GDP, Workforce Amplification, and Net Exportability

Key finding: Using organization-level adoption, realistic task multipliers (typically 1.14×–1.26×, with 1.67× a task-specific ceiling), and a default 40% utilization factor, the near-term level effect on GDP for Canada ranges from roughly +0.8% in conservative cases to +8.36% in ambitious, well-executed scenarios — before counting any net-export contribution from selling compute abroad.

Two Reinforcing AI Channels to Increase GDP

At its core, GDP = C + I + G + (X − M), and GDP per capita can be decomposed into productivity × employment/population × hours/worker. 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, 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 can lift GDP through two reinforcing AI channels, each described in more detail in the following pages:

1. Domestic productivity — firms and public bodies doing more per hour, and
2. Net exports impact — selling Canadian data-centre GPU capacity to foreign buyers while global supply is tight. (To provide a more conservative estimate of GDP uplift potential, we did not account for the net export impact.) However, we believe that, if executed properly, this could generate a significant impact on GDP, while increasing Canada's geopolitical leverage.

Using organization-level adoption (not personal use counts), realistic task multipliers (typical 1.14×–1.26×; 1.67× as a task-specific ceiling), and a utilization factor for the share of hours actually augmented, the near-term level effect on GDP ranges from well under 1% in conservative cases to high single digits in ambitious, well-executed scenarios. Over time, those level gains should be phased into annual growth paths consistent with widely cited macro guardrails (roughly +0.4–0.9 pp/year to labour-productivity if adoption, complements, and utilization keep rising).

In parallel, if Canada scales AI infrastructure faster than most OECD peers (ex-US/China), compute services exports could add a separate, material boost to (X − M) while the global market remains supply-constrained.

Domestic Productivity

OECD Scenario Bands (Context for Peers)

Across advanced economies, enterprise adoption today spans roughly low → high = ~6% → ~28% of firms, with credible within-task productivity effects clustering around 1.14×–1.26× and 1.67× as an upper bound on well-scoped tasks. Crossing those adoption bands with the multipliers (and a default u = 40%) produces a clean grid of ΔGDP levels that stays agnostic to national structures — useful for benchmarking Canada's numbers against peers.

Canada Scenario Bands + Grid (What the Numbers Mean)

For Canada we use org-level adoption only: private-sector bands anchored on current use and near-term plans; public-sector bands on implemented/piloting shares. We weight by employment (≈ 78% private / 22% public) and hold u = 40% in the base grid. The result is three intuitive bands:

• A — "implemented footprint" delivers roughly ≈ 0.8% → 3.8% depending on whether tasks average 1.14× → 1.67×.
• B — "near-term ramp (incl. partial pilots)" yields ≈ 1.1% → 5.3%.
• C — "implemented + pilots convert" reaches ≈ 1.8% → 8.4%.

These are level effects, not one-year growth rates; they materialize over several years as tools, processes, data, skills, and governance catch up. They also scale linearly: if only 25% of hours are AI-augmented near-term, multiply the grid by 0.25 ÷ 0.40 = 0.625; if average task lift is 1.20×, scale by (0.20 ÷ 0.14) from the 1.14× column, and so on.

How to Read the Pages That Follow

1. The OECD bands set the outer guardrails (adoption × multiplier).
2. The Canada grid applies Canadian adoption and employment structure to quantify plausible ΔGDP levels, with clear dials for m and u.
3. The compute-export section shows how to translate capacity plans into net exports and, therefore, into GDP via (X − M).

Together, they give you a coherent, spreadsheet-ready way to stress-test Canada's AI upside across both productivity and trade channels — without double-counting and with assumptions you can transparently tune.

How to Account for Domestic GDP Uplift

First-order effect on the level of GDP:

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

…where:

• a_priv, a_pub = org-level adoption shares (private/public).
• s_priv, s_pub = employment shares (private/public).
• m = task-level productivity multiple for augmented hours (e.g., 1.14, 1.26, 1.67).
• u = share of a typical worker's hours that are actually AI-augmented (25–60% typical early-use range).

OECD Scenario Bands (Adoption × Multiplier)

Benchmarks (enterprises, org-level)

• Adoption bands: A = 6% (low), B = 13.5% (mid), C = 27.6% (high).
• Multipliers: Low = ×1.14, Base = ×1.26, High = ×1.67.
• Default utilization: u = 40% of hours augmented (scale cells by u/0.40 to adjust).

Mechanical level effects (percent of GDP), using ΔGDP ≈ adoption × u × (m − 1):

Adoption ↓ \ Multiplier →	Low ×1.14	Base ×1.26	High ×1.67
A — 6%	0.34%	0.62%	1.61%
B — 13.5%	0.76%	1.40%	3.62%
C — 27.6%	1.55%	2.87%	7.40%

Notes

• These are OECD-style enterprise bands — they do not weight public vs. private employment.
• Treat ×1.67 as a task-specific ceiling (e.g., pro writing, well-scoped coding).
• For annual paths, phase adoption/utilization over time and cross-check against macro guardrails (~0.4–0.9 pp/yr to labour productivity over 10 years).

Canada Scenario Bands (Org-Level Only, Not Personal Use)

Employment weights (Canada)

• s_priv = 78.4%, s_pub = 21.6% of total employment.

Adoption definitions (org-level)

• Private (A → C):
  • A: 12.2% of businesses used AI to produce/deliver in the last 12 months (Q2-2025).
  • B: 14.5% plan to use AI in the next 12 months (Q3-2025).
  • C: 24.9% high-case = 12.2% + 14.5% among non-users → 0.122 + 0.145·(1−0.122).
• Public (A → C):
  • A: 22% of public-sector organizations implemented AI.
  • B: 38% = implemented + half of pilots (22% + 0.5×32%).
  • C: 54% = implemented or piloting.

Multipliers. Low = ×1.14, Base = ×1.26, High = ×1.67. Default utilization. u = 40% of hours augmented.

Canada Grid — ΔGDP level effect (percent of GDP), 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%

Quick scalers.

• Change u: multiply the whole table by (u / 0.40).
• Change m: replace (m − 1). Example: with ×1.20, scale the Low column by (0.20 / 0.14) ≈ 1.43×.
• If you prefer Q2-2025 plans = 17.9% instead of 14.5% for private-sector B/C, replace a_priv accordingly and recompute C as: 0.122 + 0.179·(1−0.122) ≈ 27.4%.

Interpretation & timing.

• These are level effects holding other macro factors constant; realized macro paths typically accumulate over several years as adoption, complements (skills, data, compute), and process redesign mature.
• Use OECD's 10-year guardrail (≈ +0.4–0.9 pp/yr to labour-productivity) to map these levels into annual growth paths.

Key Takeaways

• Across OECD peers, enterprise-level AI adoption spans roughly low → high = 6% → 28%.
• Credible task/team-level productivity effects cluster around ×1.14 → ×1.26, with ×1.67 a ceiling for well-scoped tasks.
• For Canada, using org-level adoption only (no user-level stats), and a default utilization u = 40%, the implied level effects range from ≈ +0.8% (Conservative ×1.14) to ≈ +8.36% (Ambitious ×1.67) on a mechanical, first-order basis.
• Macro guardrails (OECD) suggest annual labour-productivity growth could be lifted by ~0.4–0.9 pp/yr over a decade as adoption and utilization scale — so phase these bands over time.

Sources Backing the Inputs Used

• Canada (private sector, org-level): Used AI to produce/deliver (Q2-2025): 12.2%; plus 17.9% plan to adopt AI software in the next 12 months (same release) — Statistics Canada. Plans to use AI (Q3-2025): 14.5% (The Daily) — Statistics Canada.
• Canada (public sector, org-level): 22% implemented, +32% piloting (i.e., 54% implemented-or-piloting); 48% of public servants use AI tools at work (KPMG survey, Sep 18, 2025) — KPMG.
• Employment weights (Canada): Public-sector share of total employment ≈ 21.6% — C.D. Howe Institute.
• OECD/EU adoption benchmarks (enterprises): EU-27 enterprises using AI in 2024: 13.5%; highs Denmark 27.6%, Sweden 25.1%, Belgium 24.7% — European Commission.
• Task/team-level productivity multipliers: ×1.14 (+14%): "Generative AI at Work" (customer-support agents, field deployment) — NBER. ×1.26 (+26%): Three field experiments with software developers (Microsoft, Accenture, Fortune-100), RCTs — MIT Economics. ×1.67 (~+67% implied): GitHub Copilot controlled experiment (55.8% faster on coding task) — arXiv.
• Macro guardrail (10-year horizon): OECD micro-to-macro model: +0.25–0.6 pp to TFP per year, roughly +0.4–0.9 pp to labour-productivity growth — OECD.

Net Export Impact (Selling Compute Capacity to Foreign Countries)

The second engine lifts (X − M) directly. Treat GPU capacity sold from servers on Canadian soil to non-residents as exports of computer services (digitally delivered). Model supply in H100-equivalent GPU-hours:

Exportable GPU-hours = installed GPUs × 8,760 × utilization × export share × performance factor

Multiply by a price path ($/H100-hr) to get exports. Then subtract a levelised import content per GPU-hour — amortized imported hardware and any foreign software/licensing — to get net exports: Net exports ≈ GPU-hours × (price − LIC).

While global compute remains scarce, assume high take-up at target prices (bounded by latency and digital-trade rules); as markets mature, shift to a share-of-market model. Either way, this channel is additional to domestic productivity gains and can be material if Canada accelerates data-centre build-out, interconnection, and clean-power supply faster than peers.

We firmly believe that Canada has the potential to become a net exporter of AI computing power. However, we intentionally omitted a comprehensive GDP analysis of this export value to present a more cautious estimate of the overall impact on GDP. Nonetheless, we highly recommend prioritizing this sector as a crucial part of Canada's future export strategies.

What Accounts for "Export of Compute"?

Treat the sale of GPU/AI compute from servers located in Canada to non-residents as digitally delivered services (WTO Mode 1: cross-border). In standard trade stats this sits under EBOPS 2010 "Telecommunications, computer and information services → Computer services (SI2)" and is part of digitally deliverable services. That covers cloud/hosting, data processing, and platform infrastructure (IaaS/PaaS).

Why it matters: counting it this way means the export shows up in X (services), directly boosting (X − M); any imported GPUs or foreign licences hit M when they occur, keeping GDP consistent with domestic value added.

Supply-Side Capacity: From GPUs to Exportable GPU-Hours

Define a uniform unit (e.g., H100-equivalent GPU-hour) to combine different accelerators. For each GPU class i, exportable GPU-hours in year t are the sum across classes of:

Exportable GPU-hours = Σ ( N_i,t × 8,760 hrs/yr × u_i,t × s_i,t^export × φ_i )

where N is installed GPUs, u is utilization, s^export is the share sold to non-residents, and φ is the performance factor versus H100.

• u (utilization) captures uptime and booking efficiency (training is batchable, so 60–90% is common in steady state).
• s^export is the fraction of capacity sold abroad (constrained by data-localisation / digital services restrictions in destination markets; use the OECD Digital STRI to parameterise reachable demand by country/region).
• Check grid/power ramp feasibility while you plan capacity — recent IEA work flags electricity and interconnection as binding constraints in some locales.

Price Path: $p Per H100-Equivalent GPU-Hour

Benchmark with hyperscaler list prices (upper bound) and apply your wholesale discount:

• AWS p5.48xlarge (8× H100) on-demand (us-east-1) ≈ US$98.32/hr → ≈ US$12.29 per H100-hr.
• Azure ND H100 v5 (8× H100) on-demand ≈ US$98.32/hr (region-dependent).
• Google Cloud A3 (8× H100) pricing varies by region; Google's GPU price page confirms A3 on-demand/Spot frameworks (use region SKU for your target markets).

Gross Export Revenue

Exports of compute services_t = Exportable GPU-hours_t × p_t ⇒ ΔX_t = Exports_t

This books to "Computer services (SI2)" within services exports. For context, Canada's total services exports were $219.9 B in 2024; compute exports would add to that line.

Imports to Subtract (To Get Net Exports)

Compute services have heavy imported capital content (accelerators, servers, optics) and some imported OPEX (foreign licences/support). To keep alignment with GDP by expenditure (C + I + G + (X − M)), track:

1. Imported CAPEX (goods) in the year it arrives (one-off spikes).
2. Imported services required to deliver compute (e.g., foreign software licences, specialized support).

A convenient way — especially for 10-year scenarios — is to compute a levelised import content per GPU-hour (LIC) so you can net it out against price each year:

LIC_t = ( Imported CAPEX_t × CRF + Imported OPEX_t ) ÷ Exportable GPU-hours_t

• CRF is a capital-recovery factor using asset life L and discount rate r.
• Electricity is domestic and therefore not subtracted in (X − M) (though it matters for profitability).
• The national-accounts community is actively refining guidance for cloud computing in SNA/BPM — use the above as a first-best approximation aligned with current practice.

Then:

Net exports of compute_t = Exports_t − Imports_t ≈ Exportable GPU-hours_t × (p_t − LIC_t)

and the incremental GDP effect from trade in year t is Δ(X − M)_t.

Data Sources

• Statistical framework & classification: OECD/IMF/UN/WTO Handbook on Measuring Digital Trade (2nd ed.) — lists digitally deliverable services and maps EBOPS 2010 SI2 ("Computer services") — OECD. UN MSITS 2010 (modes of supply & trade-in-services concepts) — UNSD. WTO/UN docs on Mode 1 cross-border supply — UNSD.
• Canada baselines: Services exports headline series (to benchmark your add-on) — Statistics Canada.
• Market constraints & access: OECD Digital STRI (cross-border digital trade barriers) — OECD. IEA on data-centre & AI electricity demand and grid limits (feeds u and ramp) — IEA.
• Pricing benchmarks: AWS p5.48xlarge ≈ $98.32/hr (us-east-1); Azure ND H100 v5 ≈ $98.32/hr; Google Cloud A3 H100 pricing by region — Cost Calculator, Vantage.