Natural Gas Data Centers in Canada: The Cost Accounting and the Feasibility Test

Canada is about to find out what a natural gas data center actually costs, because it is about to build more of them per capita than anywhere else in North America. The country has roughly 337 MW of AI data center capacity today and more than 20 GW of projects announced[16]. Almost all of the announced capacity is in Alberta, and almost all of it is gas-fired, behind the meter, because the provincial grid cannot connect it[31].

The first one is no longer a concept. Meta is building a C$13 billion, 1 GW campus in Sturgeon County, powered by the Greenlight Electricity Centre - a 932 MW combined cycle plant that reached final investment decision in July 2026 at a total cost of about C$4.6 billion[1][2]. That single line item, C$4.6 billion for 932 MW, is the most important number in Canadian data center power right now, and most feasibility decks still model gas plants at a third of it.

This article builds the cost stack the way an underwriter would: fuel, carbon, operating cost, capital recovery, and the provincial levy, line by line, in Canadian dollars, against numbers that are actually published. Then it runs the feasibility test that decides whether the project survives contact with a board.

Why the gas plant exists at all: the grid door is closed

Start with the queue, because the queue is the reason the cost stack exists. Data centers have asked the Alberta Electric System Operator for roughly 19,600 MW of new grid connection as of July 2026 - more than the province's entire existing generating fleet[8]. The AESO's interim large-load framework allows about 1,200 MW of new large load to connect through 2028. That allocation is fully spoken for: a 970 MW site tied to the Meta campus and a 230 MW Keephills campus. Everything else waits for a Phase 2 that has not been sized[31].

Alberta's answer is not a bigger queue. It is a rule that changes what a data center is. The Data Centre Regulation (Alta Reg 117/2026), in force June 9, 2026, governs any facility seeking 75 MW or more of transmission service and requires every developer to demonstrate a plan to meet its own electricity needs rather than relying entirely on the grid - the "bring your own power" model[17]. Bill 8 gives the Minister and the AESO authority to set data-center-specific connection rules, load management obligations and exemptions that are not subject to the market-fairness tests that constrain every other AESO rule[18]. Bill 12 adds a levy on computing equipment, with the rate set by how much grid you use: 2% for grid-connected facilities at or above 75 MW, 1% for self-generation with grid redundancy, and 0% for fully off-grid supply[29][32].

So the decision has already been made for you. In Alberta you are not choosing between grid power and a gas plant. You are choosing which flavor of self-supply, and the levy schedule is the province telling you which one it prefers.

Build the cost stack line by line

The reference project is well documented, which is unusual and useful. Greenlight is a two-by-one H-class combined cycle plant - two Siemens Energy SGT6-8000H gas turbines, two SST6-5000 steam turbines and two SGen6-3000W generators - under a fixed-price equipment agreement and a long-term service agreement, with an Aecon-led EPC consortium covering about 85% of project cost under fixed price[1][2]. The Class III capital estimate is about C$4.0 billion; with roughly C$0.6 billion of financing costs the total is about C$4.6 billion[2]. Pembina's share of run-rate EBITDA is disclosed at about C$310 million on a 47.5% interest[1].

Those disclosures are enough to build the stack. Four lines, in Canadian dollars per MWh at the generator bus, at a 90% capacity factor:

Convert that to the IT load and it becomes the number a CFO signs: divide by the electrical efficiency of the delivery chain (about 97% through transformers and switchgear) and multiply by PUE. At PUE 1.20 the gas plant costs 14.2 ¢/kWh of IT energy, or about C$1,240 per kW of IT load per year. A 500 MW hall spends about C$620 million a year on power before a single GPU is financed.

Now put that next to what the alternatives actually cost, because the comparison is not flattering and it is the most important thing in this article:

Read that table honestly. The Alberta wholesale pool averaged 4.6 ¢/kWh in August 2026[9] - less than half the cost of the gas plant that is supposed to be the cheap option. Grid power in Alberta is not expensive. It is unavailable. The gas plant is a price of admission, not a bargain, and every feasibility model that presents it as a cost saving should be sent back.

Fuel: the line that matters least

Alberta has the cheapest gas in North America and it still barely shows up in the price. AECO-C averaged C$1.45/GJ in 2024 and C$1.74/GJ in 2025; the Alberta Energy Regulator's ST98 outlook forecasts C$2.73/GJ for 2026, C$3.22/GJ for 2027 and C$3.99/GJ by 2035[7][8]. Actual 2026 spot has been running below the forecast - C$1.51/GJ in August[9] - while the forward curve sits near C$1.74/GJ for 2027[9]. At a 6.0 GJ/MWh heat rate, the entire range from today's spot to the 2035 forecast spans C$9 to C$24/MWh.

Against a C$700/kW-yr capacity charge, that is noise. A 2.6× move in gas price moves the delivered cost by 14%[Fig. 1]. Any supplier conversation that spends more time on the gas price than on the utilization rate is avoiding the real number.

Do one sanity check on any fuel figure you are handed, because it exposes assumptions nothing else will. Greenlight will burn about 150 million cubic feet per day[1]. Convert it: 4.25 million m³/day at 38.3 MJ/m³ is about 162,700 GJ/day. At 932 MW and a 90% capacity factor the plant makes about 20,100 MWh/day, so the contracted fuel implies a heat rate of about 8.1 GJ(HHV)/MWh - 7,660 Btu/kWh, or roughly 49% LHV efficiency. A datasheet H-class combined cycle runs near 6,300 Btu/kWh, which would need only about 126 MMcf/d at the same output.

The 20% gap is not an error. It is a firm transportation contract sized for peak-day output, degradation between overhauls, and the gas a plant burns starting up and shutting down - and you cannot curtail a hyperscaler's contract to save fuel. The lesson generalizes: when a developer quotes a fuel number, divide it by output and capacity factor and see what heat rate falls out. If the implied efficiency is below about 52% LHV on a new combined cycle, someone is either buying headroom you did not price or building a plant that will not perform.

Carbon: the line everyone prices wrong

Alberta's TIER system is an output-based benchmark, not a tax on every tonne, which means the carbon line item is a function of one question: what is your emission intensity, and what is the benchmark you are measured against[6]?

The headline prices are now fixed for a decade. Under the Canada-Alberta implementation agreement signed in May 2026, the industrial carbon price holds at C$95/tonne through 2026, rises to C$100 for 2027-2029, C$115 in 2030, then about C$3/tonne per year to C$130 in 2035 and C$140 by 2040[3][4][5]. A credit price floor starts in 2030 at C$60/tonne and climbs to C$110 by 2040[4].

The benchmark is where the money is. Alberta's electricity high-performance benchmark was set at 0.37 tCO₂e/MWh in 2019 and had tightened to about 0.355 tCO₂e/MWh by 2024, declining roughly 1% a year under the new agreement[6]. Call it 0.35 today and 0.32 by 2035. A modern H-class combined cycle at 60% LHV efficiency emits about 0.33 t/MWh - below the benchmark, generating credits rather than obligations.

But run the same plant on the fuel the contract actually buys and the answer flips. The 8.1 GJ/MWh heat rate implied by the 150 MMcf/d contract is an intensity of about 0.41 t/MWh - above the benchmark. On 7.35 TWh a year, that is a 410 kt compliance obligation:

That is a C$43 million a year swing on a single interpretation, and neither number is published. Whether a behind-the-meter plant selling all its output to one customer is measured against the electricity benchmark or on a facility-specific basis is exactly the kind of clause that gets discovered in year three.

Beyond CO₂, the plant inherits air-permitting obligations like any generator. In Alberta that means an approval under the Environmental Protection and Enhancement Act with an emissions inventory, dispersion modelling and stack testing - a process that begins with the equipment selection, not after it[27]. Even a 4.9 MW five-engine data center supply now goes through a public EPEA application[30]. At 932 MW the approval is a project, not a form, and continuous emissions monitoring becomes a permanent operating cost[26].

Capital: the line that decides everything

Here is the uncomfortable arithmetic. Capital recovery is C$58.6/MWh of the C$84.3/MWh bus cost in a transparent build-up - about 70% of the plant's cost of electricity. Fuel is 23%. In Canada in 2026, a behind-the-meter gas plant is a capital-recovery business that happens to burn gas.

And the capital has tripled. Greenlight's C$4.6 billion for 932 MW is about C$4,936/kW, against roughly C$1,965/kW implied for the 900 MW Cascade plant completed two years earlier - a 2.6× step in two years[24]. Wood Mackenzie has turbine prices heading toward US$600/kW by the end of 2027, up 195% on 2019, with 110 GW of global orders in 2025 against 60-70 GW of annual manufacturing capacity[14]. Lazard's 2026 LCOE for US combined cycle is US$51-129/MWh, a 17-year high, and its head of power expects further increases as data center demand meets equipment shortage[15].

The schedule is worse than the price. GE Vernova ended Q2 2026 with 116 GW of gas equipment backlog and slot reservations and is now booking 2031 deliveries; Siemens Energy carries a 69 GW firm backlog; Mitsubishi Heavy Industries 35 GW. Large frames run about six years from order, smaller machines 18-36 months, and the US Department of Energy now quotes up to seven years for equipment that used to take two to three[14]. OEMs require reservation fees simply to hold a slot - one utility disclosed US$25 million for a single turbine[13].

The choice is a schedule decision disguised as an efficiency decision. Goldman Sachs weights time-to-market at 20% and availability at 15% in its on-site power scoring, against 25% combined for LCOE and capex[13] - because a month of delayed energization on a large AI campus is worth more than the entire efficiency delta between a recip and a turbine[12].

Reciprocating engines are the quiet workhorse of this cycle: highest simple-cycle efficiency of any combustion technology, near-flat plant efficiency across a wide load range because you switch units off instead of throttling them, and a delivery queue measured in quarters rather than years[12]. The trade-off is that a 500 MW campus built from 2.5 MW-class units is two hundred engines, and every one of them is a maintenance event, an air permit line and a methane slip source.

The levy: cheap until it is not

Alberta's data center levy looks like a rounding error and behaves like a trap. The rate is up to 2% of computing equipment value[29]. On a 1 GW campus carrying C$6 billion of GPUs, that is C$120 million a year at the grid-connected rate, C$60 million at the self-generation-with-redundancy rate, and zero off-grid.

The catch is the credit mechanism. The levy is fully creditable against Alberta's 8% corporate income tax, with a seven-year carryforward - and the credit is non-refundable[32]. A data center operator running accelerated depreciation on that equipment sits in net operating loss positions for years and may never generate enough Alberta tax payable inside seven years to absorb it. For those operators the levy is not a prepayment. It is a final tax.

The market has already responded. The first material project to break ground under the levy is one engineered around it - Bitdeer's Fox Creek campus pairs an on-site gas plant with a grid connection deliberately held below the 75 MW threshold[32]. Structure the power solution before you sign the lease, because the levy rate is a function of that structure and not the other way round.

What the provinces actually charge

Every province has now written a data center rule, and they are not variations of the same idea. They are five different answers to the same question - who pays for the megawatt - and the answer changes the project's cost stack before you have a site.

The pressure is not hypothetical. Independent analysis of the Meta project puts the impact on a typical Alberta household bill at C$267-462 a year between 2027 and 2031, once market pressure from the new load is counted against the transmission savings the project delivers[23]. That is the reason every province is now writing a rate class instead of just a connection queue.

Notice what happened to the "cheap Canadian hydro" pitch. Québec is proposing to roughly double what a large data center pays, to about the same 13 ¢/kWh that a brand-new Alberta gas plant costs to build and run[20]. The clean-power advantage has not disappeared - Québec's grid sits near 1 gCO₂e/kWh against an Alberta fleet that averaged about 0.43 t/MWh in 2025[6] - but the price advantage largely has. What remains is a carbon-intensity advantage, a permitting regime that is slower and more selective, and no C$4.6 billion on your balance sheet.

Run the feasibility test

Four gates, in order. A project that fails any one of them is not a project.

Gate 1 - Can you physically get the megawatts?

In Alberta, only if you generate them. Size the plant for the campus's actual profile, not its nameplate. The AESO's June 2026 guidance treats data centers as non-conforming loads and signals ramp limits in the neighbourhood of 10 MW per minute at the connection point[10] - and AI clusters do not ramp gently, because thousands of GPUs change state together. If your load profile cannot live inside that envelope, the answer is storage inside the fence, which is capital, and it belongs in the same model as the turbines.

Gate 2 - Can you get the equipment before the market moves?

A heavy-duty turbine slot, not a turbine price, decides when a campus ships[14]. If your energization date is inside four years, your realistic options are aeroderivatives, reciprocating engines or a bridged configuration under the Alberta regulation - and each of those carries a higher heat rate, a higher emission intensity and a different carbon answer than the combined cycle in your base case.

Gate 3 - Does the carbon line hold under both readings?

Model the plant at the datasheet heat rate and at the contracted fuel rate, against both the electricity benchmark and a facility-specific benchmark, at the C$95 headline and at market credit prices. That is eight numbers, and the spread between the best and worst of them was about C$43 million a year on a 932 MW plant[6]. Get the benchmark treatment in writing before final investment decision, not after the first true-up.

Gate 4 - Does the utilization rate survive the demand forecast?

This is the gate that kills projects. The delivered cost is dominated by capacity recovery, which means it is dominated by capacity factor[Fig. 1]. At 90% utilization the model says 11.4 ¢/kWh. At 70% it says 14.1 ¢. At 55% it says 17.3 ¢ - and at that point you are paying a premium over Hydro-Québec's proposed tariff to burn gas at 0.41 t/MWh.

A behind-the-meter plant sized to one customer has no merchant market to sell into when that customer's demand disappoints. This is the same arithmetic that makes low IT load quietly hostile to liquid cooling plants - the equipment you bought for full load becomes the wrong equipment at partial load[see our analysis of part-load effects]. Model the downside demand case explicitly, and ask what the generation asset is worth on the day the tenant does not renew.

The risk register

The decision checklist

• Price the power as capital recovery, not fuel. If your model has gas as the biggest line, the model is wrong. At C$4,936/kW, capital is 70% of the cost of electricity.
• Buy the slot before the site. Turbine reservations, not land options, set your energization date[14].
• Demand the fuel number and divide it. Contracted gas volume ÷ output ÷ capacity factor = implied heat rate. Below ~52% LHV on a new combined cycle means headroom or underperformance.
• Get the TIER benchmark determination in writing and model it at the headline price and at market credit prices[6].
• Structure for the levy rate - 2%, 1% or 0% - before the lease, and stress-test the non-refundable credit against your tax profile[32].
• Model 55% and 70% capacity factors as first-class scenarios, not sensitivities. They are the difference between 11 ¢/kWh and 17 ¢/kWh.
• Compare against Québec's proposed Rate CD at ~13 ¢/kWh[20] as your clean-power alternative, and against the Alberta pool at 4.6 ¢/kWh[9] as the price you are paying to skip a queue.
• Write the CER-reversal scenario into the financing, with a CCS-ready design and a dated compliance plan[24].

Natural gas data centers in Canada are buildable, bankable and already under construction. They are also the most capital-intensive way to make a kilowatt-hour in this country, and the numbers above are the ones that decide whether yours survives its first true-up. If you are sizing generation, cooling and electrical infrastructure for a Canadian or North American AI campus, send us your load profile and site parameters - we will put real numbers against the equipment side of that stack.

Related reading: how GPU compute rental economics set the revenue side of the same model, why 800 VDC distribution changes the electrical design behind the meter, and how modular container builds compress the schedule that gas plants stretch.

References

1. Pembina Pipeline Corporation — "Pembina Pipeline Announces Positive Final Investment Decision on the Greenlight Electricity Centre" (July 2, 2026)
2. Construction Review Online — "Greenlight Electricity Centre Reaches FID for $4.6B, 932 MW Alberta Data Centre Power Plant"
3. Government of Alberta — "Advancing an oil pipeline to the west coast" (industrial carbon pricing agreement)
4. EY Tax News Update — "Canada | Alberta revises industrial carbon prices"
5. International Carbon Action Partnership — "Canada publishes new carbon price trajectory" (May 19, 2026)
6. Climate Decode — "Alberta Power: A Flat TIER Obligation" (electricity benchmark and stringency)
7. Alberta Energy Regulator — ST98: Alberta Energy Outlook (AECO-C price history and forecast)
8. Get Energy — "Alberta Energy Watch — week of September 28, 2026" (ST98 summary, AESO connection list)
9. DNE Resources — "Alberta electricity & gas market update — August 2026" (pool price, AECO spot and forwards)
10. Bankable — "Meta's $13B Alberta AI Bet Runs on Gas" (AESO ramp guidance, transmission charges, fuel economics)
11. Stout — "Data Center Power: Firm and Expensive" (CCGT busbar cost build-up)
12. Pantheon — "Gas turbine vs reciprocating engine for data centers" (efficiency and heat-rate comparison)
13. Macrostream — "Goldman Sachs Predicts On-Site Data Center Power Plants to Become Mainstream" (time-to-market weighting, engine backlogs)
14. Acquire.fi — "Extended Gas Turbine Procurement: Why New Turbines Won't Arrive Until 2031" (OEM backlogs, lead times, Wood Mackenzie pricing)
15. Energy Connects / Bloomberg — "US Cost for Gas Power at 17-Year High and Climbing, Lazard Says" (July 2026)
16. BNN Bloomberg / The Canadian Press — "Canada data centre capacity could soar to 20 GW in planned projects: government document" (July 8, 2026)
17. EnergyRates.ca — "Alberta's Data Centre Regulation, Explained: Tethered, Bridged, and More"
18. Gowling WLG — "Status of data centre development in BC and Alberta" (BYOP framework, equipment bottleneck)
19. Legal 500 — "Powering Up: Grid Access, Strategic Screening And Canada's Emerging Data Centre Regimes"
20. Data Center Dynamics — "Canadian utility Hydro-Québec proposes electricity tariff for data centers"
21. Torys — "Ontario writes the playbook: a new framework for data centre grid connections"
22. City of Medicine Hat — "Gas, electric, wind energy rates set for June 2026" (Alberta RoLR industrial rate)
23. Pembina Institute — "New Meta data centre could add up to $460 a year to Albertans' electricity bills"
24. The Receipts — "Canada Pitched U.S. Tech Giants on Clean Energy It Doesn't Have" (TIER compliance question, Greenlight cost escalation)
25. Kalkine — "Kalina Power Strengthens Cash Reserves Amid Booming Alberta Data Centre Market" (180 MW allocation sale)
26. VL Energy — "Powering the AI Boom: How ES-PEMS Keeps Data Center Generation Compliant"
27. Air Research Group — "Alberta Environment Air Permits for Industrial Facilities" (EPEA approval process)
28. Gowling WLG — "Three power shifts shaping Canada's data centre market" (BC allocation, Ontario consultation, Québec hearing)
29. EY Tax Alert 2026 No. 12 — Alberta Budget 2026 (data centre levy framework)
30. Red Deer Advocate — "Armada Nowlit Power Ltd. — Environmental Protection and Enhancement Act" public notice (Sept 24, 2026)
31. Colin Smillie — "Power Is the Bottleneck: What's Really Constraining Canada's AI Data-Centre Boom" (AESO queue math, project list)
32. Future Summit Intelligence — "The Sovereign-Capability Reframe: Alberta in the AI Data Centre Buildout" (levy mechanics and credit treatment)