The Demand Shock: Data Centres, Electrification, and the Coordination Gap

The Collision

For roughly twenty years, the demand side of the Canadian grid was the quiet side. National generation in 2024 — about 609,000 GWh — was not materially different from what the country produced in the mid-2000s (CER; StatCan). Efficiency gains, deindustrialization in central Canada, and the offshoring of energy-intensive manufacturing cancelled out population growth almost exactly. Utilities learned to plan for a flat line, regulators learned to litigate a flat line, and an entire generation of load-forecasting practice calcified around the assumption that next year would look like last year plus or minus weather. That assumption is now dead on both sides of the border. In the United States, the five-year national load-growth forecast embedded in utility regulatory filings quintupled in under two years — from roughly 23 GW of expected growth to 128 GW — the fastest upward revision in the modern record (Grid Strategies, 2024). NERC’s 2024 Long-Term Reliability Assessment recorded the largest year-over-year jump in ten-year demand projections in the assessment’s history and flagged accelerating large-load additions as a primary reliability risk (NERC, 2024). The International Energy Agency put the global driver in one number: data centres consumed roughly 415 TWh in 2024, about 1.5% of world electricity, and are projected to approach 945 TWh by 2030 — more than Japan’s entire present consumption (IEA, 2025).

Canada is not watching this from a distance; it is a preferred destination for it. By mid-2025 the Alberta Electric System Operator was carrying more than 16 GW of large-load interconnection requests — overwhelmingly data centres — against a system whose all-time peak demand is roughly 12.4 GW (AESO, 2025). Hydro-Québec has reported industrial and data-centre connection requests exceeding 10 GW against a system built around a winter peak near 43 GW, and has moved from courting such load to rationing it (Hydro-Québec, 2024). Ontario’s government, introducing legislation in June 2025 to give the province control over which data centres connect at all, cited connection requests on the order of 6.5 GW (Government of Ontario, 2025). British Columbia suspended new cryptocurrency-mining connections outright and is running its first competitive call for new power in fifteen years partly because large new loads arrived faster than supply (BC Hydro, 2024). Summed naively, announced and requested data-centre load in Canada is comparable to the entire winter peak of Québec — the largest single system in the country. Summed carefully, as this document will argue, it is something else entirely.

The second wave is slower, larger, and — unlike the first — politically committed. The federal Electric Vehicle Availability Standard requires zero-emission vehicles to reach 20% of new light-duty sales in 2026, 60% in 2030, and 100% in 2035 (Transport Canada, 2023). Federal and provincial programs are pushing heat pumps into a housing stock that heats, in most of the country, with gas and oil (NRCan, 2024). Industrial electrification — electric-arc steelmaking in Hamilton and Sault Ste. Marie, hydrogen electrolysis proposals in Atlantic Canada, electrified LNG compression on the BC coast — adds loads that are individually the size of small cities (ECCC, 2022; CER, 2023). The Canada Energy Regulator’s net-zero scenarios have national electricity demand roughly doubling by 2050 (CER, 2023). None of this is speculative in the way a hyperscaler’s press release is speculative. It is statute, regulation, and committed capital — and it lands disproportionately on the winter peak, which is precisely where Canadian systems are tightest.

What makes this a coordination problem rather than merely a construction problem is a structural asymmetry: the loads are mobile and the grids are not. A hyperscale campus can be sited in Alberta, Ontario, Québec, Manitoba, or Ohio; the developer will file connection requests in several of them simultaneously and build in one, or none. But under Section 92A of the Constitution each provincial system evaluates its slice of that request in isolation, with no visibility into the duplicate filings sitting in a neighbour’s queue, no common definition of what counts as a committed load, and no mechanism to reconcile the aggregate (Constitution Act, 1867; C.D. Howe Institute, 2025). The same megawatt of speculative demand can appear in three provincial plans at once — and trigger real transmission spending in all three. Meanwhile the electrification wave arrives through millions of uncoordinated household decisions whose aggregate is knowable only statistically, on systems that increasingly share weather, and therefore share peaks, across seams that were documented in this programme’s founding brief as some of the thinnest in the industrialized world (BGL/HG—01, 2026).

This document maps the demand shock end to end: the flat decades that disarmed the planning apparatus; the mechanics of the machine load and of the queues it has flooded in Alberta, Ontario, and Québec; the siting calculus that decides where it lands; the phantom-load problem — the gap between announced and actual — that corrupts every forecast built on top of it; the electrification trajectories and the winter-peak collision; the allocation regimes other jurisdictions have improvised; and the reason probabilistic load forecasting, whatever its operational merits, structurally fails the governance test these decisions now face. It also reports, in summary form, an internal result. Since June 2025 the lab has maintained a load-integrity register cataloguing every discoverable Canadian data-centre commitment — 312 entries as of the June 2026 release — against evidence of land, queue position, contract, and construction (BGL Load-Integrity Register r6, internal). The register’s headline finding: announced Canadian data-centre load exceeds plausibly deliverable interconnection capacity through 2032 by a factor of approximately 4.7 (BGL Memorandum LIR-7, internal). The demand shock is real. The number attached to it is not. The gap between those two statements is where this document lives.

Two Flat Decades

To understand why the planning apparatus was caught flat-footed, it helps to be precise about how flat the flat years were. United States electricity consumption hovered near 4,000 TWh from roughly 2005 to 2020 — fifteen years of effectively zero net growth — as LED lighting, appliance standards, and industrial efficiency absorbed everything population and GDP added (EIA, 2024). Canada followed the same contour with a colder accent. Ontario’s all-time peak demand, 27,005 MW, was set on August 1, 2006; for most of the following two decades the province never came within 3,000 MW of it, and grid-connected consumption fell outright as manufacturing left and conservation programs matured (IESO). Québec grew modestly on the strength of heating load and aluminum; Alberta grew with oil-sands cogeneration; but at the national level, generation in the low-600 TWh range became a fixture (StatCan; CER). A planner who joined a Canadian utility in 2005 and retired in 2023 could have spent an entire career without once seeing sustained, structural load growth.

Flatness had institutional consequences that matter now. First, forecasting atrophied into extrapolation. When the underlying process barely moves, naive models are indistinguishable from sophisticated ones, and the discipline’s energy went into refining weather normalization and end-use decomposition rather than into the harder problem of discrete, lumpy, behaviourally-driven load additions — the academic literature of the period is dominated by probabilistic refinements to an assumed-stationary process (Hong & Fan, 2016; Hong et al., 2016). Second, the regulatory reflex inverted. Two generations of hearings trained intervenors and rate boards to treat utility demand forecasts as padded — the historical sin was over-forecasting to justify capital — so the institutional immune system points the wrong way when genuine growth appears: it attacks high forecasts precisely when high forecasts have become plausible (Grid Strategies, 2024). Third, the interconnection machinery was never engineered for volume. Queues designed as accounting formalities for a trickle of projects became the primary rationing device for a flood; the US backlog reached roughly 2,600 GW of generation and storage capacity by 2024 — more than twice the installed US fleet — with typical waits of five years and historical completion rates near one project in seven (LBNL, 2024). FERC Order 2023 was an explicit admission that the process itself, not the engineering, had become the constraint (FERC, 2023).

The whiplash, when it came, was concentrated in exactly the institutions least equipped to absorb it. Utility integrated resource plans filed in 2022 and refiled in 2024 show demand trajectories that diverge within the first five years by more than the total growth the previous plan projected over twenty (Grid Strategies, 2024; NERC, 2024). PJM’s capacity market delivered the price signal in one auction: the 2025/26 Base Residual Auction cleared at $269.92 per megawatt-day, roughly nine times the prior year, and the 2026/27 auction cleared at its administratively capped price — a combined consumer cost increase measured in the tens of billions of dollars, driven substantially by data-centre load growth colliding with retiring supply (PJM, 2024; PJM, 2025). Virginia’s legislative auditor, reviewing the world’s densest data-centre market, concluded that unconstrained demand in the state could roughly double within a decade and that existing ratepayer-protection mechanisms were not designed for load of this character (Virginia JLARC, 2024). These are not exotic foreign case studies; PJM is Ontario’s largest trading counterparty, and Virginia is where several of the operators now filing in Canadian queues learned their siting playbook.

Canada’s own institutional record of the whiplash is visible in revision histories. The IESO’s planning outlooks moved in three years from managing surplus baseload to projecting demand growth of roughly 75% by 2050 — the fastest sustained growth the province has contemplated since the 1980s — with data centres named among the fastest-growing components (IESO, 2025). The AESO’s long-term outlook, which for years treated large-load additions as a rounding term, now carries a dedicated large-load connection process created under time pressure in 2025 (AESO, 2025). Hydro-Québec’s Action Plan 2035 — an additional 8,000 to 9,000 MW of capacity and investment guidance in the range of $155–185 billion — is, among other things, a public confession that the previous decade’s planning envelope was too small for the demand now presenting itself (Hydro-Québec, 2023). The lab’s comparative read of eleven Canadian planning documents across their 2021–2025 vintages found that the median five-year-ahead demand revision between successive editions exceeded the entire forecast error band the earlier edition had published (BGL Engineering Note EN-118, internal). The instruments were not wrong at the margin. They were wrong in kind — built to measure drift, presented with steps.

One further feature of the flat decades deserves emphasis because it shapes everything downstream: flatness made the seams tolerable. When no province was growing, the thinness of inter-provincial ties documented in this programme’s founding brief — a 150 MW west–east converter here, a derated intertie there — was an inefficiency, not an emergency (BGL/HG—01, 2026; CES-Energy). Each system could balance itself because each system’s problem was small. Step-change demand breaks that tolerance asymmetrically: the provinces with the most attractive siting economics receive the load requests, while the provinces with surplus energy or surplus capacity sit across seams too thin to help. A demand shock on a coordinated grid is a construction schedule. A demand shock on a fragmented one is an allocation fight — and Canada has entered it with no referee, no common ledger, and thirteen separate definitions of what a megawatt of committed demand even means (C.D. Howe Institute, 2025; CEAC, 2024).

The Machine Load

The proximate driver of the shock is a new class of electrical load with no real precedent in utility experience: computing campuses that arrive in units of hundreds of megawatts, run at near-constant output around the clock, and are sited by a handful of global firms making portfolio decisions at continental scale. The aggregate numbers are established. Data centres consumed an estimated 415 TWh globally in 2024 — around 1.5% of world electricity — and the IEA’s base case reaches approximately 945 TWh by 2030, with AI the dominant marginal driver (IEA, 2025). In the United States, the reference national accounting put data centres at 4.4% of electricity consumption in 2023, rising to between 6.7% and 12% by 2028 depending on hardware shipment and utilization scenarios (LBNL, 2024). EPRI’s scenario band for 2030 spans 4.6% to 9.1% of US consumption — a spread whose width is itself the finding: the credible range of outcomes differs by more than the entire load of most countries (EPRI, 2024).

The physics underneath is straightforward and worth stating plainly, because it explains both the scale and the uncertainty. Training compute for frontier AI models has doubled roughly every six months since 2010 — orders of magnitude faster than Moore’s Law — which means demand growth is set by capital deployment decisions, not by any organic diffusion curve (Sevilla et al., arXiv:2202.05924). A single modern accelerator draws on the order of 700 W; racks that drew 5–10 kW in the enterprise era now ship at 80–120 kW and above for AI training, an order-of-magnitude densification that converts a warehouse into a small power plant (Uptime Institute, 2024). The efficiency counterargument is real but historically bounded: through the 2010s, global data-centre compute grew roughly 550% while energy use grew about 6%, as hyperscale consolidation and hardware gains absorbed demand (Masanet et al., Science, 2020). Whether that absorption repeats in the AI era is the single largest uncertainty in every forecast cited above — and training-efficiency shocks, in which published models achieve comparable capability at a fraction of assumed compute, have already moved markets and revised corporate siting plans mid-flight (IEA, 2025). A load class whose ten-year trajectory can be revised 30% by a single research publication is not a load class amenable to trend extrapolation.

The capital behind it, however, is not hypothetical. Combined capital expenditure by the four largest hyperscale operators exceeded US$200 billion in 2024, with public guidance above US$300 billion for 2025 — the majority directed at AI data centres and the power to feed them (company filings; IEA, 2025). The Stargate venture announced in January 2025 attached a US$500 billion figure to a single program of AI infrastructure (OpenAI, 2025). And the operators have begun buying generation directly, in transactions that would have been unthinkable to a utility planner five years ago: an 835 MW power purchase agreement to restart the Three Mile Island Unit 1 reactor for Microsoft (Constellation, 2024); a Google commitment to purchase up to 500 MW from Kairos Power’s small modular reactors by 2035 (Google, 2024); Amazon agreements targeting more than 5 GW of X-energy SMR capacity by 2039 (Amazon, 2024); a Meta campus in Louisiana whose utility proposed roughly 2.2 GW of new gas generation to serve it (Entergy, 2024); and behind-the-meter gas turbines installed at AI sites faster than air permits could follow them (Reuters, 2024). The pattern matters for Canada: these firms have demonstrated they will procure power by any available route — grid, contract, or self-supply — and will arbitrage jurisdictions against each other to do it.

Three properties distinguish machine load from every load class utilities have planning instincts for, and each one breaks a different planning assumption. First, lumpiness: a single connection decision moves a provincial load forecast by whole percentage points, so the law of large numbers — the statistical foundation of load forecasting — simply does not apply; there is no ensemble to average over when the forecast question is whether one specific 400 MW campus signs (BGL Engineering Note EN-121, internal). Second, mobility before commitment, immobility after: until financial close, the load can relocate across the continent in response to a tariff ruling or a queue delay, which makes provincial forecasts strategically gameable by the applicants themselves; after energization, the load is a fixed industrial citizen for twenty years. The planning system must therefore price a real option it cannot observe (EPRI, 2024; Virginia JLARC, 2024). Third, flatness: at 85–95% load factor, a data centre adds nearly as much demand at the winter peak hour as at 3 a.m. in May — the exact inverse of the peaky-but-small residential loads Canadian winter systems were engineered around, and the property that makes the Québec allocation problem, treated below, so acute (Hydro-Québec, 2023; Norris et al., 2025).

The lab’s interest in this load class predates the Canadian queue flood. The register that became the load-integrity programme began in June 2025 as a verification exercise: could announced North American AI-campus commitments be reconciled against observable evidence — land transactions, air and water permits, interconnection filings, construction activity — using the same derivation discipline the stack applies elsewhere (BGL Run Ledger HG-3106, internal). The first tranche of sixty entries produced the result that motivated everything since: fewer than one announced megawatt in three could be tied to any legally binding instrument whatsoever (BGL Load-Integrity Register r1, internal). Announcements, it turned out, are not a noisy signal of future load. They are a different quantity altogether — part marketing, part negotiation, part optionality — and treating them as forecast inputs is a category error with billion-dollar consequences. Section 8 quantifies this for Canada. The three sections before it examine where the load is actually trying to connect.

Alberta: The Queue That Outgrew the Grid

Alberta became the epicentre of the Canadian data-centre rush for reasons that are legible in any site-selection spreadsheet: an energy-only market with no capacity obligations and no vertically integrated gatekeeper; abundant natural gas and a policy establishment comfortable with gas-fired self-supply; a cold, dry climate that ranks among the best free-cooling environments on the continent; cheap industrial land; and a provincial government that in December 2024 published an explicit AI data-centre attraction strategy with an ambition measured in the tens of billions of dollars of investment (Government of Alberta, 2024). The market structure matters most. In Ontario or Québec, a hyperscaler must negotiate with a Crown or quasi-Crown planner that weighs its request against every other public objective. In Alberta, historically, a load that paid its interconnection costs connected. The province marketed exactly this — speed to power as a competitive product — and the market responded at a scale nobody involved appears to have modelled.

By mid-2025 the AESO’s large-load interconnection queue carried more than 16 GW of requests, dominated by data-centre projects (AESO, 2025). The number needs its denominator, per the lab’s standing discipline: Alberta’s all-time system peak is roughly 12.4 GW, set during the January 2024 cold event that also produced emergency alerts and a near-miss on rotating outages (AESO, 2024). The queue, taken at face value, proposed to more than double the largest electricity system event in the province’s history — on a grid that only months earlier had issued public appeals to reduce consumption during a cold snap, and whose April 2024 wind-collapse emergencies were documented in this programme’s founding brief (AESO Market Surveillance Administrator, 2024; BGL/HG—01, 2026). No serious participant believed all 16 GW would build. But the system operator had no instrument to determine which fraction would, and every megawatt in the queue generated real study obligations, real transmission planning consequences, and real signalling to other applicants that the queue itself was the scarce asset worth claiming early.

The AESO’s response, announced in 2025, was an interim large-load connection approach that effectively rations near-term grid connections — an initial allocation on the order of 1,200 MW of data-centre connections through 2028, assigned by readiness criteria rather than first-come-first-served, with larger volumes deferred to subsequent phases as transmission and supply catch up (AESO, 2025). Two things about this deserve notice. First, the sizing: roughly 1,200 MW deliverable against more than 16,000 MW requested is an admission ratio near 7%, which is the operator’s own implicit estimate of the phantom-load fraction — strikingly consistent with the lab’s independent register-derived multiple reported in Section 8 (BGL Memorandum LIR-7, internal). Second, the governance form: an energy-only market that spent thirty years insisting the price signal allocates everything has, under load pressure, reinvented administrative allocation — a queue triage run on criteria the operator must now defend to disappointed applicants with lawyers. Alberta’s market design has no native language for this decision. It is being improvised, mid-shock, at the same time the province executes its transition to the redesigned restructured energy market — two foundational redesigns running concurrently on one operator (AESO, 2025).

The pressure valve is self-supply, and Alberta is where the continent’s off-grid experiments are running largest. The most publicized — the “Wonder Valley” proposal in the Municipal District of Greenview near Grande Prairie — sketches up to 7.5 GW of gas-fired, off-grid AI campus development over its full buildout, a private electricity system that would, if realized, rival more than half the provincial grid it bypasses (O’Leary Ventures, 2024). Smaller behind-the-fence gas proposals dot the queue. Self-supply relieves the interconnection bottleneck but exports its consequences: gas turbines procured outside utility planning still draw on the same turbine supply chains, the same gas transmission capacity, and the same airshed, while placing hundreds of megawatts of load-plus-generation adjacent to a grid that neither dispatches nor sees them — and the carbon accounting lands in the province’s industrial-emitter regime at exactly the moment federal clean-electricity regulation tightens the envelope (ECCC, 2024; Government of Alberta, 2024). The lab’s seam dossier for Alberta flags a further asymmetry: every gas-fired megawatt built for compute in Alberta is a megawatt of clean-supply demand that Québec, Manitoba, or BC hydro systems could in principle have served — across interties that do not exist at relevant scale (BGL Seam Dossier SD-AB-2, internal). The queue crisis and the intertie deficit are the same fact viewed from different provinces.

Alberta’s episode supplies this document’s cleanest specimen of the general problem. A jurisdiction with the continent’s most laissez-faire market design received the most demand, fastest, with the least ability to distinguish real projects from queue speculation — and resolved it by inventing, under duress, an allocation mechanism whose criteria, evidence, and fairness will be contested for years. The register’s Alberta panel — 96 tracked commitments as of r6 — shows announced load in the province exceeding the AESO’s own deliverable envelope through 2032 by the widest multiple of any province (BGL Load-Integrity Register r6, internal). What Alberta lacked was not generation, gas, land, or willingness. It lacked an instrument for proving which claims on the grid were real — which is not a market failure or an engineering failure, but an accounting failure. That distinction organizes everything in the second half of this document.

Ontario: New Load on a Committed System

Ontario’s version of the shock is shaped by what the province has already promised. The IESO’s planning outlook projects demand growth of roughly 75% by 2050 — from about 151 TWh toward the mid-200s — driven by industrial electrification, EV manufacturing and charging, housing growth, and data centres, which the operator names among its fastest-growing load segments (IESO, 2025). Against that growth stands the most heavily committed supply program in the country: the Darlington and Bruce refurbishment sequences running through the 2030s, the Pickering refurbishment approved at roughly $26.8 billion, the four-unit SMR program at Darlington, expanded procurements of storage and gas capacity, and transmission buildouts in the southwest and northeast — a pipeline documented in the founding brief and enumerated in the government’s integrated plan (OPG, 2026; Government of Ontario, 2025; BGL/HG—01, 2026). Ontario is not short of a plan. It is short of slack: nearly every megawatt in that pipeline was justified by a demand case that predates the data-centre wave, and the system’s spare deliverability in the 2027–2033 window — the exact window hyperscalers care about — is thin precisely because the refurbishment schedule removes large nuclear units from service in sequence throughout it (IESO, 2025).

The connection-request arithmetic made the collision explicit. By mid-2025 the province cited approximately 6,500 MW of data-centre connection requests — equivalent to restating Ontario’s entire nuclear refurbishment program as a single new customer class — and responded with legislation: the June 2025 energy statute gives the province authority to prioritize which large loads connect, explicitly ending default first-come-first-served treatment for data centres (Government of Ontario, 2025). The stated criteria — jobs, economic value, grid impact, and alignment with provincial priorities such as AI sovereignty — convert a queue position into a policy judgment. As in Alberta, the notable fact is the form of the solution: a wholesale market with locational marginal pricing at roughly 970 nodes, the most sophisticated price system in the country, concluded that prices could not make this decision and reached for administrative allocation instead (IESO, 2025; BGL Engineering Note EN-126, internal). Prices allocate energy well. They allocate connection rights on a committed system badly, because the marginal cost of a connection is dominated by lumpy transmission and supply consequences that arrive years later and land on other customers — which is an accounting problem before it is a market problem.

Geography sharpens it. The Greater Toronto Area is simultaneously the country’s largest load centre, its most latency-relevant metro for inference workloads, and one of its most transmission-constrained regions; requests concentrate exactly where deliverability is scarcest. Southwestern Ontario offers the instructive precedent: the greenhouse boom around Leamington and Kingsville — agricultural loads adding hundreds of megawatts in a single planning region — forced a decade of reactive bulk-transmission expansion and remains the IESO’s canonical example of concentrated load growth outrunning regional planning cycles (IESO; Hydro One). Data centres replay that dynamic at triple the unit size and a tenth of the lead-time tolerance. The lab’s Ontario panel places 74 tracked commitments in the register, with the highest “evidence-grade” fraction of any province — Ontario attracts more projects with land and permits actually in hand — but also documents systematic re-filing behaviour: at least eleven register entries appear in materially identical form in a second provincial queue, and four appear in three (BGL Load-Integrity Register r6, internal; BGL Run Ledger HG-3184, internal). Ontario’s planners see their queue. Nobody sees the union of queues.

The provincial politics of allocation deserve one more paragraph, because Ontario is where they are most explicit. Prioritizing data centres by economic value requires a defensible measure of economic value — jobs per megawatt, tax base per megawatt, strategic weight of sovereign compute — and every one of those measures is contested. A hyperscale campus employs perhaps 50–100 permanent staff per hundred megawatts; an automotive plant employs thousands for the same draw; the comparison is invidious but unavoidable once connection rights are rationed, and it will be litigated by whoever loses (Virginia JLARC, 2024, documents the identical debate in mature form). Meanwhile the federal government’s sovereign-compute agenda — a national AI compute strategy backed by roughly $2 billion — presses provinces to host exactly the loads their operators are rationing (ISED, 2024). The province is thus running three incompatible mandates at once: keep rates low, win AI investment, and protect a committed system. An allocation decision under three incompatible mandates is precisely the class of decision that must survive hostile audit — and the current instrument for it is a ministerial judgment informed by a probabilistic forecast, which is to say, an instrument that cannot be re-run by the losing party (BGL Memorandum M-2088, internal).

Ontario contributes one final structural fact to the national picture. Because its system is committed, its marginal response to unexpected load is gas — the fleet built as a bridge now runs harder with every forecast miss, a dynamic the founding brief quantified in carbon terms (IESO; ECCC, 2025; BGL/HG—01, 2026). Every phantom megawatt Ontario plans for and does not receive strands capacity paid by ratepayers; every real megawatt it fails to plan for is served, at the margin, by combustion. The cost of bad load accounting is therefore not symmetric, and it is not abstract: it is measured in stranded transmission on one side and in gas dispatch and rate increases on the other. A later section returns to this asymmetry as the core argument for why capacity accounting at the seams must be provable rather than probable.

Québec: From Open Door to Allocation

For twenty years Québec was the answer to the question this document asks. Hydro-Québec’s industrial rates have ranked among the lowest on the continent for as long as its own annual comparison of North American electricity prices has been published; the power was overwhelmingly hydroelectric and therefore nearly carbon-free; and the province actively marketed both properties to energy-intensive industry, from aluminum smelters to, in the 2010s, an early generation of cloud and colocation campuses around Montréal (Hydro-Québec; Montréal International). The first stress test arrived early: the 2018 cryptocurrency rush, when blockchain operators requested connections measured in thousands of megawatts within months. The response established the template the province is now scaling — the Régie de l’énergie authorized a dedicated, limited allocation block for the sector, with dissuasive tariffs outside it and curtailment obligations inside it (Régie de l’énergie, 2018). Québec learned six years before everyone else that some load classes must be rationed by rule rather than priced by market, and that the rationing rule itself becomes the contested object.

The AI wave broke the open-door model completely. Hydro-Québec has reported connection requests from data centres and other large industrials exceeding 10 GW — against a system whose historic winter peak stands near 43 GW and whose energy surpluses, the foundation of two decades of export revenue, have already narrowed under low-hydrology years and existing industrial commitments (Hydro-Québec, 2024; Hydro-Québec, 2023). The utility’s leadership has been unusually blunt in public: the era of selling discounted power to any comer is over, and allocation will be selective, favouring projects with high economic value per megawatt, alignment with decarbonization, and — notably — flexibility at the winter peak (Hydro-Québec, 2024). The legislative machinery followed: the 2024 energy-governance statute subjects large industrial allocations to government authorization and folds them into an integrated resource-planning process, formalizing what had been ad hoc ministerial discretion (Government of Québec, 2024). Québec, in short, has converted electricity from a commodity into an instrument of industrial policy — explicitly, statutorily, and ahead of every other province.

The physical logic behind the shift is worth making precise, because it is widely misread as protectionism. Québec’s constraint is not energy in the average year; it is capacity at the winter peak hour and energy in the dry year, and a flat 24/7 data-centre load consumes both at full weight. A megawatt of data centre presents at the February peak — when the system serves electric heating for a housing stock more heating-electrified than anywhere else in North America — with essentially no diversity benefit, while the same megawatt-hour sold at peak to New England or New York through the new HVDC ties earns a premium the domestic flat load forecloses (Hydro-Québec, 2023; BGL/HG—01, 2026). Action Plan 2035’s arithmetic — 8,000–9,000 MW of new capacity, roughly 60 TWh of new supply, $155–185 billion of investment — prices the marginal megawatt of new firm winter capacity at levels that make a low-value flat load a straightforwardly bad trade (Hydro-Québec, 2023). Selectivity is not politics contaminating economics; on a winter-peaking hydro system, it is the economics. The politics enter one step later, in deciding which projects clear the bar — and that decision currently rests on internal analysis no rejected applicant can inspect or re-run.

The register’s Québec panel is the most instructive in the country precisely because the province publishes the least. Sixty-one tracked commitments; the lowest announced-to-evidence ratio of any large province — Québec’s screening deters purely speculative filings — but also the largest single cluster of “refugee re-filings”: register entries that first appear in Québec, are declined or stalled, and resurface within two quarters in the Alberta or Ontario queue with capacity figures intact and commercial narratives lightly edited (BGL Load-Integrity Register r6, internal; BGL Run Ledger HG-3197, internal). This is the coordination gap in miniature. Québec’s rigor does not extinguish phantom load; it exports it, un-flagged, to provinces with weaker screens — because no mechanism exists for one province’s determination about a project’s reality to travel with the project. Each operator re-derives, from scratch and in confidence, facts a neighbour has already established. The lab’s memorandum on the phenomenon labels it serial credulity, and estimates that a third of the gross capacity in western and central queues has already been evaluated, in some form, by another Canadian jurisdiction (BGL Memorandum M-2093, internal).

Québec also supplies the clearest preview of where every province ends up: allocation by criteria, criteria under dispute. Aluminum smelters hold decades-old contracts at legacy rates; new entrants ask why incumbency outranks jobs-per-megawatt; regional development agencies ask why Montréal’s latency economics should outrank the regions’ unemployment; the export desk asks why any domestic flat load outranks the New England peak premium; and the sovereign-compute lobby asks why commercial arithmetic outranks national capability (Government of Québec, 2024; ISED, 2024). These are legitimate, incommensurable claims, and a governance process will rank them — the only question is whether the ranking rests on analysis the losers can verify. Hydro-Québec’s advantage is that, as a single integrated utility, its internal accounting can at least be made consistent. At the seams — where a project declined in Québec becomes Alberta’s gas-fired problem, or where a Québec allocation forecloses an export that Ontario’s planners had assumed — no consistent accounting exists at all. That is the layer this programme studies.

The Siting Calculus

Where machine load lands is not mysterious; it is the output of a calculus the industry runs openly, and understanding its terms is prerequisite to predicting — or governing — the queue. The terms have shifted. A decade ago the ranking variables were power price, tax treatment, and fibre. Today the industry’s own surveys are unambiguous that the binding variable is speed to power: the calendar time from application to energized megawatts, which now varies across North American jurisdictions from under two years to over seven and dominates every other term because an idle AI campus is a depreciation event measured in millions of dollars per week of delay (Uptime Institute, 2024; LBNL, 2024). This inversion explains most of the Canadian queue geography. Alberta’s flood was a speed story; Québec’s screening pushed speed-sensitive applicants elsewhere; Ontario’s attraction survives its congestion only because of the second variable — latency.

Latency partitions the load class in a way that matters enormously for provincial planning and is routinely ignored in it. Model training is latency-tolerant: a training cluster interacts with its operators in checkpoints, not milliseconds, and can sit anywhere power, cooling, and fibre backhaul exist — which is why gigawatt-scale training proposals cluster in remote, gas-adjacent, cold-climate locations like the Alberta northwest (Government of Alberta, 2024). Model inference — serving users — inherits the physics of fibre, roughly five milliseconds of round trip per thousand kilometres, and therefore pulls toward population: the GTA, Montréal, metro Vancouver (BGL Engineering Note EN-131, internal). The planning consequence: training load is footloose and negotiable — it can, in principle, be steered toward surplus and curtailability — while inference load is geographically stubborn and will contest exactly the constrained urban transmission that housing growth and building electrification also need. A provincial allocation policy that does not distinguish the two — and none currently does, in public form — is allocating two different commodities with one rule (BGL Memorandum M-2101, internal).

Climate is Canada’s genuine edge, and it is quantifiable. Cooling overhead is captured in power-usage-effectiveness, whose global industry average has stalled around 1.56–1.58 for years; cold-climate free-air and economizer operation delivers design PUEs materially below that for most of the Canadian shield and prairie winter, a structural single-digit-percentage energy saving on loads where single digits are hundreds of gigawatt-hours (Uptime Institute, 2024; ASHRAE). Water cuts the other way and has become a first-order siting screen since the operators’ own disclosures put hyperscale evaporative consumption in the billions of gallons annually and rising — figures that generated local backlash from Arizona to the Netherlands (Google, 2023; Microsoft, 2023). Cold, water-rich, or dry-coolable Canadian sites hold real cards here. But the register documents a recurring pattern the promotional literature omits: announced Canadian campuses frequently cite free cooling in their public materials while their permit filings specify evaporative systems sized for summer peaks — the climate advantage claimed in the press release and surrendered in the engineering (BGL Load-Integrity Register r6, internal; BGL Engineering Note EN-133, internal). Marketing PUE and permitted PUE are different numbers; only one of them is load.

The remaining terms are fiscal and political, and they interact with federation structure in ways provinces individually underprice. Tax abatements, industrial rate designs, and land assembly are conventional; the newer terms are sovereignty and stability. The federal sovereign-compute strategy — roughly $2 billion behind domestic AI compute capacity — plus procurement preferences and data-residency requirements create a policy premium for Canadian siting that did not exist in 2022, and domestic carriers have moved: Bell’s AI Fabric program and Telus’s sovereign AI facilities in BC and Québec represent the first substantial Canadian-owned entries in a market otherwise sited by foreign hyperscalers (ISED, 2024; Bell, 2025; Telus, 2025). Stability, meanwhile, is the quiet variable in every 2025-vintage siting memo: tariff volatility and permitting whiplash south of the border have measurably increased the option value of Canadian sites, a fact provincial negotiators appear not to have internalized in the terms they offer (IEA, 2025; BGL Memorandum M-2104, internal). Canada is being shopped precisely because it is boring. Boring is worth basis points, and nobody at the table is charging for it.

Run the full calculus and the map draws itself — and draws the coordination gap with it. Training load wants Alberta gas and prairie cold; inference wants the GTA and Montréal; sovereignty wants Canadian flags; water screens favour the north and the dry south-west of the prairies; speed favours whoever’s queue moves fastest this quarter. No single province sees more than its own fragment of an optimization the applicants run continentally. The result, visible in the register, is systematic multi-queue arbitrage: the same corporate family holding positions in three provinces, sized not to build in all three but to preserve the option of building in the one that clears first — rational for the applicant, ruinous for planners who each count the position as prospective load (BGL Load-Integrity Register r6, internal). The siting calculus is not the pathology. The pathology is that thirteen systems are each solving one row of a matrix the applicants solve whole.

Phantom Load: The Announced and the Actual

Every number in the preceding sections — Alberta’s 16 GW, Ontario’s 6.5 GW, Québec’s 10 GW — shares a defect: it counts requests, and requests are free. The generation side of the industry has already run this experiment at continental scale and published the answer. Of all generation capacity that entered US interconnection queues over the last two decades, historically only about one project in seven or eight reached commercial operation; the queue is a lottery-ticket dispenser, not a plan (LBNL, 2024). There is no equivalent public accounting for large loads — load interconnection has never had a Queued Up — but the incentive structure is strictly worse. A generation developer must eventually post interconnection deposits scaled to real costs; a load applicant, in most jurisdictions through 2024, could reserve hundreds of megawatts of a utility’s planning attention with little more than a letter and a refundable fee. US utilities began confronting this openly in 2024: AEP Ohio’s proposed data-centre tariff — minimum-demand charges obligating large loads to pay for most of their requested capacity whether or not they use it, with exit fees and multi-year terms — was an explicit attempt to make queue positions costly enough to be informative, and versions of it are propagating through US rate cases (AEP Ohio, 2024; Virginia JLARC, 2024). The tariff innovation is telling: the industry’s first instinct for separating real from phantom load is to price the lie. It helps. It does not produce an accounting.

The lab’s load-integrity register was built to produce one, for Canada, by the only method that survives audit: evidence, graded. Every discoverable Canadian data-centre commitment since mid-2023 — corporate announcements, provincial attraction-strategy citations, municipal land dealings, queue disclosures, permit filings, procurement notices — is captured as an entry and placed on a six-rung evidence ladder: L0 announcement only; L1 land secured (title, option, or lease observable); L2 interconnection position verifiable in an operator process; L3 binding instrument (energy-services agreement, tariff contract, or authorized allocation); L4 major permits and procurement (air, water, environmental, long-lead electrical equipment); L5 construction or energization observable (BGL Load-Integrity Register, methodology annex, internal). Rungs are evidentiary, not judgmental: an entry climbs only on documents, each document hash-logged with retrieval date, and every entry carries expiry conditions — an L0 announcement uncorroborated for four quarters is flagged dormant rather than deleted, because dormant announcements have a documented habit of resurrection in new jurisdictions (BGL Run Ledger HG-3197, internal). As of release 6 — June 2026 — the register holds 312 entries totalling roughly 29 GW of announced or requested capacity: 96 in Alberta, 74 in Ontario, 61 in Québec, and 81 across BC, the prairies, and Atlantic Canada (BGL Load-Integrity Register r6, internal).

The distribution across the ladder is the finding. Roughly 58% of announced capacity sits at L0–L1 — press and land, nothing more. Less than 14% has reached L3, the first rung at which a megawatt is attached to any legally binding instrument, and under 6% shows construction (BGL Load-Integrity Register r6, internal). Cross-referencing entries against corporate families and technical fingerprints — identical capacity blocks, identical phasing language, shared engineering consultants — identifies at least 41 entries as probable duplicates of positions held in other provinces, roughly 6.1 GW of double- and triple-counted intent (BGL Run Ledger HG-3184, internal). Setting duplicated capacity to single-count, discounting each rung by empirically observed rung-to-rung survival over the register’s own thirteen quarters of history, and capping the result by each province’s stated deliverable interconnection envelope through 2032 — the AESO’s phased allocation, Ontario’s prioritized envelope, Hydro-Québec’s authorized-allocation pace — yields plausibly deliverable Canadian data-centre load of approximately 6.2 GW by 2032. Announced load exceeds it by a factor of approximately 4.7 (BGL Memorandum LIR-7, internal). The multiple is not uniform: Alberta’s panel runs above 6, Ontario’s near 3.4, Québec’s — screening at the door — below 2.5 (BGL Load-Integrity Register r6, internal). The lab publishes the method and the multiple, not the roster; entries name counterparties, and the register exists to be audited, not to embarrass.

Two objections deserve engagement because both are partly right. First: so what — everyone knows announcements are inflated. Everyone knows it qualitatively; nobody prices it consistently, and the asymmetry of errors makes vibes-level discounting dangerous. A province that over-plans builds transmission and supply that ratepayers carry for decades — the JLARC inquiry’s central anxiety (Virginia JLARC, 2024). A province that under-plans loses the investment to a neighbour and serves what does arrive with marginal gas — Ontario’s asymmetry, quantified above. When the same phantom megawatt inflates three provincial plans simultaneously, the federation over-plans in aggregate even while each province believes itself conservative — a fallacy of composition running on real capital (BGL Memorandum M-2093, internal; CEAC, 2024). Second: the register could be wrong. It could; specific entries assuredly are. The claim is not omniscience — it is that every rung assignment traces to a hash-logged document any counterparty could inspect, every discount factor to a published survival table, and every provincial cap to an operator’s own statement, so a challenger must name the entry, the document, or the factor they dispute (BGL Load-Integrity Register, methodology annex, internal). Disagreement becomes tractable. That property — not the 4.7 — is the point. The multiple will move with every release; the discipline that produces it is the deliverable, and Section 12 argues no probabilistic construction can substitute for it.

A closing observation on dynamics. Phantom load is not static noise on top of real demand; it is reflexive. Queue congestion signals scarcity; scarcity raises the option value of queue positions; higher option value attracts more speculative filings; the queue grows, and operators respond — rationally — with caps and triage that raise the value of incumbent positions further. Alberta’s 1,200 MW allocation instantly converted every earlier queue position into an appreciating asset and every criterion in the triage into contested terrain (AESO, 2025). The registry literature on generation queues documented the same spiral for a decade before reform (LBNL, 2024; FERC, 2023). Load queues are earlier on the same curve, with lumpier stakes and less transparency. The window in which Canada can install honest load accounting — before incumbent queue positions harden into property — is the current one, and it is closing at the speed of the next allocation decision (BGL Memorandum LIR-7, internal).

The Second Wave: Wheels, Heat, and Industry

If machine load is the shock’s loud half, electrification is its heavy half — smaller headlines, larger integrals, and a fundamentally different statistical character. Start with transport. Zero-emission vehicles reached roughly one in seven new light-duty registrations in Canada in 2024, led by Québec and BC where provincial mandates and rebates run strongest (StatCan, 2025). The federal Electric Vehicle Availability Standard fixes the trajectory in regulation — 20% of sales in 2026, 60% in 2030, 100% in 2035 — with Québec and BC statutes running the same road (Transport Canada, 2023; Government of Québec; Government of BC). The energy arithmetic is undramatic: a light-duty EV draws roughly 3 MWh a year, so even five million EVs — a plausible mid-2030s fleet — add on the order of 15 TWh, under 3% of national consumption (CER, 2023). The capacity arithmetic is the real story: an EV’s 7–19 kW home charger is one to two orders of magnitude above the diversified draw of the house it parks at, so the peak impact is almost entirely a question of coincidence — and coincidence is a policy variable. Ontario’s ultra-low overnight rate class exists precisely to buy that coincidence down, and managed-charging pilots across North America consistently show the overwhelming majority of charging is shiftable off-peak when priced or automated (OEB, 2023; IESO, 2025). EV load is, in the lab’s taxonomy, large, slow, and steerable — the best-behaved wave in the shock, provided the steering instruments are actually deployed at scale rather than piloted indefinitely (BGL Engineering Note EN-140, internal).

Heating is the opposite case: moderate energy, brutal capacity, and steerable only within the physics of staying warm. Heat pumps have moved from marginal to mainstream — they now lead new heating-equipment sales in the US and their Canadian stock share is climbing from a base concentrated in Atlantic Canada and Québec, pushed by the Greener Homes and Oil-to-Heat-Pump programs and provincial equivalents (NRCan, 2024; StatCan). The engineering nuance that planning cannot ignore: a cold-climate heat pump’s coefficient of performance degrades as temperature falls, and below its balance point — commonly in the −15 to −25 °C range for modern cold-climate units — auxiliary resistance heat carries the difference at COP 1 (NRCan, 2024). A house that electrifies heating adds modest annual energy and a large draw precisely at the coldest hour of the year — the design hour. Québec, the continent’s most heating-electrified system, already lives this: its record peak near 43 GW is a temperature reading as much as a load reading, and every additional electrified furnace across the country recruits its owner into the same correlated event (Hydro-Québec; ECCC). Unlike EV charging, the coldest-hour draw cannot be shifted — only insulated against, backed up, or served. Building-envelope programs are, in capacity terms, peak-shaving infrastructure, and are almost never accounted as such (Efficiency Canada, 2024; BGL Engineering Note EN-141, internal).

Industrial electrification is the wave with the largest single steps and the strongest policy anchors. The steel conversions are committed capital: the Hamilton and Sault Ste. Marie transitions from blast furnaces toward electric-arc and direct-reduction routes, federally co-funded, each add loads in the hundreds of megawatts and remove megatonnes of process emissions — loads as lumpy as data centres but with decades-long commitment horizons and thousands of jobs attached, which is exactly the comparison Ontario’s allocation criteria will be forced to make explicit (ECCC, 2022; Government of Canada, 2022). Hydrogen electrolysis proposals in Atlantic Canada — the Newfoundland and Nova Scotia wind-to-hydrogen ventures advanced under the Canada–Germany Hydrogen Alliance — carry nameplate electrical demands in the gigawatts, dwarfing the host provinces’ existing peaks, though the register applies to them the same evidence ladder it applies to data centres and finds most below L2 (NRCan, 2022; BGL Load-Integrity Register r6, internal). On the west coast, electrified LNG compression — Cedar LNG and the second-phase electrification debate at LNG Canada — concentrates several hundred megawatts of flat industrial load at the far end of BC’s north-coast transmission, the driver behind BC Hydro’s north-coast expansion line and a live case of load arriving ahead of wires (BC Hydro, 2024; CER).

The three waves differ in exactly the dimensions that matter for planning, and collapsing them into one demand line — standard practice in every provincial outlook — destroys the information a coordination layer needs. Machine load is lumpy, fast, mobile before commitment, flat after it, and substantially fictional at the announcement stage. EV load is diffuse, slow, statistically forecastable with genuine confidence — mandates fix the fleet trajectory within narrow bands — and steerable at low cost. Heating load is diffuse, slow, forecastable in energy but savage in capacity, correlated across provinces by synoptic weather, and unsteerable at the design hour. Industrial load is lumpy, slow, policy-anchored, and honest — a steel plant does not file in three provinces. A single “demand forecast” summing these four processes is adding a contract, a diffusion curve, a weather integral, and a press release, then presenting one number with one error band (BGL Memorandum M-2108, internal; Hong & Fan, 2016). The seams inherit the confusion: what crosses an intertie at the February design hour depends on which wave dominates each side of it, not on the sum. Any accounting fit for allocation decisions must carry the waves separately, with their evidence classes attached — the design requirement Section 13 takes up directly.

The Winter Peak Problem

Every thread in this document converges on a single hour: the coldest weekday morning of a February arctic outbreak, somewhere between 07:00 and 09:00 local, when the Canadian grid’s design constraint actually binds. Québec’s record — roughly 43 GW — was set in such an hour; so were the recent peaks in Alberta, Saskatchewan, Manitoba, New Brunswick, and Nova Scotia, all winter-peaking systems, with Ontario and BC trending toward dual peaks as heating electrifies (Hydro-Québec; AESO, 2024; CER). The demand shock loads this hour asymmetrically. Data centres present at full draw — flatness means no diversity relief. Heat pumps present at their worst COP, some in resistance backup. EVs present with cold-degraded batteries whose owners charged less efficiently and want range margin. And the hour is continentally correlated: the synoptic systems that freeze Montréal freeze Toronto, Winnipeg, and Minneapolis within the same seventy-two hours, which is precisely when American markets bid hardest for the imports Canadian systems also want (ECCC; NERC, 2023). Winter Storms Uri and Elliott established what correlated cold does to modern interconnections — generation failures, gas-supply failures, and demand forecast misses arriving together, with the FERC/NERC inquiries documenting demand underestimation as a recurring contributor (FERC/NERC, 2021; NERC, 2023).

The reliability accounting for this hour has a technical name worth importing into the policy debate: effective load-carrying capability, the metric by which a resource’s — or a load’s — contribution at the constrained hour is measured, not its nameplate or its annual energy (NERC, 2024; IESO, 2025). Loads deserve the symmetric treatment they rarely get. A megawatt of data centre is a full megawatt at the design hour; a megawatt of nameplate EV charging is a small fraction of one under managed charging and most of one without it; a megawatt of new heat-pump connection is more than a megawatt at the design hour once backup resistance is counted, offset only by envelope quality. Provincial plans that sum these as interchangeable demand are mispricing the only hour that sizes the system. The lab’s reconstruction of the four largest provinces’ published outlooks — re-deriving design-hour contributions wave by wave under each province’s own stated adoption trajectories — found divergences between stated peak forecasts and wave-resolved arithmetic of up to 9% at 2035, in both directions, with the sign differing by province (BGL Run Ledger HG-3221, internal). Nine percent of a 43 GW system is a Bruce Power. The discrepancies are not accusations of error; they are demonstrations that the published instruments cannot be audited to the hour that matters, because the wave-level assumptions are not published at all (BGL Memorandum M-2112, internal).

The flexibility counterargument is the most important recent development in the literature, and it deserves both its due and its limits. The Duke analysis that reframed the US debate found that if new large loads accept curtailment for as little as 0.25–1% of annual hours — the top few dozen system hours — existing US grids could host on the order of 76–126 GW of new load with no new peak capacity at all: headroom hiding in the space between average and peak utilization (Norris et al., 2025). The insight transfers to Canada with force — the design hour is rare, and a load absent from it barely stresses the system. But three frictions temper it. Economics: an AI training campus at hyperscale capital intensity resists curtailment because idle accelerators are the most expensive idle capital in industrial history; operators promise flexibility in siting negotiations and procure diesel or gas backup to avoid honouring it — the register logs backup-generation permits alongside flexibility claims and finds them, at several sites, sized to identical megawatts (BGL Load-Integrity Register r6, internal). Verification: a curtailment commitment is only capacity relief if it is contractual, telemetered, and enforced at the hour of maximum system stress — a governance instrument, not a forecast assumption (Norris et al., 2025; AESO, 2025). And correlation: heating’s design-hour draw cannot be curtailed at all, so flexible-load headroom is real only net of the unsteerable wave rising underneath it (NRCan, 2024). Flexibility is genuine capacity. Claimed flexibility is phantom capacity with better branding — and distinguishing the two is, once again, an accounting problem.

What makes the winter peak the coordination gap’s sharpest expression is that the hour is shared while the planning is not. At the February design hour, Québec historically leans on interruptible industrial load and imports; New Brunswick and Nova Scotia lean on Québec and New England ties; Alberta’s January 2024 emergency was ended in part by emergency assistance across exactly the interties this programme has documented as chronically thin (AESO, 2024; BGL/HG—01, 2026). Each province’s resource-adequacy filing quietly assumes the seams deliver at the design hour; no national process verifies that the assumptions sum. The lab’s cross-filing reconciliation — aligning each jurisdiction’s assumed design-hour imports against every counterparty’s assumed exports across the same seam — found the familiar signature of uncoordinated optimism: paired assumptions that cannot simultaneously hold at three of the eight material seams, in magnitudes from 140 to over 600 MW (BGL Run Ledger HG-3226, internal). This is not a forecast disagreement. It is double-booked capacity at the binding hour, discoverable only by an entity that reads all the filings as one ledger — and no such entity exists in the Canadian federation. The demand shock will add gigawatts to both sides of these seams before 2032. The reconciliation gap scales with it.

Allocation Regimes Abroad

Canada is not the first jurisdiction to discover that data-centre load breaks first-come-first-served. Four regimes ran the experiment earlier, under different constraints, and their trajectories converge on one lesson. Ireland ran the uncontrolled version: data centres grew to 21% of national metered electricity consumption by 2023 — the highest national share on record anywhere — before the system operator effectively closed the Dublin region to new connections for the balance of the decade and the regulator imposed connection conditions requiring dispatchable on-site generation or storage and locational flexibility (CSO, 2024; EirGrid, 2022; CRU, 2021). The Irish sequence — open door, exponential growth, regional saturation, emergency criteria — took roughly a decade, and its most instructive feature is what the emergency criteria lacked: a transparent accounting basis. Connection decisions moved from a queue to a judgment, and the judgment’s inputs remained internal to the operator, leaving every declined applicant and every burdened community arguing against a black box (CRU, 2021). Ireland got control without getting legitimacy, and the political fight has never closed.

Singapore ran the deliberate version. A 2019 moratorium froze new data-centre construction entirely; the 2022 restart admitted projects through a competitive call scoring efficiency, decarbonization, and economic contribution against an explicit capacity budget, and the 2024 Green Data Centre Roadmap published the forward allocation — roughly 300 MW of near-term headroom with additional tranches conditioned on green energy procurement (IMDA, 2022; IMDA, 2024). Singapore’s regime is the world’s most explicit admission that connection rights are a scarce public resource to be allocated by criteria — and its transferable lesson is procedural: the criteria were published before the queue formed, the budget was stated as a number, and applicants competed against a rulebook rather than lobbying against each other. The Netherlands ran the reactive version — a national moratorium on new hyperscale facilities in 2022 after the Zeewolde controversy, with siting authority pulled upward from municipalities that had been negotiating alone against global counterparties (Government of the Netherlands, 2022). And Virginia — the world’s largest concentration, transmitting roughly the load of a mid-sized country through one county — ran the mature-market version: a legislative audit concluding that demand at announced pace could roughly double within a decade, that the resource costs are largely assigned to the general body of ratepayers under current rate design, and that the state possessed no instrument to distinguish which announced load was real (Virginia JLARC, 2024). PJM’s capacity prices delivered the arrears invoice for that ambiguity across thirteen states at once (PJM, 2024; PJM, 2025).

Reduced to their governance skeletons, the four regimes differ on two axes: whether allocation criteria were explicit or discretionary, and whether the accounting behind them was published or internal. Singapore sits in the explicit-published corner and has the least litigious aftermath; Ireland explicit-internal; the Netherlands discretionary-internal; Virginia, until its audit, effectively no regime at all. Canada’s provinces are presently distributing themselves across this grid without noticing the axes: Alberta’s interim triage is explicit-internal, Ontario’s statute is discretionary-internal, Québec’s authorization regime is discretionary-internal with the strongest screening, and the remaining provinces are pre-regime (AESO, 2025; Government of Ontario, 2025; Government of Québec, 2024; BGL Memorandum M-2117, internal). The comparative record predicts the consequences with some confidence: discretionary-internal regimes control load but bleed legitimacy — every rejection becomes a lobbying campaign, every approval a suspicion — while explicit-published regimes convert the same scarcity into a rulebook competition that survives its losers. No Canadian jurisdiction currently occupies the explicit-published corner. None has committed to publishing the capacity accounting on which its allocations rest.

The federation adds a failure mode none of the comparators faced: regime arbitrage across internal borders. Ireland, Singapore, and the Netherlands are single jurisdictions; Virginia at least shares PJM’s common queue discipline and FERC’s common tariff law with its neighbours. Canada’s ten regimes share nothing — not a queue format, not an evidence standard, not a definition of committed load, not a disclosure norm — so every tightening in one province redirects the phantom flood toward the loosest neighbour, a dynamic the register has now documented across three provincial pairs (BGL Run Ledger HG-3197, internal; BGL Load-Integrity Register r6, internal). Inter-jurisdictional shopping is rational and legal; what is neither is the federation’s failure to notice it in aggregate. The European Union, facing the same dynamic among member states, at least publishes harmonized adequacy assessments through ENTSO-E; Canada’s equivalent instruments — NERC’s continental assessments and the CER’s scenario work — aggregate provincial submissions without reconciling them, inheriting every double-count this document has described (ENTSO-E; NERC, 2024; CER, 2023). The comparative conclusion is uncomfortable but precise: every jurisdiction that faced the shock eventually built an allocation regime; the regimes that worked published their accounting; and no accounting can be published that the publisher cannot defend line by line. Which returns the argument to instruments.

Where Probabilistic Forecasting Fails the Governance Test

Nothing in this document is an argument against probabilistic load forecasting as an operational discipline. For the purposes it was built for — day-ahead scheduling, weather-driven variance, diversified end-use aggregates — the probabilistic toolkit is mature, rigorous, and demonstrably effective; the tutorial literature and the forecasting competitions that shaped the field are monuments of applied statistics (Hong & Fan, 2016; Hong et al., 2016). EV diffusion and heating electrification are, as Section 9 argued, genuinely statistical processes where these methods earn their keep. The argument is narrower and structural: the decisions the demand shock forces — who connects, in what order, backed by what transmission, across which seams — are governance decisions, and probabilistic forecasts fail four tests that governance imposes, regardless of how well-calibrated they are.

The first failure is ontological: the dominant uncertainty is not a random variable. Whether a specific 400 MW campus materializes is a contractual and behavioural fact in the hands of identifiable counterparties, not a draw from a distribution; there is no population of independent trials behind it, and the applicant controls the information asymmetry. Wrapping that uncertainty in a scenario weight — assigning, say, 40% probability to the high-load case — does not model the world; it launders a negotiation into a statistic (BGL Memorandum M-2108, internal; EPRI, 2024). Worse, the act of forecasting feeds back: a utility that publishes a high-load forecast justifies transmission that lowers connection costs, which attracts filings that validate the forecast — reflexivity the queue spiral of Section 8 exhibits in the wild (LBNL, 2024; Grid Strategies, 2024). Statistical forecasting assumes the forecast does not move the process. Queue-driven load violates that assumption by construction.

The second failure is the burial of decisions in priors. A probabilistic outlook must choose scenario weights, correlation structures, and which percentile to plan to — P50 or P90 peak, weather-year ensembles, load-factor assumptions for uncommitted projects — and each choice moves billions of dollars of consequences while being presented as a technical setting (NERC, 2024; IESO, 2025). When Virginia’s auditors or Alberta’s disappointed applicants ask why this number, the honest answer traces to an analyst’s judgment encoded as a parameter — unexaminable by the parties who bear the outcome (Virginia JLARC, 2024). The third failure is seam incoherence. Probabilistic outlooks produced independently by neighbouring systems are mutually inconsistent in exactly the ways Section 10 documented: percentiles do not sum, correlated tails are assumed away, and the same phantom project appears in two distributions at once. There is no statistical operation that reconciles two sovereign forecasts; reconciliation is a ledger operation, requiring entries, not densities (BGL Run Ledger HG-3226, internal; CEAC, 2024). The fourth failure is the audit test itself, and it is the decisive one: a losing counterparty cannot re-run a judgment. The reliability establishment’s own caution about opaque models in operations — the insistence on verifiability that NERC applies to machine learning — applies with equal force to any allocation resting on unpublishable ensembles (NERC, 2024; Venzke & Chatzivasileiadis, arXiv:1910.01624). The founding brief argued that markets solved this inside each ISO by making prices the dual variables of an explicit optimization — auditable by reconstruction (O’Neill et al., 2005; PJM Manual 11). The demand side of the seams has no equivalent instrument. That absence, not any deficiency of statistical craft, is the gap.

The failure pattern has a positive image: it specifies the instrument that would pass. It must treat lumpy commitments as evidentiary objects, not distributional ones — graded by documents, as the register grades them. It must expose every assumption as a named, contestable input rather than a fitted parameter. It must operate on the union of jurisdictions, so that a claim entered twice is caught structurally rather than statistically. And it must be re-runnable: given the same evidence base, any party’s analysts must reach bit-identical conclusions, so that disagreement attaches to evidence — where governance can adjudicate it — rather than to method, where it cannot. These are not aspirations; they are the defining properties of deterministic derivation over a governed corpus, and the next section describes the working system the lab operates to those specifications (BGL Engineering Note EN-149, internal). Probabilistic methods keep their place inside the scenarios — diffusion curves and weather ensembles are inputs a deterministic accounting happily consumes, labelled as such. What they cannot be is the arbiter. Distributions inform. Ledgers decide.

Provable Capacity Accounting

The lab’s answer to the instrument gap is deterministic capacity accounting: a ledger discipline in which every claim about future load and future deliverability is an evidentiary object with lineage, and every planning-relevant conclusion is derived from those objects by a reproducible computation. The load-integrity register is the demand-side half of that discipline already in operation. Its architecture is the lab’s standard stack — M.A.D. (Marixous Architectural Derivation) with the D.A.E. at its core — applied to a corpus of commitments rather than telemetry: ingestion renders each announcement, permit, queue disclosure, and contract instrument into a governed object model without altering source records; the store is append-only, so the evidence state that produced any conclusion is permanently reconstructible; and the derivation engine computes rung assignments, duplicate detection, survival discounting, and the aggregate multiple deterministically — same corpus, same conclusions, bit-identical, every run (BGL Load-Integrity Register, methodology annex, internal; BGL Run Ledger HG-3230, internal). When release 6 states that announced load exceeds deliverable interconnection by 4.7, that number carries a hash-chained derivation record tracing through every entry, every document, and every discount factor to the register’s floor. A challenger does not argue with the lab. A challenger re-runs the derivation and names the object they dispute.

Extended to the seams, the discipline produces what this programme calls seam capacity certificates: for a named interface — the Alberta–BC intertie, the Québec–New Brunswick corridor, the Ontario–Québec ties — a stated scenario, and a stated hour class, a derived accounting of deliverable capability net of every documented claim against it: ratings and deratings from operator filings, existing contractual reservations, each side’s design-hour import assumptions, and the wave-resolved load growth of Sections 9 and 10 on both sides (BGL Engineering Note EN-152, internal). The scenario structure is deliberately finite and explicit — enumerated adoption trajectories, enumerated commitment outcomes at each register rung, enumerated weather years from the historical record — rather than sampled: the output is not a probability of adequacy but a provable statement of the form under scenario S, at hour class H, seam X delivers Y megawatts, and here is the derivation. Probabilistic inputs enter as labelled scenario axes, never as buried weights. The reconciliation failures documented earlier — the three seams whose paired assumptions cannot simultaneously hold — fall out of this construction automatically, because a single ledger cannot double-book what two sovereign forecasts can (BGL Run Ledger HG-3226, internal).

The obvious objection is institutional, not technical: the inputs are commercially sensitive, and no operator will publish its queue or its contract book. This is precisely the constraint the stack’s identity layer was engineered for. Under Tri-Con identity, counterparties contribute evidence as encrypted composite references — the derivation engine can determine that two queue entries in different provinces resolve to the same commitment, and can count it once, without any party (including the lab) observing the counterparty’s raw identity or terms (BGL Engineering Note EN-153, internal). Duplicate detection across confidential queues — the single most valuable operation in the national accounting, and the one no province can perform alone — is exactly the class of multi-party computation over sensitive corpora the architecture performs in production elsewhere: the operational deployment fuses 19 live sources and completes full-corpus analysis in 431 milliseconds with complete derivation trails, under the governed-reasoning regime the Stratum platform enforces (BGL/HG—01, 2026). The register today runs on public and semi-public evidence, which bounds its resolution; the same machinery, fed under Tri-Con by even two cooperating operators, would resolve the duplication question between them definitively — and neither would learn anything about the other’s book beyond the collision itself (BGL Memorandum M-2121, internal).

What would change, concretely, if allocation ran on provable accounting? Alberta’s triage would attach each admitted project to a published evidence grade and each deferral to a derivable capacity constraint — converting a discretionary ranking into a rulebook competition of the Singaporean form, with the rulebook’s arithmetic re-runnable by every applicant (AESO, 2025; IMDA, 2024). Ontario’s prioritization statute would gain the instrument its criteria presuppose: jobs-per-megawatt and grid-impact comparisons computed on a common evidentiary base rather than on submissions each applicant authors about itself (Government of Ontario, 2025). Québec’s determinations could travel: a project’s register standing, established once, would follow it into the next province’s queue instead of being re-derived from the applicant’s own refreshed narrative (BGL Run Ledger HG-3197, internal). And the federal intertie program — five priority corridors and a generational capital commitment, per the founding brief — would size its wires against reconciled, wave-resolved, design-hour demand instead of the sum of thirteen optimisms (NRCan, 2026; CEAC, 2024; BGL/HG—01, 2026). None of this replaces a single operator’s planning authority, exactly as the dispatch layer proposed in HG—01 replaces no ISO’s engine. It gives the authorities a shared floor of provable fact — which is the one input the federation’s structure cannot currently produce for itself.

The Research Programme

HG—04 closes the diagnostic phase of the Helios Grid demand workstream and opens its instrumentation phase. Three lines of work are funded and active. First, the register itself: release cadence moves to quarterly, coverage extends to hydrogen electrolysis and industrial conversions on the same evidence ladder, and the methodology annex — ladder definitions, survival tables, duplicate-resolution criteria — is being prepared for external publication so that the discipline can be criticized in the open, which is the only way an accounting standard hardens (BGL Load-Integrity Register r6, internal; BGL Memorandum M-2124, internal). The 4.7 multiple will move as evidence accumulates and as provincial allocation regimes force announcements to reveal themselves; the lab’s standing forecast for its own instrument is that the multiple falls as the ladder does its work — a register succeeding looks like phantom load exiting the record, not like a number defended (BGL Memorandum LIR-7, internal).

Second, seam certificates: the reconciliation runs summarized in Section 10 are being productionized as a standing computation over the eight material inter-provincial seams, refreshed with each operator filing cycle, with the three unreconcilable seam pairs re-derived first under expanded evidence (BGL Run Ledger HG-3226, internal; BGL Engineering Note EN-152, internal). The near-term deliverable is deliberately modest and deliberately public: a single worked certificate for one seam, one scenario, one hour class, with its complete derivation record — a demonstration that the form is computable, not a claim that the lab’s inputs are final. Third, the allocation-audit protocol: a specification, developed against the comparative regimes of Section 11, for what an explicit-published allocation decision must disclose to be independently re-runnable — evidence grades, capacity envelopes, criteria arithmetic — offered as a reference standard for any jurisdiction that wants to occupy the corner of the governance grid none currently does (BGL Memorandum M-2117, internal; IMDA, 2024; Virginia JLARC, 2024).

The partnership surface is the same one the founding brief mapped, sharpened by a year of federal motion. The National Electricity Strategy and the priority-intertie program create direct demand for reconciled cross-jurisdictional accounting; the federal funding architecture — NRCan’s Energy Innovation Program, the smart-grid streams renewed in Budget 2025, Innovative Solutions Canada, and IRAP — fits instrumentation work of exactly this scale and stage; and the Treasury Board’s contribution-agreement machinery, with intellectual property vesting in the recipient, remains the workable Crown-partnership form for an IP-intensive lab (NRCan, 2026; TBS; BGL/HG—01, 2026). The specific invitation this document extends is narrower than a grand coordination scheme, because narrowness is what the Wasoqonatl precedent teaches: two operators, one seam, one Tri-Con-mediated duplicate-resolution exercise over their confidential large-load queues — a bounded engagement whose entire output is the list of collisions and a derivation record each party can verify alone (Canada Infrastructure Bank; BGL Memorandum M-2121, internal). The lab estimates the exercise would resolve, in one filing cycle, more phantom capacity than any tariff reform now proposed — and would cost less than the study phase of a single transmission project (BGL Memorandum M-2124, internal).

The demand shock will be narrated, for the next decade, as a race to build — and the building is real: the wires, the reactors, the storage, and much of the compute will exist, and the country will be larger for it. But the record assembled here supports a more precise statement. Canada’s binding scarcity is not generation, land, gas, cold air, or capital; every section of this document found those in surplus somewhere in the federation. The binding scarcity is verified truth about demand — who is actually coming, where, at what hour, with what evidence — shared across thirteen systems that currently cannot even agree on what a committed megawatt is. Roughly four-fifths of the load now shaping Canadian grid politics cannot be tied to a binding instrument (BGL Load-Integrity Register r6, internal). Decisions of generational consequence are being allocated against that fog by discretion, province by province, at the exact moment the fog is thickest. The deterministic thesis, demand-side form: the grid does not need a better guess about the future. It needs an account of the present that no party can dispute — and from provable present fact, the allocation questions become answerable in the open. D.A.E. is the instrument. The register is the proof of method. The partnership conversation is open: research@blackgridlabs.com.