Quebec hydropower and the new gating of AI compute

From cheapest power on the continent to gating new load #

Quebec built its industrial identity on hydroelectricity. The James Bay complex, the Manicouagan and Outardes systems, and the Churchill Falls contract together gave Hydro-Quebec roughly 37 gigawatts of installed capacity by the early 2020s, more than 99 percent renewable, and an export book that supplied New York, New England, and Ontario at margins few utilities could match. Industrial Rate L, the tariff for large customers above 5 megawatts, sat near 5 cents Canadian per kilowatt hour for years, roughly half the equivalent industrial price in Northern Virginia and one quarter the price in Frankfurt or Dublin.

That arbitrage drew aluminum smelters first, then bitcoin miners, then a procession of data center developers. Montreal and the Beauharnois corridor became a credible challenger to Ashburn and Hillsboro for greenfield colocation. By 2023 the cumulative interconnection queue for crypto and data center loads exceeded 30 gigawatts of requests against an export-adjusted spare capacity of perhaps 2 to 3 gigawatts. Something had to give.

What gave was the political consensus that Quebec power should serve Quebec first. In late 2023 the provincial government instructed Hydro-Quebec to reserve new capacity for housing electrification, industrial decarbonization, and battery and green hydrogen projects with measurable economic spillover. By March 2024 every data center request above 5 megawatts required individual cabinet review. By early 2026 the practical message to most hyperscaler developers is that incremental Quebec capacity is gated, dated, and conditional.

Demand surge: from the bitcoin ban to the AI pipeline #

The bitcoin episode shaped the policy reflex now applied to AI. After Hydro-Quebec opened to crypto in 2017, requests reached roughly 18 gigawatts within 18 months. The Regie de l'energie capped new mining allocations at 300 megawatts in 2019, then the province froze further mining connections in 2022 after the bear market exposed how little tax and employment value the load created relative to the power consumed. The political memory is that low industrial tariffs plus undifferentiated demand can strand a generation of cheap hydro on speculative use cases.

AI data centers initially escaped that frame. The 2023 to 2024 hyperscaler surge promised long contracts, large capex, and visible job counts. Through 2024 the Ministry of Economy fielded announcements from Microsoft, Google, AWS, Meta, QScale, eStruxture, Vantage, and several sovereign-adjacent developers, totaling more than 9 gigawatts of stated intent. Allocated capacity was a tiny fraction of that. The table below reconciles the headline announcements against what Hydro-Quebec actually approved.

The gap between announced and allocated is the entire story. Developers who signed land options and ordered switchgear assuming Quebec power would arrive on the usual 24 to 36 month timeline are now facing 60 to 84 month queues, conditional letters, and tariff terms that did not exist when the model was built.

Hydro-Quebec 2035 strategic plan and the deficit by 2027 #

Hydro-Quebec's November 2023 strategic plan, refreshed in late 2025, is the document that reframed the supply picture. The utility now projects domestic Quebec demand to grow by 60 terawatt hours by 2035 against a 2022 baseline of roughly 185 terawatt hours, a 32 percent increase. Drivers include heat pump conversion, electric vehicle adoption, the Northvolt and battery cluster on the South Shore, green hydrogen and steel pilots, and a now-modest but explicitly capped data center allocation.

On the supply side, the plan calls for 8 to 9 gigawatts of new generation by 2035: roughly 3,800 megawatts of wind under a tendered procurement program, 1,500 to 2,500 megawatts of new hydro on rivers including the Petit-Mecatina, several hundred megawatts of solar, and demand response and efficiency contributions sized at 3,500 megawatts. The capital program is in the order of 155 to 185 billion Canadian dollars over the decade. Even on plan, the utility's own modeling shows a structural energy deficit emerging in 2027 and persisting through the early 2030s, met by curtailing exports to New York and New England under the existing firm contracts.

The implication for AI siting is direct. Quebec is no longer long power. The marginal megawatt has an opportunity cost equal to either an export contract margin or a domestic electrification commitment. The tariff and queue policies follow from that arithmetic, not from any anti-data-center sentiment in isolation.

Tariff redesign and the pressure on Rate L #

Rate L, the industrial tariff for customers above 5 megawatts, has been the quiet backbone of Quebec's industrial base. Through 2023 it averaged roughly 5.0 to 5.4 cents Canadian per kilowatt hour all-in, with predictable annual indexation. The Regie de l'energie's 2024 to 2025 rate case began the redesign in earnest. Three changes matter most for compute economics.

First, a new large-load category above 50 megawatts faces a separate tariff schedule with explicit scarcity pricing during winter peak hours, lifting the effective rate by roughly 25 to 35 percent for that band. Second, capacity charges have been restructured to reflect the cost of the deferred generation needed to serve incremental demand, adding a fixed component that materially changes payback for low utilization workloads. Third, interruptibility is now mandatory for new contracts in this band, with curtailment events sized at up to 100 hours per winter season and compensation tied to the exported equivalent revenue Hydro-Quebec foregoes.

The composite effect is that an AI training campus signing in 2026 should model an effective Rate L equivalent of 7.5 to 9.5 cents Canadian per kilowatt hour over the contract life, plus a curtailment risk discount on annual energy that can range from 1 to 4 percent depending on cooling and workload flexibility. That is still cheaper than Northern Virginia or most European nodes, but the Quebec discount has compressed by roughly half versus the pre-2024 baseline.

Comparative siting economics: Quebec, Texas, Virginia, Iceland #

Tariff and queue changes only matter against the alternatives. The four nodes that dominate the 2026 hyperscaler training cluster shortlist are Montreal and the Quebec corridor, ERCOT West Texas, Northern Virginia and the broader PJM footprint, and Iceland. Each trades off power cost, latency to major user populations, climate cooling load, and regulatory predictability. The table below summarizes the comparison on the dimensions that matter for a 200 to 500 megawatt training campus signed in 2026 for energization in 2028 to 2029.

The economics are no longer obvious. Quebec retains a clean power and climate advantage but loses on queue certainty and tariff trajectory. ERCOT offers the fastest interconnection and deepest renewable build but exposes operators to ancillary service volatility and water stress. PJM is structurally tight on capacity through the late 2020s with capacity auction prices reflecting that scarcity. Iceland is a niche answer for batch training that tolerates high latency to North American users.

Quebec sovereignty politics and the foreign hyperscaler question #

The economic story sits inside a political one. Quebec's relationship with foreign capital around natural resources has always been conditional. Hydropower, mining royalties, and lately rare earths and graphite have all gone through cycles of openness and retrenchment. The current cycle is retrenching. The Coalition Avenir Quebec government, the Parti Quebecois opposition, and most of the Quebec Liberal caucus agree that Quebec power should preferentially serve Quebec-headquartered or Quebec-anchored projects. That consensus is durable across the 2026 to 2030 electoral horizon.

Hyperscaler campuses are the visible flashpoint. A 300 megawatt Microsoft or Google site creates a few hundred construction jobs and perhaps 50 to 100 permanent operations roles, while consuming power that could electrify tens of thousands of homes or anchor a domestic battery plant. The political math is unfavorable, and the Ministry of Economy has been explicit that future approvals will weigh local supply chain commitments, French-language workforce development, and visible technology transfer to Quebec institutions including Mila and the IVADO network.

This is not a closed door. It is a higher and more legible bar. Hyperscalers willing to anchor research partnerships, locate French-language model work in Montreal, and accept curtailment terms can still get approved. Hyperscalers expecting to drop a hyperscale campus on the same terms that worked in 2021 will not.

What hyperscalers should plan for #

We see five planning shifts that follow from the 2026 picture. First, treat Quebec as a constrained allocation, not a default green-power answer. Anchor capacity assumptions in the allocated column of Table 1, not the announced column. Second, model Rate L at the 7.5 to 9.5 cent Canadian per kilowatt hour band, with explicit curtailment scenarios at 50, 100, and 150 hours per winter, and validate cooling and workload flexibility against those bounds.

Third, design siting portfolios rather than single-node bets. The 2026 to 2028 AI training buildout should assume that no single node will absorb more than 30 to 40 percent of new capacity, with the remainder distributed across ERCOT, PJM-adjacent secondary markets, and selected international nodes including Iceland, Norway, and emerging Gulf locations. Fourth, invest early in the political and research footprint that makes Quebec approvals viable. Mila partnerships, French-language model work, and visible Quebec procurement are now part of the diligence file, not adjacent marketing.

Fifth, watch the export contract calendar. The Hydro-Quebec firm export contracts to New York and New England roll through the late 2020s and early 2030s. Renegotiation outcomes will set the marginal opportunity cost of Quebec power and therefore the ceiling on what new domestic load can pay. A favorable renegotiation tightens the gate further; an unfavorable one creates a narrow window for incremental allocation. Either way the answer is to plan now and avoid the assumption that 2021 economics return.

Athena and Promethean run this analysis at the campus level and the portfolio level. If you are evaluating a Quebec allocation, restructuring a stranded queue position, or building a multi-node siting strategy for the next training generation, /engage and we will scope a focused diligence sprint.

Sources #

• Hydro-Quebec 2035 strategic plan and annual report
• Regie de l'energie de Quebec rate case filings
• Quebec Ministry of Economy, Innovation and Energy
• IEA Electricity 2025 report
• Uptime Institute global data center survey
• Reuters coverage of Hydro-Quebec data center policy
• Bloomberg coverage of North American AI power siting
• Financial Times coverage of hyperscaler power constraints