 3. Singapore deployed S$1B+ over 5 years with specific target to triple practitioners to 15,000. France committed €2.22B over 10 years targeting double specialists and 3 institutions in global top tier. UK committed £100M BridgeAI with 20x AIRR expansion by 2030 and "AI maker not taker" positioning. Even China's reduced pace maintains strategic focus with $2B R&D, complete vertical stack (Samsung chips+software), and cultural receptiveness (10.72M ChatGPT users, 76% believe AI will impact entire economy). Canada risks permanent structural disadvantage if competitors lock in advantages during current deployment surge.

The energy advantage represents time-limited opportunity. Current 15 GW application backlog (20x existing capacity) demonstrates demand exists now. Hyperscalers making multi-decade data center location decisions in 2024-2027 period. AWS committed $18B through 2037 for Calgary region — decisions made, capital allocated. Microsoft's $80B FY2025 global AI data center spend and Google's $75B 2025 capex (42% YoY increase) represent investment cycle happening in real time. If Canada fails to solve grid interconnection and establish First Nations partnership frameworks during this window, hyperscalers will build capacity elsewhere with 20-30 year commitments. The energy advantage doesn't disappear but the call option expires worthless if not exercised during capital deployment cycle.

Productivity crisis compounds. Census 2021 found 3 in 5 Canadian workers in occupations with high AI exposure potential. McKinsey estimates $4.4 trillion global productivity potential from AI corporate use cases with 92% of companies planning increased investment — yet only 1% report being "mature" in deployment. Vanguard projects 20% automation rate across US jobs by 2035 potentially raising GDP growth to 3% in 2030s — fastest since late 1990s. If US achieves 3% growth through AI deployment while Canada remains at 1-1.5%, the productivity gap (currently 72% of US levels) will become unbridgeable. Current 9-percentage-point decline since 2000 (81% to 72%) accelerates to 20+ points by 2035 creating permanent structural divergence. At that stage, brain drain becomes brain hemorrhage as entire cohorts see no viable Canadian career paths.

Recommended immediate actions for Privy Council Office coordination:

Month 1-6: Establish Canadian Digital Centre with spend control authority modeled on UK GDS. Deputy Minister Committee mandate for AI inference deployment across government demonstrating use cases. Treasury Board approval for differentiated data sovereignty framework (classified/Protected B/Protected A tiers). Negotiate First Nations capacity/equity funding templates with BC, Ontario, Manitoba, Quebec leadership.

Month 3-12: Launch fast-track grid interconnection process targeting 12-18 months maximum with performance guarantees. Identify data center corridors in northern Quebec, northern Ontario, Manitoba, BC for transmission co-investment. Deploy first $300M of AI Compute Access Fund removing SME capital barriers. Establish 5 sector-specific AI Centers of Excellence (agriculture, manufacturing, healthcare, financial services, natural resources).

Year 2-5: Scale successful approaches nationally. Achieve 35%+ business AI adoption from current 12.2%. Commission gigawatts of new data center capacity from 15 GW application backlog. Establish Canada as preferred North American AI inference location leveraging energy advantages (5.23-12 cents/kWh, 50% cooling cost reduction, 82% non-GHG). Reverse brain drain through competitive domestic opportunities. Achieve NATO GDP defense target by 2027 with NORAD modernization deploying cloud-based AI-enabled command systems on sovereign compute infrastructure.

The choice is stark: orchestrate AI inference ecosystem deployment at scale capturing productivity gains and energy advantages, or continue current trajectory toward permanent structural disadvantage relative to peer nations. Government's role is creating conditions for private sector success — not buying vendor solutions, not imposing blanket rules preventing innovation, not nurturing research that never commercializes. Platform provision, grid infrastructure, First Nations partnerships, procurement transformation, talent retention, and risk-based governance enable the 87.8% of businesses not yet using AI to achieve documented 30-66% productivity improvements. This is framing an entire industry productivity flashpoint — Canada either seizes the moment or watches competitors lock in advantages during current deployment cycle while productivity gap becomes unbridgeable.

With the five key themes discussed — geopolitics, productivity imperative, AI opportunity, strategy shift, and system reforms — we have a blueprint of both why and how Canada must act. In essence, we face a "productivity flashpoint": the convergence of global pressures and technological potential has created a brief window in which bold action can dramatically change our trajectory. If we align policy, capital, and will, Canada can leap forward, turning its perceived disadvantages (small market, proximity to a giant neighbor) into advantages (agility, niche focus, trusted partner status). If we do not, the compounding effect of inaction will be very hard to reverse — every year of delay means further entrenchment of foreign tech dominance in our economy, more domestic talent leaving for better opportunities, and continued erosion of our relative wealth.

This is a nation-building project for the 21st century. Just as past generations built railways, highways, and power grids to secure Canada's economic future, our generation must build the digital, AI, and productivity infrastructure that will undergird prosperity for decades to come. It's nation-building in a modern form: less visibly physical, but equally transformative. And it requires a "whole-of-nation" effort — government policy, private sector dynamism, academia's innovations, and indigenous and community participation (like the First Nations data center ventures) all moving in concert. The federal government, under the Privy Council's coordination, is uniquely positioned to spearhead this alignment. We must move with urgency and purpose, for the world will not wait.

Appendix A — List of Recommendations

Five Immediate Priority Actions

1. Launch an Energy-to-Compute Pathway. Create pre-permitted sites with transparent power blocks and standardized "revenue-per-watt" contracts that unlock fully private financing for AI inference campuses. This enables provinces to convert their clean energy advantage into exportable compute infrastructure without public capital.

2. Align Policy for Optimal Capital Allocation. Establish a Scale-Up Co-Investment Window for strategic late-stage rounds, introduce immediate expensing for AI infrastructure, launch an AI Adoption Tax Credit for SMEs, and provide financing guarantees rather than subsidies. Publish a quarterly Capital Velocity Scorecard to track how quickly policy converts into productive assets.

3. Announce an Ecosystem Procurement Pilot. Launch in two priority domains (digital service delivery and government optimization) with outcomes-based tenders, mandatory interoperability, and multi-phase "pilot → scale" awards. This moves away from single-vendor lock-ins toward orchestrated ecosystems that deliver results.

4. Sovereign Computing Architecture. Build a five-province, Tier IV, post-quantum-secure architecture with fragmented storage, high-availability pairing, and lights-out operations. This keeps sensitive data resident and services online while enabling cost-efficient inference at a national scale.

5. Create a Clerk-led Productivity & Sovereignty Taskforce. Establish transitory leadership to align federal levers (procurement, skills, data governance) and measure quarterly progress toward adoption targets. This ensures whole-of-government coordination rather than fragmented departmental efforts.

GDP Growth Recommendations

6. Set Organization-Level Adoption Targets. Establish "AI-in-work" KPIs by sector measuring share of organizations with implemented copilots, share of hours augmented, and average task multiplier by function. Manage these inputs deliberately to achieve the 2-5% GDP mechanical lift.

7. Make Public Sector the Throughput Proof-Point. Prioritize copilots in high-document, high-queue domains (benefits, permits, casework) and re-engineer workflows to realize throughput not just time-saved. Track cases/agent/day and backlog-days as public administration represents 21.6% of employment.

8. Tie Adoption to Hours and Work Design. Mandate task inventories and prompt playbooks at team level while funding manager training to redesign queues, handoffs, and QA around AI. Moving utilization from 25% to 50% of hours touched doubles the mechanical GDP effect.

9. Build a National Compute-Export Lane. Create a "Compute-to-Export" program bundling expedited zoning/permitting, long-term clean-power PPAs, interconnection prioritization, and trade facilitation. Global demand is strong but the bottleneck is power plus interconnect.

10. Publish Compute-Exports Satellite Account. Break out Computer services (SI2) with a compute-exports memorandum item in national accounts releases. Adopt a levelized import content framework so policy can track net exports without double-counting.

Infrastructure & Energy Recommendations

11. Publish AI-Ready Infrastructure Baseline. Direct NRCan and provincial utilities to produce public inventory of AI-ready MW and GPU slots by campus, including rack-density limits and liquid cooling readiness. Use this as the capacity ledger for allocations and incentives.

12. Stand Up Energy Call Option Auctions. Launch pilots in provinces with headroom (QC/MB/BC/NL) with 5- and 10-year tranches and 24/7 carbon-free energy floors. Reserve portions for AI campuses that meet heat-reuse and response-time standards.

13. Designate AI Free Trade Zones. Create two pilots — one brownfield (retrofit with liquid loops) and one greenfield megasite (pre-zoned + modular substation). Set binding permit shot-clocks and publish interconnection queues and tariffs.

14. Power-Aware Siting and Procurement. Tie data-center siting to grid nodes with surplus low-carbon power, enable zonal pricing and behind-the-meter options. Power is the cap; clean power protects both margins and climate goals.

Capital Velocity Instruments

15. 90-Day Innovation Procurement Fast Lane. Create standardized, challenge-based federal procurement that moves from problem to pilot in ≤90 days. Include milestone payments, pre-negotiated IP terms, and cloud/AI safety guardrails.

16. Sovereign Co-Investment Facility. Implement automatic 40% public co-investment matching for Series B+ rounds above $25M led by qualified funds. Government takes common/preferred equity with tag-along rights and 8-year exit horizon.

17. SR&ED Real-Time Refund. Convert to monthly cash rebates based on validated R&D spending with simplified digital filing. Raise rates to 45% for AI/productivity software while lowering for marginal improvements.

18. Frontier Tech Visa Program. Create 2-week approval for AI/quantum/biotech talent with 3-year open work permits and permanent residence pathway. Include "friendshored R&D grants" covering 40% of salary for researchers relocating from partner nations.

19. Senior Lender Government Pool. Deploy $5B federal guarantee pool enabling banks to lend against intangibles, ARR, and forward contracts. Target 5-10x private leverage with portfolio approach to risk.

Sovereign Computing Architecture

20. Build Five-Province Sovereign Infrastructure. Create BC, AB, ON, QC, NS footprint for geo-diversity with Tier IV facilities, independent power/cooling, and fault-tolerant networks. Implement fragmented storage, zero-trust networking, and post-quantum cryptography by default.

21. Deploy National Productivity Stack. Allocate capital where it turns into GDP fastest through inference at the point of work. Standardize sovereign stack paying for cycle-time reductions, defects avoided, and hours returned to service.

22. Implement Post-Quantum Cryptography. Adopt NIST's PQC suite (FIPS 203/204/205) aligned with CCCS federal migration roadmap. Validate performance on CANARIE 400-Gbps backbone before full production rollout.

Procurement Transformation

23. Platform-Not-Products Strategy. Buy shared platform layers (vector DBs, model routers, monitoring) behind open interfaces. Purchase applications as plugins meeting interface, security, and portability tests.

24. Modular Contracting with Objective Tests. Replace big-bang RFPs with short, iterative task orders competing capabilities against common test suites. Align to NIST AI RMF and ISO/IEC 42001 practices.

25. Mandate Open Interfaces and Portability. Insert standard no-lock-in clauses requiring data export in open formats, API conformance to GC standards, and model portability. Make exit and reversibility real with annual exit drills.

26. Institute Spend Controls. Require central approval for digital/AI spend duplicating platform capabilities. Grant earned autonomy to mature teams meeting policy-as-code, security, and portability baselines.

Governance & Measurement

27. Establish Canadian Digital Centre. Create entity with spend control authority modeled on UK GDS. Give Deputy Minister Committee mandate for AI inference deployment demonstrating use cases.

28. Create Sector-Specific AI Centers. Launch 5 Centers of Excellence in agriculture, manufacturing, healthcare, financial services, and natural resources. Focus on deployment and adoption rather than research.

29. Treasury Board Sovereignty Framework. Define sovereignty as Evidence-of-Control with minimum proof objects for keys, policy, attestation, and exit. Make these gating criteria for all awards.

30. Launch Cross-Sector Innovation Sandboxes. Stand up sandboxes for predictive maintenance, workflow automation, smart infrastructure with fixed governance playbook. Require portability to GC platform as graduation criterion.

Sources: OECD (GDP per hour worked; "Miracle or Myth?" macroeconomic productivity gains); Fraser Institute and TD Economics on Canada's productivity gap; Statistics Canada (real GDP per capita; Q2/Q3 2025 AI adoption; international trade in services); KPMG Canada (public-sector AI adoption); McKinsey Global Institute ("Out of balance"); Aschenbrenner, "Situational Awareness: The Decade Ahead" (2024); NVIDIA GTC 2024 keynote; Meta Q4-2023 earnings call; IEA (Energy and AI; Electricity 2024/2025); Dell'Oro Group; CBRE; Deloitte; Hydro-Québec, IESO, BC Hydro, Keeyask, Alberta Energy Regulator, Government of Nova Scotia, NB Power, and provincial releases; Government of Canada (Cohere investment; AI Compute Access Fund; CIC blueprint); Bell AI Fabric; NIST PQC standards (FIPS 203/204/205) and CCCS federal PQC migration roadmap; CANARIE; NBER and GitHub/Microsoft productivity studies; and the supporting analyses listed in Appendix B.