Primary-source Canadian power, regulatory, and emissions data — the CER cost spine and ECCC grid factors above.
Canada’s Compute
What actually determines whether a Canadian region can host the next generation of large-scale AI and data-centre development?
An 8-page research brief and a source-backed dataset: grid emissions, the 2026 large-load connection process, and the interconnection bottleneck across five provinces — plus a playable six-round model of the same trade-offs. Every value carries a source and a date.
At a glance.
5provinces compared on grid emissions and large-load access
~230×range in grid carbon intensity, Alberta vs Québec
5 / 5regions where large-load connection is rationed, not on request
On paper, several provinces plainly have enough electricity for AI. The finding is that in every one of the five examined, a new large compute load in 2026 has to get through a rationing mechanism — a selection process (Québec), a competitive call (British Columbia), an interim allocation block that is already fully committed (Alberta), or a system study with no published duration (Ontario). Manitoba has the most near-term room and a forecast capacity shortfall around 2029/30. How clean or how large the provincial grid is does not change this.
Download the research.
Research release 2026-09 (v1.0). The brief, the dataset, and the model read the same sourced values. The dataset is regenerated from a documented pipeline and passes a schema / source-id / rubric validation gate.
- Research brief — 8 pages, PDFDownload ↓
- Regional dataset — 5 provinces, CSVcanada_compute_regions.csv ↓
- Technology cost spine — 7 technologies, CSVtechnology_costs.csv ↓
- Source manifest — 35 sources with URL and retrieval datesource_manifest.csv ↓
- Claim registry — each headline claim → source + typeclaims.csv ↓
- Methodology — pipeline, gates, and known limitationsMETHODOLOGY.md ↓
- Data dictionary — every field, unit, and source idDATA_DICTIONARY.md ↓
The live model.
What the power costs.
The model's cost spine is the Canada Energy Regulator's Canada Energy Future 2023 Electricity Supply Model, verified line-by-line against the live table. Capacity factors are 2025 U.S. EIA utility-scale empirical proxies — performance references, not site forecasts. Together these decide how much firm compute a dollar of capital actually energizes.
| Generation | Overnight CAPEX | Fixed O&M | Capacity factor |
|---|---|---|---|
| Natural gas combined cycle | $1,450/kW | $50/kW-yr | 58.4% |
| Utility solar PV | $1,700/kW | $39/kW-yr | 24.4% |
| Onshore wind | $1,900/kW | $72/kW-yr | 34.2% |
| Gas combined cycle + CCS | $3,705/kW | $110/kW-yr | 58.4% |
| Large hydro > 100 MW | $4,500/kW | $51/kW-yr | 35.3% |
| Geothermal | $7,100/kW | $137/kW-yr | 65.9% |
| Small modular reactor | $9,262/kW | $63/kW-yr | 91.0% |
Capacity MW = committed power capital ÷ benchmark CAPEX. No crowd-in multiplier, no province/technology "fit" multiplier — the physical result stays traceable to the capital allocation and the benchmark.
The Canadian regional landscape.
Aggregate provincial capacity means little if the power cannot be interconnected where the campus is being built. Grid emissions intensity is observed (ECCC 2026); the large-load access column is the current connection reality quoted from each system operator, not a published queue time. Full rows, with the categorical access assessment and its rubric, are in canada_compute_regions.csv.
| Region | Grid intensity | Large-load access today |
|---|---|---|
| Québec | 1.9 gCO₂e/kWh | Data centres above 5 MW face a selection process — connection is allocated, not open headroom |
| British Columbia | 18 gCO₂e/kWh | Competitive process for emerging-industry connections; capability is project- and location-specific |
| Ontario | 59 gCO₂e/kWh | IESO handles large computational loads through a system-specific technical study, not a queue with a set duration |
| Manitoba | 2.5 gCO₂e/kWh | More near-term room — Manitoba Hydro projects peak demand exceeds capacity only around 2029–30 |
| Alberta | 438 gCO₂e/kWh | AESO's interim 1,200 MW Phase-1 large-load limit is fully allocated; further projects move to the next phase |
The earlier model treated Alberta as fastest and assumed a 1,200 MW province-wide "headroom" figure. Checked against AESO, that block is fully committed across two projects — corrected to a constraint, not available capacity.
Observed, proxy, or calibration.
The model separates primary-source inputs, external proxies, and explicit scenario assumptions rather than treating every number as equally certain. A test gate enforces it: every region must cite the ECCC grid source, every primary benchmark must cite CER, and every calibration must declare itself as one.
Comparable external inputs where no Canadian benchmark exists — EIA reference-plant lead times and capacity factors.
Disclosed scenario parameters — regional interconnection-delay years, event-deck effects, classification thresholds. Labelled, never presented as measured.
Sources.
Every technology and region in the executable model carries source metadata and an evidence-kind tag, surfaced as clickable links in the interface and enforced by an integrity test.
Research direction.
- Practitioner review
The timing, tariff-classification, and archetype thresholds have not yet had an energy or grid-planning professional review them — the highest-value next validation step.
- Interconnection
Regional transmission and large-load queue constraints at more resolution than a single province-wide calibration.
- Site-level analysis
Regional deployment cases for specific hyperscale campuses.