AI Data Center Economics in Canada and the US: Build Models, Power Costs & Site Selection

The single most common analytical error in North American AI infrastructure conversations is treating "AI data centre" as one thing. A powered shell developer, a neocloud operator and a mining company converting a hall are running three different businesses that happen to share a transformer. They carry different capital structures, different revenue mechanisms, different staffing loads, different build durations and completely different risk profiles.

Below are the seven configurations the SCR model prices. The distinctions matter because a per-megawatt figure quoted without a model attached is meaningless.

The table below sets out indicative all-in build cost per megawatt of IT load, in Canadian dollars with the US-dollar equivalent before state adjustments, excluding accelerator procurement. "Shell" covers structure, electrical distribution and mechanical plant. "Fit-out" covers white space, racks, cooling terminations and commissioning. Land, utility contribution and network are separate line items because they behave differently by site, province and state.

USD figures are the CAD constants restated at 1.38, the rate the benchmarks were originally converted at, so they are the underlying US benchmarks rather than a second conversion. For a specific US state, the configurator then multiplies shell, fit-out, cooling and network by a construction location factor running from 0.86 in Arkansas and Mississippi to 1.30 in Hawaii, and adds equipment sales and use tax where the state has no exemption open to new projects.

Three observations that matter more than the absolute figures.

Edge deployments are expensive per megawatt and that is not a defect. A 500 kW edge site carries the same design, permitting and commissioning overhead as a much larger one, spread across a fraction of the load. Edge projects are almost never justified on cost per megawatt; they are justified on latency, sovereignty or a compliance requirement that no central facility can satisfy. If an edge business case is being argued on unit economics, it is being argued wrongly.

Retrofits look cheap in the shell column and are not cheap overall. The shell figure is low because the structure and the interconnect already exist. The fit-out and network figures carry the real work, and in practice a retrofit's largest schedule risk is not construction but the discovery that the existing electrical topology cannot support redundant paths without substantial rework.

Accelerator procurement dwarfs all of this where the operator buys the silicon. In the neocloud and hyperscale models, the hardware line is typically the dominant capital item and depreciates on a schedule that does not care whether the building is finished. That timing mismatch — capital committed to silicon against a facility that cannot energise for two years — is the single most common way an otherwise sound AI infrastructure model fails.

Operating cost tracks capital intensity but not proportionally. Annual OpEx typically runs between 1.8% of CapEx for a powered shell and 4.0% for an edge deployment, with staffing ranging from roughly CAD 31,000 (about USD 22,500) per megawatt annually for a shell to CAD 99,000 (about USD 71,700) per megawatt for edge sites where headcount cannot be spread across load. For a line-item treatment of the mining-side equivalent, see our 300-unit crypto mining facility cost breakdown.

Electricity price is the number everyone asks for first and the number that decides the fewest projects. The variable that actually decides sites in Canada right now is the interconnection queue — the indicative time from a signed service agreement to energisation. The spread across the provinces is roughly 16 months, from 20 in New Brunswick to 36 in British Columbia and Quebec, which is long enough to invalidate an entire pro forma. United States sites are covered in the state section below.

The constraint behind each number

British Columbia. BC Hydro concluded the stepped Rate Schedule 1823 on 1 April 2026, and RS 1830 is now the default transmission rate at 4.914 cents per kWh energy plus a demand charge, blending near 6.6 cents at high load factor. New data centre capacity is competed for through the emerging-industries allocation, which drew 15 applications for roughly 800 MW, with successful applicants notified from mid-September 2026. The rate is excellent and the power is genuinely clean; the question is whether a given project wins an allocation at all. Our BC Hydro electricity rate analysis covers the tariff structure in detail, and power grid strategies for Vancouver-area operators covers the practical service application path.

Alberta. No longer the fast lane. The AESO Phase 1 large-load cap of 1,200 MW was fully allocated to two projects, and requested data centre load now exceeds 20,600 MW against a provincial peak near 13,000 MW, so grid-supplied projects now model at about 26 months. Bringing your own generation is the practical route: the Data Centre Regulation (AR 117/2026, in force 9 June 2026) gives tethered bring-your-own-generation projects of 75 MW and above priority, with bridging grid access of up to three years under heightened curtailment risk. The AESO's first bring-your-own-generation intake slipped to later this fall. Pool prices averaged about CAD 32 per MWh in Q1 2026 while 2029 forwards traded near CAD 63, so model the forward curve rather than the spot. Carbon intensity remains a real trade-off for tenants with emissions-linked procurement standards.

Quebec. Clean, heavily spoken for and now materially repriced. Hydro-Québec has asked the Régie de l'énergie for a dedicated rate for data centres above 5 MW of about 13 cents per kWh, roughly 91% above the current rate, to take effect on 1 November 2026, with existing sites phased in over five years. Operators are contesting it, Régie hearings ran from 1 to 9 October 2026, and no decision has been issued. Blockchain load is proposed at 19.5 cents. Every project above 5 MW still passes through a selection process, and a 36-month indicative queue makes Quebec a site for projects that can wait.

Manitoba. Among the lowest industrial rates in North America. The binding constraint is capacity headroom and transmission access rather than price, which is a materially different problem from an expensive grid and requires a different diligence approach.

Ontario. Global Adjustment mechanics dominate large-load economics to the point that the posted rate is close to uninformative. Class A participation materially changes the effective rate, and any Ontario model built on a headline number rather than a Class A analysis will be wrong by a wide margin. Under Bill 40 (Royal Assent December 2025), data centres above a prescribed threshold will need provincial approval to connect, and the draft Data Centre Playbook released on 13 August 2026 scores projects on economic benefit, data sovereignty and community confidence; the final Playbook and thresholds were not yet out as of 1 October 2026. The model now uses a 36-month indicative queue.

Saskatchewan, New Brunswick, Nova Scotia and Newfoundland & Labrador. Smaller grids where an individual project can be material to system planning. New Brunswick is actively pursuing data centre load. Saskatchewan has limited large-load precedent, so expect bespoke negotiation and utility contribution requirements. Nova Scotia's higher rates and tighter capacity generally suit edge and modular scale rather than campus builds. Labrador holds substantial hydro capacity, with transmission and fibre as the practical limiters rather than generation.

Since version 1.3.0 the SCR AI Build Configurator covers all 50 states and the District of Columbia alongside the nine provinces. Selecting a state switches every figure to US dollars. The model then applies that state's indicative large-load rate, queue, construction cost factor and property tax level, plus two things that barely exist in the Canadian data: equipment sales tax where no exemption is open, and utility take-or-pay tariffs for large loads.

The binding constraints differ from Canada. South of the border, capacity is rarely rationed by a provincial crown utility. It is gated by utility large-load tariffs, by regional capacity markets, and by state tax policy that moved faster in 2026 than in any year before it. The National Conference of State Legislatures counts 38 states offering data centre incentives, and lawmakers in 28 of them introduced measures to scale those back or modify them this year.

What decides US sites right now

Virginia. The largest data centre market in the world, and the clearest example of tariff structure outweighing price. From 1 January 2027, Dominion Energy Virginia customers of 25 MW and above move to the GS-5 class. It carries 14-year contracts, minimum charges of 85% of contracted transmission and distribution demand and 60% of generation demand, plus collateral and exit fees. The equipment sales tax exemption was kept, but a data centre electricity consumption tax of US$0.011 per kWh has applied since 1 July 2026 (HB 30) on grid and self-generated power, capped at US$600 million a year and scheduled through 30 June 2028; the configurator adds it to the energy cost. The State Corporation Commission ordered on 31 July 2026 that new large loads pay directly for transmission built only to serve them, and Prince William County ended by-right data centre development on 22 September 2026. Northern Virginia energisation queues model at about 48 months.

PJM states. Thirteen states and DC sit inside PJM, where the 2028/2029 capacity auction cleared at the cap of roughly $325 per MW-day. That cost is already inside the indicative rates for Delaware, DC, Maryland, New Jersey, Ohio, Pennsylvania, Virginia and West Virginia, and for the PJM portions of Illinois, Indiana, Kentucky and Michigan.

Texas. ERCOT offers one of the lowest large-load rates in the model at about 7.1 cents, but the modelled queue is now about 42 months. Senate Bill 6 governs loads of 75 MW and above: interconnection study fees, disclosure of backup generation, and mandatory curtailment in grid emergencies. ERCOT is tracking about 474 GW of large-load requests, roughly 90% of them data centres. On 21 September 2026 the governor directed state agencies to halt data centre permits and approvals until a PUCT and ERCOT audit finishes, with no fixed end date, and he has said he will seek repeal of the sales tax exemption in the 2027 session; the exemption is still law. 4CP transmission charges can move a real invoice well off the average.

Federal and state review. FERC issued show-cause orders on large-load interconnection to all six RTOs and ISOs on 18 June 2026, and PJM filed a framework on 13 August 2026 under which new loads of 50 MW or more that do not bring their own capacity face curtailment ahead of other demand response from 1 June 2027. Maryland now reviews projects of 25 MW and above under a 23 September 2026 executive order, and Nevada's Executive Order 2026-005 ties abatements to paying the school support tax in full, covering grid costs and protecting water.

Ohio and take-or-pay. AEP Ohio's data centre tariff bills 85% of contracted capacity above 25 MW, and Ohio paused new data centre tax incentive approvals on 27 May 2026. Similar minimum-demand structures are modelled for Georgia, Indiana, Kansas, Missouri, New Jersey and Oregon. The practical lesson is that the power reservation should be sized to a realistic load ramp. An oversized reservation is paid for whether it is used or not.

The 2026 tax pullback. Arizona's exemption is closed to new applications from 1 July 2026 to 30 June 2029. Illinois paused its incentives for two years from 1 July 2026. Nebraska ended access to its data centre incentives by executive order in July. Massachusetts stopped accepting new exemption applications, New Jersey froze its remaining credits, and Oklahoma ends incentives for centres not operating by January 2027. Washington ended its exemptions for replacement servers and refurbished urban sites on 1 July 2026 (ESSB 6231), and Minnesota repealed its data centre sales tax exemption for purchases from 1 July 2026. Maine now excludes new data centres from its business equipment property tax exemption (LD 713), and New York paused environmental permits for new sites of 50 MW or more for a year from 14 July 2026. For a build that owns its accelerators, a closed exemption can add more to capital cost than any realistic difference in power price.

Cheapest and dearest. Louisiana and New Mexico lead on rate at about 6.5 cents, followed by Oklahoma, Texas, South Carolina and Tennessee. Hawaii's island grids sit at about 38 cents, with California and Alaska near 20.5. North Dakota has the shortest indicative queue at about 24 months. California, Hawaii and Virginia have the longest at about 48.

Bitcoin mining companies hold something genuinely scarce: energised interconnects at scale, secured years before the current demand cycle. That is a real advantage and it is why the retrofit model exists. It is also routinely overstated.

A mining hall is engineered around an assumption that a failed hash costs almost nothing. Dust ingress is tolerated. Temperature swings are tolerated. Single utility feeds and minimal redundancy are normal. Networking is sized for a trickle of pool traffic rather than for east-west accelerator communication. None of that is a defect in a mining context; all of it is disqualifying in an AI hosting context.

Converting therefore means conditioned space, redundant power paths, carrier-grade fibre, structured cabling, and in most cases a cooling architecture the site was never designed to accommodate. Our model prices retrofit shell at CAD 6.0 million per megawatt against CAD 8.5 million for a purpose-built powered shell — a genuine saving — but adds CAD 2.2 million per megawatt of fit-out against the shell model's CAD 1.0 million, plus a substantially higher network allocation. The saving is real and it is smaller than the pitch decks suggest.

The hybrid model exists for operators unwilling to choose. Hashing continues to monetise megawatts while AI capacity is built out, and power yields progressively to the higher-value tenant as conversion completes. The model prices a blended revenue stream with an adjustable AI share, defaulting to 55%. It carries the lowest capital intensity of the seven at roughly CAD 7.35 million per megawatt, and the operational complexity of running two businesses on one interconnect. For background on the mining side of that equation, see Bitcoin mining on Vancouver Island and our crypto data centre overview.

Figures on this page and in the configurator are derived from published utility tariffs and rate schedules, publicly available interconnection guidance, and observed large-load precedent in Canadian and US markets, together with U.S. Energy Information Administration state price data, state data centre tax incentive status and utility large-load tariff filings, blended into indicative planning constants and stamped with a vintage date. The current vintage is Q4 2026 (rev. 1 October 2026).

What this model is

A screening tool. It exists to answer, in minutes rather than weeks, whether a given combination of province or state, build model, scale and contract rate is worth taking to a feasibility study. It is designed to be wrong in ways that are visible and directional rather than precise and misleading.

What this model is not

It is not a substitute for a utility service agreement, an engineering estimate, a geotechnical assessment or a tax and incentive analysis. Provincial, county and municipal incentives, negotiated property tax abatements, water availability, fibre routes, seismic classification and local permitting posture all move real project economics, and none of them are in this model. For US states, state-level equipment sales tax status, a property tax index and large-load take-or-pay floors are modelled; negotiated local deals are not. A screening output that looks good should trigger diligence, not a commitment.

Known limitations we would rather state than hide

Blended rates compress a great deal of complexity — demand charges, time-of-use structures, Ontario's Global Adjustment, US capacity market pass-throughs such as PJM and ERCOT 4CP transmission charges, negotiated industrial arrangements — into a single figure. Queue estimates are indicative and vary substantially by specific interconnection point, not merely by province or state. Accelerator pricing and revenue-per-hour assumptions move faster than any published dataset can track and should always be overridden with your own figures where you have them. The Enterprise licence tier exists specifically so that operators can replace our cost constants with theirs.

If you find a figure that disagrees with what you know from direct experience, we would rather hear about it than defend it. Corrections go to support@strategiccryptoreserve.ca and are folded into the next dataset vintage, which propagates to every licensed deployment on the next page load.