Canada Data Centres and GPU Capacity

Key finding: Only about 11% of Canada's installed data-centre IT load is GPU-accelerated today — roughly 89% remains CPU-only. National readiness should be assessed in GPU slots and AI-ready megawatts, not floor space or CPU colocation.

Why This Matters

• In the context of AI, CPUs do not matter; only GPUs can process AI.

• We need to evaluate the capacity of generic data centres and focus solely on those that provide GPUs. Currently, only about 11% of the installed IT load in our proxies is GPU-accelerated. This aligns with the idea that CPU-only colocation does not represent AI capacity, and national readiness should be assessed in terms of GPU slots and AI-ready megawatts. By the same logic, data-centre floor space should not be used as a metric. (CER)

• Canada's "AI-ready" growth is approaching — but mostly in the planning stage: examples include Bell's AI Fabric (up to 500 MW) and other large-scale campus projects; Quebec has restrictions on new power today, which influences where the next AI developments may begin. (DataCenterKnowledge)

Below is a proxy "grid" of Canada's data-centre landscape showing counts, current (installed) IT power, hosting mix, and an estimated % CPU vs. % GPU of installed compute. Where public numbers don't exist, I've used transparent proxies and stated the hypotheses behind them.

Canada Data Centres: Counts, Installed IT Load, Hosting Mix, and CPU vs. GPU (2024/25 Baseline)

Region	Est. # of Data Centres	Est. Installed IT Capacity (MW)	Colocation (%)	Hyperscale/Dedicated (%)	Enterprise/Captive (%)	AI/HPC (%)	Est. % GPU (of IT load)	Est. % CPU (of IT load)
Ontario (ON)	105	360.0	65.0	25.0	8.0	2.0	8.1	91.9
Quebec (QC)	57	180.0	50.0	30.0	5.0	15.0	18.9	81.0
British Columbia (BC)	35	109.8	60.0	25.0	10.0	5.0	10.5	89.5
Alberta (AB)	22	69.0	45.0	40.0	12.0	3.0	9.8	90.2
Rest of Canada	20	31.1	75.0	8.0	15.0	2.0	6.7	93.3
Canada (Total)	239	750.0	59.2	26.9	8.2	5.7	11.2	88.8

How to read this table: "Installed IT capacity (MW)" is the currently deployed/energized IT load, not building "full-build" potential. Hosting mix expresses the share of installed IT load by use type. % GPU vs. % CPU estimates the share of installed IT load attributable to GPU-accelerated servers vs. CPU-only servers.

What This Says at a Glance

• Canada has ≈239 operating data centres (all types). Ontario has the largest number (105), followed by Quebec (57), BC (35), and Alberta (22). (CER)

• We proxy current installed IT load at ≈750 MW nationally (not "full build"), with Toronto ≈335 MW and Montreal ≈169 MW anchoring Ontario and Quebec. (CleanBridge)

• Hosting today still skews colocation (≈59%), with hyperscale/dedicated (≈27%), enterprise (≈8%), and AI/HPC-specialized (≈6%) of installed IT load. (Mix varies by province.) (Mordor Intelligence)

• On this basis, only ≈11% of Canada's installed IT load is GPU-accelerated today (and ≈89% remains CPU-only) — consistent with the point that CPU-only capacity is not AI capacity. (See assumptions below.)

Method and Hypotheses (Transparent Proxies)

1) Counts of Data Centres (by Province)

• National count: 239 operating data centres (all types), from the Canada Energy Regulator (CER). Their snapshot also cites Toronto (81) and Montreal (48) as the largest city clusters, based on DataCenterMap. (CER)

• Provincial breakdown: S&P Global (Oct 2, 2024) reports Ontario 105, Quebec 57, BC 35, Alberta 22; I set "Rest of Canada" = 239 − (105+57+35+22) = 20. (S&P Global)

2) Installed IT Power (MW), "Current Usage" Proxy

• National baseline: I use 750 MW of installed IT capacity in 2024 (not full build), per CleanBridge (citing ENCOR). CleanBridge also provides anchors: Toronto ≈335 MW (≈45% of national installed) and Montreal ≈169 MW (≈22.5%). (CleanBridge)

• Provincial split:

  • Ontario = 360 MW (Toronto 335 + ≈25 for Ottawa/others) — conservative given Toronto's anchor.
  • Quebec = 180 MW (Montreal 169 + ≈11 for Quebec City/others).
  • BC = ≈110 MW and Alberta = ≈69 MW by applying each province's share of national site count to 750 MW (35/239 and 22/239 respectively). Remaining ≈31 MW allocated to "Rest of Canada." (This keeps the national sum at ≈750 MW and reflects smaller average site sizes outside Toronto/Montreal.) (S&P Global)

• Note: Other sources discuss "full-build" colocation capacity >800 MW and large future pipelines (10–12 GW) — those are potential rather than installed and were not used for the "current usage" row to avoid inflating today's picture. (Businesswire)

3) Hosting Mix by Province (Percent of Installed IT Load)

We assigned a proxy mix per province, reflecting local realities and recent market notes:

• Ontario 65% colocation / 25% hyperscale / 8% enterprise / 2% AI/HPC — AI leasing is rising but still a small share of installed load. (CBRE)

• Quebec 50 / 30 / 5 / 15 — higher AI/HPC share thanks to hydro-powered AI campuses (e.g., QScale) and hyperscale expansion (but note 2024 power-procurement constraints throttled new projects). (CER)

• BC 60 / 25 / 10 / 5 — new AI builds (e.g., Bell AI Fabric) begin in B.C., but today's installed AI/HPC load is still modest. (DataCenterDynamics)

• Alberta 45 / 40 / 12 / 3 — presence of AWS (Canada West) and growing pipeline; some policy headwinds noted recently. (S&P Global)

• Rest 75 / 8 / 15 / 2 — smaller sites dominated by colocation and enterprise.

These mixes are qualitative proxies cross-checked against: (i) market structure (colocation remains dominant by revenue share), (ii) recent local trends (e.g., AI-led leasing in Toronto; Quebec's power constraints), and (iii) announced AI initiatives (e.g., Bell AI Fabric). (Mordor Intelligence)

4) % GPU vs. % CPU (of Installed IT Load)

Because operators rarely publish GPU counts, we estimated GPU share within each hosting type and then weighted by the provincial hosting mix:

• Colocation: 5% GPU / 95% CPU (mix largely general-purpose servers; limited accelerator density today).

• Hyperscale/dedicated: 12% GPU / 88% CPU (GPU clusters exist, but still minority of installed load across Canadian regions).

• Enterprise/captive: 2% GPU / 98% CPU (most enterprise racks remain CPU-centric).

• AI/HPC: 85% GPU / 15% CPU (purpose-built AI/HPC designs are accelerator-heavy).

• Applying these shares to the hosting mix yields the provincial and Canada-level GPU/CPU estimates in the grid (e.g., Canada ≈11% GPU of installed IT load).

• Policy relevance: GPU slots, power density and interconnect — not building footprint or CPU-only racks — determine a country's AI capacity. (CBRE)

Caveats and Update Path

• Counts: CER's 239 facilities reflect multiple sources (including DataCenterMap). Provincial breakouts rely on S&P's Oct-2024 snapshot. Some campuses comprise multiple buildings; counts will drift with openings/closures. (CER)

• "Current energy usage": I used installed IT MW (≈load available to IT gear) — not building "full build" or utility draw — with 750 MW national baseline and city anchors (Toronto/Montreal). Where city/province detail was missing, I allocated by share of facility count. You can replace proxies with operator disclosures as they become public. (CleanBridge)

• Hosting mix and GPU/CPU shares: No comprehensive public dataset exists for Canada. The mixes reflect qualitative evidence (market reports, leasing notes, power policies) and a conservative GPU assumption. As AI-specific builds (e.g., AI inference farms) come online, GPU share will rise quickly. (CBRE)

Sources for the key baselines: 239 operating data centres, Toronto (81), Montreal (48) — Canada Energy Regulator (Oct. 2, 2024), citing DataCenterMap. Provincial counts (ON 105, QC 57, BC 35, AB 22) and market ≈750 MW installed IT — S&P Global Commodity Insights; CleanBridge (ENCOR). Colocation footprint ("full build" >800 MW; 118 existing colo sites) — ResearchAndMarkets database notices. Market conditions: AI-led leasing in Toronto; Quebec 2024 power-procurement pauses; announced Bell AI Fabric (to 500 MW).