This book began as a classroom exercise. On 3 September 2026 the incoming Full-time MBA, Master of Management and Master of Business Analytics cohorts at the UBC Sauder School of Business were given twenty-five minutes, a team of four and any artificial intelligence tool they liked, and asked one open question: the wave of artificial intelligence data centres is coming to Canada, so what should it mean for Canada, and for whom?

It seemed only fair to do the exercise myself. This book is the result. Fourteen research agents worked in parallel, one per dimension of the question, each producing a written brief, a machine-readable dataset and a folder of downloaded primary documents. Between them they ran roughly five hundred searches, saved 884 files and 716 megabytes of source material, and recorded every figure with a link. What you are reading is those fourteen briefs, edited into a single argument, with the charts and the connective tissue added.

I want to be exact about what that means, because the method is part of the subject. The research was gathered by machines. The framing, the ordering, the judgement calls and the errors are mine. Where two sources disagreed, the disagreement is reported rather than resolved. Where a number could not be verified, it is marked as contested rather than quietly dropped. Two of the scores that appear in the companion software are explicitly labelled as my judgement rather than published data, and the reasoning is shown beside them. Readers who want to check anything will find the full source for every figure in Appendix G.

The subject deserves that care, because it is moving quickly and it is being argued about badly. In the fifteen months to August 2026 the Alberta Electric System Operator's data-centre connection queue was reported at 16.2, 19.6 and 20.7 gigawatts by three different counters on three different dates. All three numbers are true of the moment they describe. Quoted without a date, any of them is misleading. Much of the public conversation about data centres in Canada consists of numbers quoted without their dates.

There is a second reason for care. A great deal of what has been announced will not be built. Roughly twenty-one gigawatts of Canadian capacity has been announced or proposed; about two point three gigawatts is actually under construction. The federal government's own documents say most of the proposed pipeline is not expected to proceed. An argument built on the announcement number and an argument built on the construction number are arguments about different countries.

The book is organised in three parts. Part I sets the global picture and the physics, because you cannot argue about megawatts without knowing what one is. Part II is Canada on the ground: the projects, the rules and the ledger of costs and benefits. Part III is the strategic frame: the border, the inheritance and the money. The appendices hold the inventory, the timelines, a glossary and every source.

My own conclusion, which readers are free to reject, appears at the end of the executive summary. The students were not asked to reach it. They were asked to reach one they could defend.

Chunhua Wu
Marketing and Behavioural Science, UBC Sauder School of Business
Vancouver, September 2026

Artificial intelligence has to live somewhere. That somewhere is a building full of processors that consumes electricity, water, capital and land, and it has to be built in a specific place, under a specific government, on a specific grid. The question of where those buildings go is now one of the larger industrial questions in the world, and Canada is unusually well placed to answer it and unusually unprepared to.

What is happening

World data-centre electricity consumption was about 415 terawatt hours in 2024 and 485 in 2025, and the International Energy Agency's base case puts it near 945 by 2030. The growth alone is roughly another Japan. Forecasts keep being revised upward: BloombergNEF's estimate of American data-centre demand in 2035 went from 78 gigawatts in April 2025 to 106 in December to 194 in July 2026. The four largest hyperscale operators plan about 725 billion United States dollars of capital spending in 2026, up from 410 billion in 2025, and almost a third of it is now debt funded, against nine per cent two years earlier.

The constraint has moved. It is no longer processors. It is electricity, and the iron that delivers it. A power transformer takes about 160 weeks. A large gas turbine takes about five years. Average grid connection in the Americas takes five years. Capacity that exists is essentially full: North American vacancy has sat near one per cent for three years, and ninety-five per cent of the sixty-six gigawatts under construction is already leased.

Why Canada

Canada's advantage is specific and measurable. Six of the seven cheapest large-power cities in North America are Canadian: Winnipeg at 5.09 cents per kilowatt hour, Montreal at 5.52, Vancouver at 6.72, against Toronto at 12.66 and New York at 22.88. Seventy-eight per cent of Canadian generation is non-emitting. The prairie cities sit at or below fifteen degrees for nine months of the year, which is free cooling. The country is politically stable and legally predictable, and it is not the United States, which since 2025 has become a commercial argument in itself.

Why it is harder than it looks

There is almost no spare room. Every large province is now rationing rather than recruiting. British Columbia caps new data-centre load at 400 megawatts over two years and 145 megawatts per project, written into regulation, against about 800 megawatts requested. Quebec requires a ministerial decision for any load above five megawatts and has asked its regulator to charge data centres about 13 cents per kilowatt hour, roughly double the general large-power rate. Ontario makes data centres pay the full cost of their grid impact and screens every connection at the minister's desk. Alberta received requests near twenty thousand megawatts against a record provincial peak of 12,785 and allocated 1,200 megawatts of firm capacity in total, all of which went to two projects.

The consequence is the central fact of this book. The only route to gigawatt scale in Canada this decade is to build your own power station, and the only province that will let you is the one whose grid runs at about 490 grams of carbon dioxide per kilowatt hour. Meta's campus in Sturgeon County, Alberta, is the template: thirteen billion dollars, one gigawatt, groundbreaking in July 2026, with its own 932 megawatt gas plant. Canada's cleanest power and Canada's only buildable gigawatt sit in different provinces. A one gigawatt campus on the Alberta grid emits about 3.9 megatonnes of carbon dioxide a year. The same machine in Quebec emits about 0.01. That is a factor of nearly three hundred, decided entirely by which province signs the contract.

What it is worth, and to whom

The investment is real and large: Meta's project alone is about 17.6 billion Canadian dollars including its power plant, with roughly 250 million a year in provincial revenue and 40 million in municipal tax by government estimate, and a 60 million dollar community package. The employment case is much weaker: about thirty permanent jobs per hundred megawatts, against construction crews ten to a hundred times larger that then leave. Electricity costs can be shifted onto other customers, as they were in one American capacity auction to the tune of 6.5 billion dollars. This asymmetry, large capital and tax against small permanent employment, is why the argument is so often about whether the community got a fair price rather than about technology.

Communities have begun to answer. Rocky View County refused a campus and later imposed a moratorium. Mississauga and Oakville paused approvals. The Alberta Utilities Commission refused a 1.4 gigawatt gas plant at Olds. Manitoba's premier dropped a proposal after a petition gathered more than thirteen thousand signatures. There were coordinated demonstrations in fourteen cities in June 2026. Social licence is no longer a communications problem; in several provinces it is the binding constraint.

The two arguments that could overturn everything

The first is the bubble. Bain estimates that artificial intelligence needs about two trillion dollars of annual revenue by 2030 to fund the compute being built, and will fall roughly 800 billion short. A widely cited study from MIT found that ninety-five per cent of enterprise pilots produced no measurable profit. The International Energy Agency notes that only about a fifth of American grid connection requests ever materialise. A province that builds transmission for a load that never arrives has made an expensive mistake, which is precisely why Ontario and Alberta now make the developer carry that risk.

The second is repricing. Canada's advantage is the price of firm clean power, and Canada is in the process of raising it. If Quebec's proposed 13 cent rate takes effect, the country's cheapest clean electricity becomes roughly as expensive as Ontario's, and much of the arithmetic in Chapter 8 changes. Any recommendation that rests on 5.5 cents should say what happens at 13.

The conclusion

Four answers are defensible on this evidence, and they disagree about weights rather than facts. Alberta is the only province that can deliver scale this decade, and the price is carbon and contested local consent. Quebec has the cleanest and cheapest firm power and is rationing it, so it suits a hundred-megawatt project and not a gigawatt one. Ontario has the largest system, a nuclear build-out and the customers, and it charges full freight and takes three to five years. And not building at all is defensible too: leasing capacity rather than owning it may be the right call if the revenue does not arrive.

For a student, an investor or a minister, the useful discipline is the same. Name the number your case depends on. Say which date and which counter it came from. Then say what would have to be true for you to change your mind.

Each of the eight chapters follows the same shape, so the book can be read straight through or opened at any point.

• An opener giving the chapter's kicker number, the questions it answers and a short standfirst.
• Key takeaways, six to ten of them, each carrying a figure and the source that produced it.
• The picture in numbers, the tables and charts, with the unit, the year, the region and the source for every value.
• The narrative, roughly a thousand words of plain-English explanation at the level of a graduate business student, defining each technical term the first time it appears.
• Contested or uncertain figures, which is the section to read before quoting anything.
• Open questions, written for students looking for a research topic.
• Sources, numbered, with publisher, date and link.

Notation and conventions

• Currency is Canadian dollars unless the unit says otherwise. Where a figure is in United States dollars it is written US$ or stated in the surrounding text.
• Energy is quoted in the unit the source used. A megawatt (MW) is a rate of power; a megawatt hour (MWh) is an amount of energy. A gigawatt (GW) is a thousand megawatts, and a terawatt hour (TWh) is a million megawatt hours. Appendix F defines the rest.
• Project capacity is given as announced. Announced, under construction and operating are different things and the book keeps them apart.
• Figures were verified to 2 September 2026. Several move monthly, particularly the Alberta connection queue.
• Where a claim rests on a single secondary source, or where sources conflict, it is flagged in the chapter's contested section rather than presented as settled.

Narrative

What is being built, and why the numbers are so large. A data centre is a building full of servers; AI has filled it with racks of accelerators such as Nvidia graphics processing units (GPUs) that draw far more power per rack. A terawatt-hour (TWh) measures energy used over time: one TWh is one billion kilowatt-hours, and 415 TWh, the IEA's estimate of data-centre electricity in 2024, is roughly 1.5% of everything the world consumed that year. A gigawatt (GW) measures power drawn at an instant: 1 GW is 1,000 megawatts (MW), about the output of a large nuclear reactor, and a 100 MW data centre running all year uses roughly the electricity of 100,000 households (IEA 2025). Capacity figures from real-estate firms are usually expressed in MW of "critical IT load", the power available to servers; total power drawn is higher by the facility's power usage effectiveness (PUE), the ratio of total energy to server energy, typically 1.1 to 1.5.

Demand. The IEA's April 2025 Energy and AI report set the reference numbers: 415 TWh in 2024, rising about 15% a year to 945 TWh in 2030 and 1,200 TWh in 2035 in its Base Case, with 2035 scenarios spanning 700 to 1,720 TWh. Its April 2026 update reported 485 TWh in 2025, a 17% rise, with AI-focused facilities up 50%, and left the 2030 figure at about 950 TWh. The United States dominates (45% of 2024 consumption, 180 TWh), followed by China (25%) and Europe (15%). US-specific work by Lawrence Berkeley National Laboratory (LBNL) shows the acceleration: the December 2024 report had 176 TWh in 2023 (4.4% of US electricity) and 325 to 580 TWh by 2028; the 2025 Update, published in 2026, has 192 TWh in 2024 and a 649 TWh Reference Case for 2030 (11.8%), with data centres supplying a third of all US load growth this decade. The important nuance is between energy and peak power: utilities plan for peak demand, and the aggregate five-year peak forecast filed by US utilities has risen six-fold since 2022 to 166 GW, about 90 GW of it data centres (Grid Strategies, November 2025).

Capacity and markets. Real-estate data describe what has actually been built. JLL counts about 103 GW of data-centre capacity worldwide at the start of 2026 and expects 200 GW by 2030, with the Americas holding half and the US about 90% of the Americas. Cushman & Wakefield's 2026 comparison shows the pipeline's scale: 43.4 GW operating in the Americas, 25.3 GW under construction and 191.3 GW planned. The market has no slack. "Vacancy" is the share of built capacity available to lease; "absorption" is the amount newly leased; "pre-leasing" means signing tenants before a building is finished. North American vacancy has been about 1% for three years, H1 2026 absorption doubled to 25 GW, and 95% of the 66 GW under construction is already committed, so tenants signing today are getting 2028 deliveries (JLL Midyear 2026); Northern Virginia, the world's largest market, had 0.3% vacancy in Q1 2026 (CBRE). Hyperscalers, the largest cloud and internet companies, now run 1,360 large data centres and hold 48% of all capacity, heading to 67% by 2031 (Synergy Research). Outside North America the story is one of clusters and constraints: Johor in Malaysia and Mumbai in India are growing fastest; Ireland, where data centres take 22% of electricity, only ended a four-year connection moratorium in December 2025 by requiring new sites to bring their own generation and 80% renewables; Amsterdam allows no new data centres; the Gulf is building the 5 GW UAE-US AI campus with a 1 GW Stargate cluster.

Money. Four companies, Amazon, Alphabet, Microsoft and Meta, spent about USD 410 billion of capital expenditure (capex, spending on long-lived assets) in 2025 and guide to about USD 725 billion in 2026; with Oracle the IEA counts more than USD 400 billion in 2025 and a 75% rise. Nvidia's data-centre revenue was USD 193.7 billion in its fiscal 2026 and USD 89.0 billion in the quarter to July 2026. The money is increasingly borrowed: FactSet finds debt funded 32% of hyperscaler capex in the last twelve months versus 9% in fiscal 2024, and JLL counts USD 250 billion of AI-related bonds in the first half of 2026. Flagship projects show the gap between announcement and operation: OpenAI's Stargate has more than 9 GW planned by 2029 but only 0.3 GW operating in April 2026.

The shortages. Chips were the first bottleneck and are still tight at the memory layer: high-bandwidth memory (HBM), the stacked memory bonded to each AI accelerator, is sold out at Micron and SK hynix through 2026, and the IEA expects the shortage to last through at least end-2027, capping AI-server additions at roughly 25 GW a year. Power has become the binding constraint, captured in the industry phrase "time to power". An interconnection queue is the waiting list on which a utility or grid operator studies requests to connect large new loads or generators; in PJM, the grid covering 13 US states around Northern Virginia, AI projects entering service in 2025 waited more than seven years from request to operation, and Cushman & Wakefield puts average connection times at 5.0 years in the Americas. Behind the queue sit physical shortages: GE Vernova's gas-turbine backlog and reservations reached 116 GW with new orders delivering in 2031; large power transformers take more than 160 weeks and cost 77% more than in 2019; more than 60% of operators report difficulty hiring, and electricians are the tightest trade. Water and permitting add local friction: US data centres consumed 66 billion litres of water directly in 2023 and about 800 billion litres indirectly through the power they used, and LBNL expects hyperscale sites alone to use 60 to 124 billion litres in 2028; a JLL survey found only 14% of Americans support a data centre near them.

The demand gap. Put the two halves together and a gap appears. Demand-side models (McKinsey 219 GW of capacity demand by 2030; Bain 200 GW of incremental AI compute by 2030; Epoch AI's finding that frontier training compute grows four to five times a year) exceed what the supply side can plausibly deliver (JLL's 200 GW of total capacity by 2030 in its base case, HBM at 25 GW a year, turbines booked to 2031). Bain frames the gap financially: USD 2 trillion of annual AI revenue would be needed by 2030 to fund the build, and it expects an USD 800 billion shortfall. BloombergNEF sees a 19 GW US power shortfall even at record connection rates. For a country with spare clean electricity and land, the implication is that power delivered within two to three years is now worth more than cheap power delivered in seven.

Where Canada fits. Canada ranks fifth in the world by number of data centres (337) but is small by capacity: about 1.4 GW live against 40.6 GW in the US. The pipeline is disproportionate: DC Byte counts 10.3 GW across all stages, and federal documents list 20 to 22 GW of proposed projects, more than 18 GW of them in Alberta, against a government estimate that the country needs 5.5 GW of AI compute by 2030 and a warning that most proposals will not proceed. The advantage is cheap, clean power in the hydro provinces: Montreal's large-power rate of 5.55 cents per kWh compares with 12.01 in Houston and 18.07 in New York, and Vancouver (7.91) and Toronto (8.88) also undercut most US hubs, though Calgary and Edmonton were above 23 cents in the 2023 comparison. The constraint is allocation and time to power. Alberta received more than 20 GW of connection requests, more than its peak demand, and capped interim connections at 1,200 MW to 2028, pushing developers toward "bring your own generation". BC Hydro is releasing just 400 MW over two years. Quebec, where data-centre load is expected to rise from 200 MW to more than 1,000 MW by 2035, proposes to charge data centres 13 cents per kWh, about double its industrial rate, from late 2026. Canada's opportunity is therefore less about total scale than about whether it can convert a small number of large, well-sited, quickly powered projects while the global gap is widest.

The picture in numbers

A. Electricity demand

B. Capacity, markets and the pipeline

C. Money: capex, chips and flagship projects

D. Bottlenecks

E. Where Canada fits

Contested or uncertain figures

• What counts as "capacity". Goldman Sachs counted about 55-59 GW of global data-centre power demand in 2023-2025, McKinsey 82 GW of capacity demand in 2025, JLL 103 GW of total capacity at the start of 2026 and Cushman & Wakefield 43.4 GW operating in the Americas alone. The differences come from definitions (critical IT load versus total facility power, and whether enterprise on-premise sites are included), not disagreement about the trend. Compare growth rates across sources, not levels.
• US 2030 electricity use. LBNL's 2025 Reference Case (649 TWh) is far above the IEA's US estimate for 2030 (about 430 TWh, as cited by LBNL) and inside EPRI's 380-790 TWh range. LBNL attributes the divergence to assumptions about accelerator shipments, so the range, not the point estimate, is the honest answer.
• Speculative and duplicated load requests. The IEA finds only about 20% of US data-centre connection requests materialise in the short to medium term; Grid Strategies says utility forecasts overstate data-centre load by roughly 25 GW and rest on unrealistic 96% load factors; Canary Media reports developers pitching the same project to several utilities. Grid Strategies argues NERC's reliability shortfalls largely disappear once queued generation is counted; NERC disputes this. Alberta's 20 GW of requests against a 12 GW peak illustrate the same phenomenon in Canada.
• Forecast whiplash. BloombergNEF's 2035 US figure rose 2.5 times in fifteen months; NERC's ten-year peak growth rose 69% in a year. Rapid upward revision is itself evidence of low forecast reliability, in either direction.
• Hyperscaler capex totals differ by USD 100 billion or more depending on whether finance leases and customer prepayments are included and on Microsoft's June fiscal year (USD 630 billion, 690 billion, 725 billion and 800 billion all appear in reputable compilations for 2026). Nvidia's fiscal 2025 data-centre revenue (USD 115.3 billion) in the data file is derived from reported growth rather than quoted directly.
• Goldman Sachs' US path (31 GW in 2025, 41 GW in 2026, 66 GW in 2027) and Meta's USD 50 billion Louisiana investment come from search summaries of pages that blocked automated access and are marked unverified in the data file.
• Canadian market statistics. CBRE's H2 2025 North America report has no Canadian markets; its H1 2025 Toronto profile is descriptive, and the widely repeated 7.6% Toronto vacancy and 223 of 236 MW pre-committed figures could not be verified in the page text. Synergy, JLL and CBRE do not publish Canada's share of global capacity; the 1.4 GW versus 40.6 GW comparison relies on DC Byte via BLG. The "337 MW of AI capacity" in the federal document uses a narrower definition than the 1.4 GW of live capacity.
• Electricity price comparison. The Hydro-Quebec city comparison served at the public URL is the edition with rates in effect April 2023; Alberta's figures reflect the 2022-2023 pool-price spike and have since fallen, and Quebec's proposed 13-cent data-centre rate would erase most of Montreal's headline advantage for new AI sites.
• The bubble. MIT's 95% figure is based on 52 interviews, 153 survey responses and 300 deployments and has been criticised as unrepresentative; Goldman's finding that only 2% of S&P 500 firms quantify AI benefits sits alongside hyperscaler earnings growing 54%. Circular financing (Nvidia investing up to USD 100 billion in OpenAI, AMD warrants, Oracle's USD 300 billion contract) inflates the apparent size of demand.

Open questions students could investigate

1. If only about 20% of US connection requests materialise, what is the equivalent conversion rate for Canada's 20-22 GW pipeline, and which project features (secured generation, water, permits, tenant) predict conversion?
2. What is the all-in cost per MW-year of a 300 MW AI campus in Montreal, Toronto, Calgary and Vancouver versus Northern Virginia, Dallas and Phoenix once Quebec's 13-cent rate, Virginia's new data-centre tax and Alberta's bring-your-own-generation requirement are included?
3. How much is "time to power" worth? Build a model of the value of a 2027 versus 2031 connection date to a hyperscaler, given GPU depreciation over four to six years.
4. Does Canada's clean-power advantage survive if data centres are powered by on-site gas turbines in Alberta, and how would that interact with the federal Clean Electricity Regulations and buyers' carbon commitments?
5. Which of the demand forecasts (IEA, LBNL, EPRI, BloombergNEF, McKinsey) has the best track record so far, and what does the pattern of revisions say about how to weight them?
6. If Bain is right about an USD 800 billion revenue gap by 2030, which parts of the buildout get cancelled first, and would Canadian projects be early or late casualties?
7. How should a province allocate scarce grid capacity among data centres, electrification and industry: first come first served, auctions, or economic scoring like BC Hydro's 145 MW per-project cap?
8. What share of Canada's proposed projects are training clusters (latency tolerant, could sit anywhere with power) versus inference capacity (needs proximity to users), and what does that imply for Alberta versus Toronto?
9. Could Canada capture value beyond hosting: transformers, switchgear, turbines, cooling and skilled trades, given 160-week transformer lead times and a 116 GW turbine backlog?
10. What community-acceptance conditions (water, noise, local rates, jobs) have derailed projects in Ireland, the Netherlands and US counties, and are Canadian municipalities prepared?

Sources

Narrative

Here is the shape of the thing, in plain English.

For the first twenty years of this century, computing got enormously more useful without getting much hungrier. That is the fact almost everyone leaves out. Between 2010 and 2018 the world's data centres ran roughly six times as much computing work while their electricity bill rose about 6 per cent, from 194 to 205 terawatt-hours. Storage capacity went up twenty-five-fold over the same period. Internet traffic went up more than tenfold. Electricity went sideways. As a share of world electricity, data centres moved from about 1 per cent in 2005 to about 1.5 per cent in 2024. That is nineteen years for half a percentage point.

This did not happen by luck, and it did not happen because chips got better, though they did. It happened for a mostly boring reason: the work moved buildings. In 2014 more than sixty per cent of American server energy sat in company-owned rooms and closets where the average server ran at about ten per cent utilisation and the building wasted about as much electricity on cooling as it delivered to the computers. By 2023, nearly ninety per cent of that work had moved into hyperscale and colocation facilities where servers run at forty-five per cent utilisation and the best buildings waste ten per cent instead of one hundred per cent. Virtualisation, which lets one machine carry many workloads, meant that servicing more demand needed fewer physical boxes: American server shipment growth fell from fifteen per cent a year in the early 2000s to three per cent after 2010, even as the internet exploded.

Lawrence Berkeley National Laboratory put a price on all of this. It ran the counterfactual: freeze American efficiency practice at 2010 levels and let only the chips improve. On that path, 2020 consumption would have been 170 per cent higher, and the industry's efficiency work saved about 620 terawatt-hours over the decade. That is roughly Canada's entire annual electricity generation, saved by engineering that got no press at all.

Now the second half of the story. Every one of those offsets is either exhausted or exhausting.

The building migration is nearly complete: with enterprise rooms down to single-digit percentages of server energy, there is almost nothing left to move. Power usage effectiveness, the cooling-overhead ratio, has fallen from 2.0 in 2007 to about 1.45 across the American fleet, and at facilities built for AI it is already 1.145 and forecast to reach only 1.136 by 2030. A building at 1.1 has ten per cent of overhead left to remove, ever. Koomey's law, the fifty-year trend of computations per kilowatt-hour doubling every eighteen months, slowed to a doubling every 2.6 years after 2000, when the physics of shrinking transistors changed.

And into that exhausted set of levers walked a workload with the opposite characteristics. An NVIDIA P100 accelerator drew 300 watts in 2016. The B300 and AMD's MI355X draw 1,400 watts in 2025. The power density of AI servers rose elevenfold between 2020 and 2025 and is expected to rise fourfold again by 2027, at which point a single rack the size of a large refrigerator could peak at the electricity demand of sixty-five households. The chips got more numerous, each one got heavier, and the building could not get any more efficient. Three multiplicative terms, all pointing the same way, at the same time.

So the curve bent. Global data-centre electricity grew about twelve per cent a year from 2017, then seventeen per cent in 2025 alone. In the United States it grew fourteen per cent in 2024 with twenty-two and twenty-nine per cent forecast for 2025 and 2026. Accelerated servers were fifteen per cent of global data-centre electricity in 2024; by 2030 AI-focused facilities are forecast to be about half the total.

Three things a business audience should take from this rather than the headline.

First, the numbers are softer than they look. Every historical global figure in this brief is a model output, not a meter reading. The IEA says so explicitly. When an independent review graded over a hundred studies, the estimates for 2020, a year that has already finished, spanned two hundred to twelve hundred terawatt-hours, a factor of six. Projections for 2030 diverge by nearly forty times. And there is a pattern: studies with weak methods produced systematically bigger numbers, averaging 1,013 terawatt-hours for 2023 against 335 from the strong ones. Peer review did not filter this; five of the nine peer-reviewed global estimates since 2014 were graded low or low-medium.

Second, the famous per-query numbers are the weakest part of the field, and the failure is instructive. The figure most people have heard, that a ChatGPT query uses about ten times a Google search, decomposes into three separate errors. The "Google search uses 0.3 watt-hours" anchor comes from a Google blog post of January 2009 which has never been updated. The "ten times" came from an Alphabet chairman talking to Reuters about DOLLARS, not electricity. And the resulting 2.9 watt-hours was published with the words "at most," a hedge that vanished when a consultancy repeated it. When researchers sampled a hundred news articles about AI energy, seventy-five per cent gave a figure with no uncertainty or source and five per cent urged caution. Meanwhile the three organisations that actually did the work by different methods, Google with production telemetry, Microsoft with a peer-reviewed bottom-up model, and Epoch with an independent one, landed at 0.24, 0.31 and 0.30 watt-hours. The convergence is the news; nobody reported it. The honest caveat is that Google's own paper computes 0.24 and 0.10 watt-hours for the same prompts on the same day, purely by changing where the boundary sits, and that a reasoning query measures thirteen times a short one.

Third, cryptocurrency is the control case, and it teaches in both directions. Bitcoin grew from thirteen terawatt-hours in 2017 to about 189 in 2025, which is more than the entire global electric vehicle fleet used in 2024. It is driven by coin price, not by anyone's demand for a service. Then on 15 September 2022 Ethereum changed a consensus rule and removed roughly twenty-one terawatt-hours a year, about two thirds of Ireland's whole annual generation, in a single day. Digital electricity demand is not a law of nature; it is a set of choices. But the disanalogy is just as important. Proof-of-work mining was wasteful by design, which is exactly why it could be deleted without losing anything. AI compute does work that customers are paying for, so there is no equivalent switch to flip, and anyone who suggests otherwise is selling something.

Where does that leave a decision-maker? With a load that is real, growing fast, geographically concentrated, and roughly a sixth AI today heading toward roughly a half by 2030. Also with a load that is smaller than air conditioning's growth, smaller than the world's televisions, and roughly one and a half per cent of world electricity. Both of those sentences are true, and anyone using only one of them is arguing rather than informing.

The long series

Read the boundary column before the number column. Four questions decide whether two rows can share an axis: does it count storage and network equipment or only servers; does it count facility infrastructure such as cooling; does it count cryptocurrency mining; does it count telecom networks and end-user devices. A fifth question applies only to crypto, and it bites hard: is this a calendar-year total or an annualised instantaneous rate.

Global, by source lineage

Three warnings on that table. First, Koomey's 2010 midpoint of 237.6 TWh and Masanet's 194 TWh are different models, not a revision of one another, and the 44 TWh gap for the same year is the honest measure of how much model choice matters. Second, the 220 TWh often attributed to Malmodin and Lunden for 2015 is a third party's back-out; the paper's own table says 245 TWh for data centres plus enterprise networks. Third, the 2030 rows are scenarios, and the IEA 4E review advises against projections reaching more than five years ahead.

United States, the best documented series in the world

The Energy Act of 2020 requires the United States Department of Energy to publish this, which is why it exists.

Note two disagreements that are worth showing students. The same laboratory put 2014 at 70 TWh in 2016 and at about 60 TWh in 2024, a 17 per cent revision from updated shipment data. And the IEA and LBNL differ by 12 TWh for the United States in 2024, the same country in the same year.

The inflection points

Mainframes and early servers, 1946 to 1999. No credible global electricity series exists. What is documented is efficiency: computations per kilowatt-hour doubled about every 1.57 years from ENIAC in 1946 to 2009, measured across 76 machines with an R-squared of 0.983 (Koomey, Berard, Sanchez and Wong, 2009 and 2011). ENIAC managed 437 computations per kilowatt-hour on 150,000 watts. The whole modern debate rests on a series that only begins in 2000.

The dot-com build-out and the internet-electricity myth, 2000 to 2005. World consumption roughly doubled, from 70.8 to 152.5 TWh, about 16.6 per cent a year. Almost all of it came from the sheer number of cheap volume servers, which went from 19.7 to 50.5 TWh worldwide; power per box barely moved. United States server shipments grew 15 per cent a year, nearly doubling the installed base. Why the curve bent upward: the internet was being built, and each new service needed its own physical machine. This is also the era of the field's biggest public error, covered in its own section below.

Rapid growth meets the financial crisis, 2005 to 2010. World consumption rose about 56 per cent rather than doubling, to a midpoint of 237.6 TWh, about 9.3 per cent a year. The United States rose about 36 per cent. Why the curve bent downward: the 2008 to 2009 recession cut server purchases hard, server virtualisation arrived and let one physical box carry many workloads, and cost pressure made operators care about efficiency for the first time. American server shipment growth fell from 15 per cent a year to 5 per cent. The IEA dates the decoupling of data-centre electricity from digital service demand to 2007 and 2008.

The efficiency plateau, 2010 to 2018. From 194 to 205 TWh, about 0.7 per cent a year, while compute instances grew about 550 per cent, storage capacity 25-fold and internet traffic more than tenfold. Why the curve went flat: four offsets, quantified in the next section. This is the central fact of this dimension and the thing most commentary still gets wrong in both directions.

The cryptocurrency surge, 2017 onward. Bitcoin went from about 13 TWh in 2017 to about 96 TWh in 2022 and about 189 TWh in 2025, roughly 49 per cent a year over the first five years. Why the curve bent: nothing to do with demand for a service. Mining hardware is added whenever the coin price exceeds the electricity cost of running it. It appeared outside every official data-centre statistic and remains outside them.

The Merge, 15 September 2022. Ethereum switched from proof of work to proof of stake and removed roughly 21 TWh a year of demand in one day, about two thirds of Ireland's entire annual generation. Why the curve bent: a rule change. This is the counter-example that stops anyone treating digital demand as a law of nature, and the disanalogy matters as much as the analogy, because proof of work was wasteful by design while AI compute does work people are paying for.

The artificial intelligence inflection, 2022 to 2026. From 350 TWh in 2022 to 485 TWh in 2025, about 11.5 per cent a year, accelerating to 17 per cent in 2025 alone, with AI-focused facilities up 50 per cent in that year. In the United States: 14 per cent in 2024, a forecast 22 per cent in 2025 and 29 per cent in 2026. Why the curve bent: the offsets ran out at the same moment the demand arrived. That is the whole story, and it is the next section.

The projected decade, 2026 to 2030. The IEA Base Case doubles global consumption to 950 TWh, about 14.4 per cent a year, with AI-focused facilities more than tripling to about 465 TWh. Accelerated server electricity grows 30 per cent a year against 9 per cent for conventional servers. Treat every number after 2026 as a scenario.

Why it was flat, and why it broke

The four offsets that held the line, quantified

1. Migration out of enterprise rooms. In 2014 more than 60 per cent of American server energy sat in internal company data centres. By 2023 that was nearly 10 per cent, with hyperscale and colocation facilities taking almost 80 per cent, and by 2028 internal facilities fall below 2 per cent (LBNL, 2024). This mattered because of utilisation: LBNL modelled internal data centres at 10 per cent average server utilisation against 45 per cent for hyperscale. Moving a workload from a 10 per cent box to a 45 per cent box removes roughly four servers.

2. Virtualisation and utilisation. Virtualisation, which lets one physical server carry many independent workloads, arrived between 2005 and 2010 and is the main reason American server shipment growth fell from 15 per cent a year to 5 per cent and then to 3 per cent after 2010 (LBNL, 2016).

3. Falling power usage effectiveness. At the time of the 2007 EPA Report to Congress the average was 2.0, meaning every kilowatt-hour reaching the servers needed another for cooling and power conversion. Some sites measured above 3.0. By 2014 the American fleet average was 1.6, by 2024 it was 1.45, and by 2030 it is forecast at 1.36. The gains came almost entirely from the shift to big buildings, not from improving small ones: LBNL's 2014 modelling put hyperscale at 1.2 and a server room at 2.5, and the best hyperscale operators were already recording 1.1 or less by 2016.

4. Chip efficiency, Koomey's law. Computations per kilowatt-hour doubled about every 1.57 years from 1946 to 2009. Energy per compute instance fell about 20 per cent a year through the 2010 to 2018 plateau (Masanet et al., 2020).

The counterfactual that prices all four. LBNL froze American industry efficiency at 2010 levels (server utilisation 14 per cent, idle draw 59 per cent of maximum, 11.3 watts per hard disk drive, weighted average PUE 1.89) while letting chip performance keep improving. Result: 2014 consumption would have been 60 per cent higher and 2020 consumption 170 per cent higher, a saving of about 620 TWh over the decade (LBNL, 2016). That is roughly Canada's entire annual generation, saved by engineering that nobody wrote about.

Why they stopped working

Koomey's law slowed first. After about 2000 the doubling time for peak-output computations per kilowatt-hour stretched from roughly 1.5 years to about 2.6 years, as Dennard scaling ended (Koomey and Naffziger, 2016). Note also the metric's own caveat: Koomey's law measures peak output, and servers in typical business applications run near full load only 5 to 15 per cent of the time.

The migration lever is spent. With internal data centres already down to a few per cent of American server energy, there is almost nothing left to move.

The PUE lever is spent. American facilities serving AI equipment were at 1.145 in 2024 and are forecast to reach 1.136 in 2030. That is a 0.8 per cent improvement over six years. A hyperscaler at 1.1 has at most 10 per cent of overhead left to remove, ever.

Accelerator power went vertical. NVIDIA's Tesla P100 drew 300 watts in 2016 and its V100 the same in 2017. The A100 reached 400 watts in 2020, the H100 700 watts in 2022, the B200 1,200 watts in 2024, and the B300 and AMD's MI355X 1,400 watts in 2025. That is a 4.7-fold rise in nine years, at the same time as the number of chips was rising and PUE had stopped falling. Three multiplicative terms, all pointing the same way.

Density followed. AI server power density rose elevenfold between 2020 and 2025 and is set to rise a further fourfold by 2027, when a single rack the physical size of a large refrigerator could peak at the demand of 65 households (IEA, 2026). At server level, average annual power draw of an eight-GPU server passes 4 kilowatts by 2026 (LBNL, 2025 Update).

And idle power turned out to matter more than anyone assumed. LBNL cut its assumed AI inference utilisation from 40 per cent to about 20 per cent by 2030, because inference is user-facing and must be sized for peak demand, so it idles between peaks. Industry experts told LBNL idle draw could reach 50 per cent of rated power in latency-sensitive facilities against the 20 per cent LBNL assumes. That single assumption moves the 2030 American forecast by more than 20 per cent, from 649 to 782 TWh.

The breakdown

By segment, the chart that matters

No published source gives a clean four-way split of traditional, cloud, crypto and AI in terawatt-hours across a run of years. The IEA says why: there is no comprehensive data on workload shares, colocation and cloud operators often cannot see what runs in their own halls, and definitions of AI differ. What follows is assembled from published totals and published AI shares, with every derivation labelled. It should be charted with the caveat on the face of the chart.

Sources: IEA Energy and AI (2025) and Key Questions (2026) for totals and the 15 per cent accelerated-server share; Kamiya and Coroama, IEA 4E EDNA (2025) for the AI ranges; Cambridge Bitcoin Electricity Consumption Index v1.8.0 for crypto calendar-year totals. Crypto sits OUTSIDE the IEA and LBNL data-centre totals, so it is a band on top, never inside.

The 2030 row deserves a moment. The IEA's 465 TWh counts AI-FOCUSED FACILITIES, including their cooling and their conventional servers. EDNA's 200 to 400 TWh counts AI HARDWARE. Both are defensible; they are answering different questions, and the gap between them is roughly the non-AI equipment sitting inside AI buildings.

By share

Training against inference, where two authoritative sources disagree

Eighteen months apart, the same laboratory forecast training rising to about 52 per cent by 2028 and then falling to about 20 per cent by 2030. This is not a data correction. It is a change of view about where AI activity and spending go, and the newer report follows industry feedback that inference now dominates both. Chart both lines and let students argue about which is right. Use the newer one and say so.

The perspective number: every large AI model ever trained accounts for about 1.7 TWh cumulatively across 283 models. Against 485 TWh of data-centre electricity a year, all frontier training in history is about 0.35 per cent of a single year.

By data-centre type

Two different bases. Do not mix them.

The only published terawatt-hour split by data-centre type that could be verified is the World Bank and ITU's 2022 figure: colocation 110 TWh, cloud 53 TWh, content 32 TWh, totalling 195 TWh. It EXCLUDES enterprise data centres entirely, which is why it sits well below the IEA's 350 TWh for the same year. The roughly 155 TWh gap is a rough measure of the enterprise and unreported remainder.

By server type

Globally, accelerated server electricity grows 30 per cent a year to 2030 against 9 per cent for conventional servers. Accelerated servers deliver almost half the net increase in total data-centre electricity, and about 70 per cent of the increase in server electricity alone (IEA, 2025).

By end-use industry

Not verified, and not manufactured. No credible published breakdown of data-centre electricity by customer industry (finance, telecoms, healthcare, government, retail, media and streaming) could be found. This cut is genuinely absent, for a structural reason the IEA states plainly: colocation and cloud operators often have limited visibility over the workloads running in their own facilities, so nobody holds the data. Vendor reports on IT SPENDING by industry vertical exist, but spending is not electricity, and converting one to the other would be invention. Any chart a student meets showing data-centre electricity by end-use industry should be treated as unsourced until proven otherwise.

Context anchors

Against whole countries and the world

So: global data centres in 2025 used about 74 per cent of everything Canada generates, about 97 per cent of Germany's output, and about 3.6 times the Netherlands'. The 2030 Base Case is about 1.5 times Canada's entire system and about 92 per cent of Japan's. Ireland already gives about 22 per cent of its metered electricity to data centres.

Against other loads

Three of these are worth saying out loud in a classroom. Data centres in 2024 used more than twice the electricity of every electric vehicle on earth. In 2023 end-user devices used MORE electricity than all the world's data centres, and televisions used more still. And over the decade in which data centres added 250 TWh, air conditioning added 700 TWh, which got a fraction of the coverage.

On heavy industry and aviation, honesty is required. Global steel and cement electricity in terawatt-hours could not be verified from a primary source, and cement is dominated by thermal fuel rather than electricity, so an electricity-only comparison would mislead. The defensible statement is the IEA's own: by 2030 the United States is set to consume more electricity for data centres than for the production of aluminium, steel, cement, chemicals and all other energy-intensive goods combined. Aviation runs on jet fuel, so there is no honest terawatt-hour comparison at all; on carbon, data-centre electricity emitted about 180 million tonnes of CO2 in 2024, but the corresponding aviation figure was not verified here and is therefore not given.

Per unit

Useful ratios on Google's measured figure: one kettle boil is about 458 prompts, one kilometre of electric driving about 775 prompts, one phone charge about 62 prompts, and one prompt is about nine seconds of television.

Training energy, for completeness: GPT-3 took 1,287 megawatt-hours and 552.1 tonnes of CO2e (Patterson et al., 2021). The IEA estimates GPT-4 at 42.4 gigawatt-hours on a full-facility basis for training compute of 2.2e25 calculations. Llama 3.1 405B took 30.84 million GPU-hours and 8,930 tonnes of CO2e location-based, or zero market-based because Meta matches its electricity with renewables; note that the widely quoted 11,390 tonnes is the whole model family, not the 405B alone, and Meta publishes no megawatt-hour figure at all.

Contested or uncertain figures

This section matters more here than in any other dimension of the pack.

The whole historical series is modelled, not measured. The IEA states plainly that all historical data on global data-centre consumption are estimates. Very few governments require reporting, and operators do not disclose data-centre electricity specifically.

The 2020 estimates span a factor of six. From about 200 terawatt-hours to 1,200. For 2030 the spread is nearly forty times. Studies graded low quality averaged 1,013 terawatt-hours for 2023; high-quality studies averaged 335 (IEA 4E EDNA, 2025).

Koomey's 2010 figure and Masanet's 2010 figure differ by 44 TWh (237.6 against 194) for the same year. They are different models, not a revision. Never treat one as superseding the other without saying so.

LBNL revised its own 2014 estimate down by about 17 per cent between its 2016 and 2024 reports, from 70 to about 60 TWh.

The IEA and LBNL disagree about the United States in 2024 by 12 TWh, 180 against 192.

LBNL contradicts itself on training against inference. The 2024 report forecast training rising to 50 to 53 per cent of AI server energy by 2028. The 2025 Update forecasts training falling to about 20 per cent by 2030. Both are current publications from the same laboratory.

The AI share depends entirely on what "AI" means. Accelerated-server electricity in 2024 was about 62 TWh. AI-focused FACILITIES in 2030 are 465 TWh on the IEA's basis and 200 to 400 TWh on EDNA's. These are different boundaries answering different questions.

Per-query energy figures span a factor of 400, from 0.10 to about 40 watt-hours, and almost none of that is disagreement about physics. The drivers are: where the boundary sits (Google gets 0.10 or 0.24 for the same prompts), whether idle reserve capacity is counted, whether requests are batched (academic benchmarks often are not, and batching roughly halves per-task energy), how long the answer is, whether a reasoning model is thinking, whether a median or a mean is reported, and the fact that no closed model discloses its size, hardware or location.

The 2.9 watt-hour figure is disputed and should not be used as current. Its chain is documented in full in the data file and the manifest.

The "500 millilitres per conversation" water claim is a misquotation. The paper says 500 millilitres per ten to fifty medium-length responses, which is roughly 10 to 28 millilitres each, for GPT-3 in specific American data centres. The paper's own headline total moved 54 per cent between its versions. Water figures are the least comparable numbers in the field: Google's measured 0.26 millilitres is about 173 times below Mistral's audited 45 millilitres, mainly because of boundary.

The Llama 3.1 emissions figure is routinely misattributed. 11,390 tonnes is the whole model family; the 405B model alone is 8,930 tonnes location-based and zero market-based. Meta publishes no energy figure, so any megawatt-hour number for Llama is somebody's arithmetic, and the two published attempts differ by a factor of two.

Cryptocurrency has two separate traps. First, the headline Cambridge number is an ANNUALISED INSTANTANEOUS RATE, not a calendar-year total: on 2 September 2026 the live rate was 150.9 terawatt-hours a year while the 2025 calendar total was 188.6. Second, Cambridge serves five model vintages simultaneously, and 2021 Bitcoin is 89.0 terawatt-hours on the current model against 104.0 on the old one, a 17 per cent difference from methodology alone. Digiconomist reads about 35 per cent higher than Cambridge and has been frozen at an unchanged value for 259 days; its Ethereum index reads exactly zero from 6 May 2026 and should not be cited live.

The four published Merge reduction percentages are not variants of one number. 99.84 per cent is de Vries's most conservative bound; 99.95 per cent is an Ethereum Foundation forecast made sixteen months BEFORE the Merge; 99.988 per cent is the Crypto Carbon Ratings Institute's measurement; 99.99 per cent is Cambridge's day-of-Merge figure. Name the owner when you quote one.

The IEA has no crypto figure after 2023. Its only published number is about 125 terawatt-hours for 2023. Also, on the same page, DATA CENTRES grew by over 90 terawatt-hours between 2020 and 2023 while CRYPTOCURRENCY grew by over 50; the two are frequently swapped.

Could not be verified at all: any breakdown of data-centre electricity by end-use industry; global steel and cement electricity in terawatt-hours; a global level (as opposed to growth) for residential air conditioning electricity; global aviation for a carbon comparison; the energy of an LED bulb-hour or a laundry load from a primary source. None of these were invented.

Could not be obtained: the Masanet et al. Science paper itself, the Koomey et al. IEEE Annals paper itself, and the Malmodin et al. Telecommunications Policy paper itself are all paywalled. Their headline figures here come from freely available companion documents by the same authors and from the IEA 4E review's reading, both of which are in the downloads folder. Two citation corrections came out of this: Malmodin et al. 2024 is in Telecommunications Policy, not the Journal of Industrial Ecology, and the "1.57 years" of Koomey's law is verified in the 2009 companion report rather than in the 2011 article text.

Open questions students could investigate

1. LBNL reversed its own training-versus-inference forecast in eighteen months. Which report is right, and what evidence would settle it? What would a cloud provider's revenue mix tell you that a shipment forecast cannot?
2. Idle power at AI inference facilities is assumed at 20 per cent of rated power by LBNL and hypothesised at 50 per cent by industry practitioners. That single number moves the 2030 American forecast by more than 20 per cent. How would you measure it, and who has the incentive to?
3. The four offsets that produced the flat decade are exhausted. Name the next four. Which of liquid cooling, custom silicon, model right-sizing, sparsity, load shifting or something not yet invented could plausibly deliver another 620 terawatt-hours of avoided demand?
4. If per-task AI energy really is falling by an order of magnitude a year, as the IEA says, why is total energy rising? Work through the Jevons argument with the numbers in this file and decide whether it holds.
5. Ethereum deleted 21 terawatt-hours a year with a rule change. Is there any analogous lever in AI? Consider inference-time compute budgets, mandatory model right-sizing, or disclosure requirements, and estimate what each could plausibly save.
6. Build the segment chart yourself from the data file, then write the two-sentence caveat that has to appear on it. Which parts of the split are published, which are derived, and which do not exist?
7. Studies with weak methods systematically produced bigger numbers. Why would that be? What does it imply about how to read the next forecast that crosses your desk?
8. Canada generates 652 terawatt-hours a year. Global data centres will use around 950 by 2030. What share of that could Canada realistically host, given the provincial reserve margins in dimension 10, and what would it do to Canadian electricity prices?
9. Data centres in 2024 used more than twice the electricity of the global electric vehicle fleet, and end-user devices used more than data centres. Why does one of these get the coverage and the others do not, and does that misallocation of attention have policy consequences?

Sources

Narrative

The claim that Canada should attract a large share of the world's new AI data centres rests on three adjectives: cheap power, clean power, lots of it. Each one is testable. Two of the three hold up. The third does not, and the way the first two interact with the third is the actual strategic problem.

Start with size, because it disciplines everything else. Canada generated 652 terawatt hours of electricity in 2025 from about 156 gigawatts of installed capacity. That makes it the seventh largest producer in the world and the tenth largest by capacity, with 2.06 per cent of global generation. Statistics Canada and Ember, working independently, agree on the capacity figure to within one per cent, which is reassuring. But the ranking flatters. Canada is roughly the same size as South Korea and smaller than Russia, Japan and India. China generates sixteen times more.

The number that should reframe the whole conversation is the rate of addition. In 2025, China added 315 gigawatts of solar capacity and 119 gigawatts of wind. That single year of Chinese additions is 2.8 times the entire Canadian generating fleet, built from scratch, in twelve months. In the same year Canada added about 1.1 gigawatts of non-emitting capacity: 0.1 gigawatts of solar and 0.3 of wind. Canada's whole national build rate is a rounding error against the pace of its largest competitor. Any strategy that assumes Canada can out-build anyone is not a strategy.

Now price, which is where the real advantage lives and where the debate is most often misled. There are at least four different numbers that people call "the electricity price" and they differ by a factor of five. A residential customer in Montreal pays 8.29 Canadian cents per kilowatt hour. A small business pays 12.08. A medium industrial customer at 500 kilowatts pays 14.70. A 5,000 kilowatt industrial customer pays 5.83, and a 50,000 kilowatt customer pays 5.52. The large-power price is one third of the medium-power price in the same city, in the same month, from the same utility. Comparing a residential number in one country with an industrial number in another produces conclusions that are wrong by hundreds of per cent, and this happens constantly in press coverage of data centres.

Done properly, Canada's advantage is real. Statistics Canada reports an implied industrial price for mining and manufacturing of 6.72 Canadian cents per kilowatt hour in 2024, which is 4.91 United States cents. The United States Energy Information Administration reports 8.13 United States cents for the same year on the same conceptual basis, revenue divided by volume for the industrial class. Canadian industry pays about sixty per cent of what American industry pays. Hydro-Quebec's annual city comparison, the only source that puts twenty-two North American utilities on a single customer profile and a single exchange rate, gives the same answer at the level that matters for a data centre: at a 50,000 kilowatt profile, Montreal is 5.52 Canadian cents, Winnipeg 5.09 and Vancouver 6.72, against Toronto 12.66, Houston 12.90, New York 22.88 and Boston 24.58.

But Canada is not the world's cheapest. Eurostat's band IF and band IG series, which cover exactly the consumption levels a data centre reaches, put Finland at 4.16 euro cents per kilowatt hour and Norway at 4.68 in the second half of 2025, excluding taxes. Converted at the 2025 average rate those are 4.70 and 5.28 United States cents. Norway's band IG price, for consumers above 150 gigawatt hours a year, is 4.08 euro cents, or 4.61 United States cents, the cheapest large-industrial power in Europe. France sits at 6.86 United States cents on that basis. So the honest ranking puts Norway and Finland ahead of Canada's national average, with Quebec alone beating them.

Europe's price story since 2020 is worth understanding because it is what created the current competitive map. The European Union's band IF industrial price roughly tripled between the first half of 2020 and the second half of 2022, from 5.30 to 18.92 euro cents. It has since fallen back to 10.95, but that is still double the pre-crisis level and it is not falling much further. Germany is stuck at 12.83 against 4.34 in 2020, a 196 per cent rise, which is why Berlin has legislated a subsidised industrial price of 50 euros per megawatt hour retroactive to January 2026. Italy has done something similar at 65 euros. The International Energy Agency's assessment is blunt: European prices for energy-intensive industry are more than double United States prices and more than fifty per cent above China and India, and the European Central Bank estimates that a permanent ten per cent price rise costs two per cent of employment in those industries. This is why Europe's AI gigafactory programme, for all its 20 billion euros, faces a structural cost disadvantage that money does not fix.

Ireland is the cautionary tale inside that story. It is the European data-centre capital and it now has the most expensive large-industrial electricity in the European Union: 16.80 euro cents at band IF and 15.62 at band IG. Data centres take about twenty-two per cent of Ireland's metered electricity. Concentrating a lot of inflexible load on a small system does not lower prices; it raises them and then triggers connection moratoria. Singapore reached the same place by a different route: about 1.46 gigawatts live, only about 20 megawatts under construction, and 5.5 per cent non-emitting generation.

On cleanliness, Canada is genuinely strong but the usual framing overstates it. Ember, the only source that applies one method to every country, puts Canada at 77.0 per cent non-emitting generation in 2025, fifth in the world among large producers behind Norway, Sweden, France and Brazil. Its carbon intensity figure for Canada is 190.7 grams of carbon dioxide per kilowatt hour, which is higher than the 100 to 120 grams often quoted, because Ember uses lifecycle emission factors and assigns non-zero emissions to hydro, wind, solar and nuclear. Environment and Climate Change Canada, counting only what comes out of stacks, reports about 86 grams for 2024. Both numbers are defensible; mixing them is not. What matters commercially is that Canada is three to seven times cleaner than every large market it competes with for AI load, and that the provincial range runs from 1.7 grams in Quebec to 730 in Saskatchewan.

The set of jurisdictions that are simultaneously cheap and clean at scale is therefore very short: Norway, Sweden, Finland, France, and four Canadian provinces. Everywhere else forces a trade. Texas is cheap and dirty. Malaysia is cheap and among the dirtiest destinations at 603 grams. The Gulf is the cheapest gas in the world and the dirtiest grid, at 692 grams in Saudi Arabia. Denmark and Ireland are clean and expensive. That short list is Canada's actual competitive set, and it is small enough that Canada's position in it is a genuine asset.

Which brings the argument to its uncomfortable end. The North American Electric Reliability Corporation assesses Canadian and American systems on the same reserve-margin basis. Alberta has an anticipated reserve margin of 36.2 per cent against a reference level of 11.8, and holds above 33 per cent through 2030-31. Saskatchewan and British Columbia are similarly comfortable. Quebec is not: its margin falls from 15.6 per cent to 11.5 by 2030-31 against a 12.2 per cent reference, with an explicit warning of winter shortfalls. Manitoba falls to 1.2 per cent. The Maritimes are already below target.

So the province with abundant spare capacity is Alberta, where industrial power costs 10.01 Canadian cents, the highest in the country, and the grid runs at 335 to 540 grams, comparable to Germany. The provinces where the pitch of cheap, clean power is true have no room. That is the real Canadian constraint, and it is not solved by attracting more announcements. It is solved by building generation and transmission in Quebec, British Columbia and Manitoba, or by accepting that Canadian AI capacity will largely be gas-fired Alberta capacity with a carbon story to manage. Given that Canada added 1.1 gigawatts of non-emitting capacity last year and the pipeline is 20 gigawatts, the arithmetic does not currently work.

The picture in numbers

A. Installed capacity and generation

All Ember figures are from the 2026 release of the Yearly Electricity Data file, downloaded 2 September 2026 and stored in this pack. Ember's capacity total is clean plus fossil capacity; it is not a nameplate registry and should be treated as within a few per cent.

Net capacity additions, the pace test:

B. Generation mix and carbon intensity

Non-emitting share is Ember's "clean" category: hydro, nuclear, wind, solar, bioenergy and other renewables. Carbon intensity is Ember's published figure, which is a LIFECYCLE number (see "Contested or uncertain figures").

Canada on the Canadian method, and the provincial spread that the national number hides:

Which countries are both cheap and clean. Naming them explicitly, using large-industrial price below about 8 United States cents per kWh and non-emitting share above 75 per cent:

• Norway. 4.61 United States cents at band IG, 99.0 per cent non-emitting, 28 gCO2/kWh. The best combination in the world, and the reason Norwegian data-centre capacity keeps growing.
• Sweden. 7.21 United States cents at band IG, 98.8 per cent non-emitting, 35 gCO2/kWh.
• Finland. 4.70 United States cents at band IF, 96.3 per cent non-emitting, 58 gCO2/kWh. The cheapest band IF price in the European Union.
• France. 6.86 United States cents at band IG, 94.8 per cent non-emitting, 42 gCO2/kWh, and the only one of these at real scale (570 TWh).
• Canada, but only in four provinces. Quebec at 3.59 and Newfoundland and Labrador at 2.55 United States cents industrial, Manitoba at 4.45 and British Columbia at 4.86, all with grids at or below 15 gCO2e/kWh on the Canadian method. Alberta at 7.31 United States cents is the country's most expensive province and its dirtiest at 335 to 540 gCO2e/kWh.
• Iceland and Paraguay qualify on the metrics (100 per cent non-emitting) but are too small to matter: 2.2 GW and 8.9 GW of capacity respectively.
• Brazil is a borderline case: 88.7 per cent non-emitting at 110 gCO2/kWh and historically low industrial prices, but no comparable published industrial tariff in this dataset.

Everywhere else fails one test. Texas is cheap (6.65 United States cents) and dirty. Malaysia is cheap and very dirty (603 gCO2/kWh). Denmark and Ireland are clean-ish and expensive. Germany is neither. Singapore is dirty (497 gCO2/kWh from 5.5 per cent non-emitting) and constrained. The Gulf states are the cheapest gas power in the world and among the dirtiest grids anywhere, at 692 gCO2/kWh in Saudi Arabia.

C. Industrial electricity price, the part that matters most

Every row states the basis. Residential prices appear only in the clearly labelled block at the end and are not comparable to anything above it. Conversions use the Bank of Canada 2025 annual averages of 1.3978 Canadian dollars per United States dollar and 1.5782 Canadian dollars per euro, implying 1.1291 United States dollars per euro, except where the source uses its own rate.

C1. Eurostat, industrial band IF (70,000 to 149,999 MWh a year, roughly an 8 to 17 MW continuous load), second half of 2025, EXCLUDING taxes and levies.

C2. Eurostat band IG (150,000 MWh a year and over, roughly 17 MW and up, the right band for a real data centre), second half of 2025, EXCLUDING taxes.

C3. North America and the rest of the world, industrial or large-power basis.

C4. The 2020 to 2026 price trend, Eurostat band IF, EUR/kWh excluding taxes. This is the European price spike and where it has settled.

Source for the whole table: Eurostat nrg_pc_205, extracted 2 September 2026, dataset last updated 11 August 2026.

North America over the same period, for contrast: the United States industrial price rose from 6.67 cents in 2020 to 8.62 in 2025 and 8.89 year to date through June 2026, a 33 per cent rise (EIA Table 5.3). Montreal's large-power rate rose from 5.24 Canadian cents in 2021 to 5.83 in 2025, an 11 per cent rise (Hydro-Quebec). Alberta spiked to 23.77 Canadian cents in 2023 and fell back to 8.02 by 2025.

C5. Residential prices, kept separate on purpose. These are a different and much higher number and must not be compared with anything above.

The last row is the number most often quoted as "European industrial electricity prices." It is 68 per cent higher than the band IF figure for the same countries in the same half-year, because it covers small businesses and includes taxes. Using it to describe a data centre's cost is wrong by more than half.

C6. Data-centre specific tariffs and large-load rules. The important trend is that regulators everywhere are creating a separate price and a separate queue for data centres, which means the general industrial tariffs above are increasingly the wrong number for a new AI facility.

Texas and ERCOT belong in this table and are not fully covered here. ERCOT's industrial price of 6.65 United States cents makes it the cheapest large United States market, its anticipated reserve margin of 28 to 30 per cent gives it the most headroom of any large North American system, and Texas passed large-load legislation in 2025 imposing curtailment obligations on data centres. The specific thresholds and curtailment terms of that legislation could not be verified from a primary source within this dimension and should be checked against the Public Utility Commission of Texas and ERCOT protocols before being quoted.

D. Data-centre capacity and where the AI capacity is going

Read the metric column before quoting anything. The four main research houses measure four different things and their global totals differ by nearly a factor of two.

E. Sovereign and national AI compute programmes

Where Canada sits

The four questions, answered as rankings.

Is Canada cheap? Yes, and it is one of the cheapest in the developed world, but it is not the cheapest and the advantage is concentrated in four provinces.

On the most defensible like-for-like comparison, Statistics Canada's implied industrial price against the EIA's, Canada was at 4.91 United States cents per kWh in 2024 against the United States at 8.13. Canada was 40 per cent cheaper. On the Hydro-Quebec comparison of posted large-power tariffs, Montreal at the 50,000 kW profile was 3.85 United States cents against Houston at 8.99, Toronto at 8.82, New York at 15.95 and Boston at 17.13. But on the Eurostat band IF and IG measures for large industrial consumers in Europe, Finland at 4.70 and Norway at 4.61 to 5.28 United States cents undercut Canada's national average. Ranking the large-industrial prices in this brief from cheapest to most expensive: Norway (4.61 band IG), Finland (4.70), Canada national average (4.91), Norway band IF (5.28), Texas (6.65), France (6.86 band IG), Sweden (7.21 band IG), United States average (8.62), China (about 8.7), Japan (about 11.7), European Union band IF average (12.36), South Korea (about 13.2), Germany (14.49), Ireland (18.97). Canada is third. Quebec alone, at 3.59 United States cents, would be first.

Two honest caveats. Alberta, the province with by far the largest data-centre pipeline, is Canada's most expensive at 7.31 United States cents industrial, above the Nordics and above Texas. And Quebec has proposed a data-centre specific rate of about 13 Canadian cents per kWh from late 2026, roughly double its industrial rate, which would erase most of Montreal's headline advantage for new AI load.

Is Canada clean? Yes, and it is a genuine top-five country, but not top-three and not uniform.

Canada is fifth in the world on non-emitting share among countries generating more than 100 TWh, at 77.0 per cent in 2025, behind Norway (99.0), Sweden (98.8), France (94.8) and Brazil (88.7). On Ember's lifecycle carbon intensity it ranks twelfth of the 48 countries generating more than 50 TWh, at 190.7 grams, behind Norway (28), Sweden (35), Switzerland (39), France (42), Finland (58), Belgium (109), Brazil (110), Austria (117), Portugal (128), Spain (154) and Colombia (187). On Canada's own operational method the number is about 86 grams, which would place it much higher. The provincial spread from 1.7 grams in Quebec to 730 in Saskatchewan is a factor of more than 400 and makes the national figure meaningless for siting.

Canada is unambiguously cleaner than every large competitor it actually competes with: the United States at 384 grams, China at 526, Japan at 477, Korea at 417, Malaysia at 603, Singapore at 497 and Saudi Arabia at 692. Against the Nordics and France, Canada loses.

Is Canada big? No.

Canada is seventh in the world by generation with 2.06 per cent of the world total, and tenth by capacity with 1.53 per cent. It is smaller than Russia, Japan and India and about the same size as South Korea, a country with 40 per cent more people on a quarter of the land. China generates 16 times as much electricity as Canada. The United States generates seven times as much. The more telling comparison is the rate of change: China added 434 GW of solar and wind capacity in 2025, 2.8 times Canada's entire fleet, while Canada added 1.1 GW of non-emitting capacity, or one four-hundredth of China's addition. Canada is not big and it is not adding capacity quickly.

Does Canada have spare room? Yes in Alberta, British Columbia and Saskatchewan. No in Quebec, Manitoba and the Maritimes.

NERC's 2025 Long-Term Reliability Assessment, which covers Canadian and United States areas on the same basis, gives Alberta an anticipated reserve margin of 36.2 per cent in 2026-27 falling only to 33.0 per cent in 2030-31, against a reference margin level of 11.8 to 17.6 per cent. Saskatchewan runs at 25.9 to 35.0 per cent against 15.0. British Columbia is 19.1 to 24.2 per cent against 11.6 to 12.1. Those three have real headroom.

Quebec does not. Its anticipated margin falls from 15.6 per cent in 2026-27 to 11.5 per cent in 2030-31 against a reference level of 12.2 per cent, and NERC's own summary states that "demand growth projections are outpacing planned resource additions, leading to projected resource shortfalls in the winter season." Manitoba collapses to 1.2 per cent by 2030-31 against a 12.0 per cent target. The Maritimes sit below their 20 per cent target in the first two years. Ontario is adequate but thin, at 18.7 per cent in 2029 against a 15.8 per cent reference.

For context, the United States comparators are worse in the two biggest markets: MISO falls to 4.3 per cent by 2030 against an 8.5 per cent reference, and PJM to 13.9 per cent against a reference rising to 26.3 per cent. ERCOT and California have large margins. So Canada's spare-room position is better than PJM's and MISO's, and worse than ERCOT's and California's, and inside Canada the headroom sits in exactly the province where the power is dirtiest and dearest.

The uncomfortable synthesis: Canada's competitive advantage is real but it is a Quebec, Manitoba, British Columbia and Newfoundland advantage, and those systems are either short of reserve margin or committed. The province with room to grow is the one where the pitch of cheap, clean power does not hold.

Contested or uncertain figures

1. Carbon intensity for Canada: 190.7 or 86 grams? Ember reports 190.7 grams of CO2 per kilowatt hour for 2025 on a lifecycle basis; Environment and Climate Change Canada reports about 86 grams of CO2 equivalent for 2024 on an operational basis. Reverse-engineering Ember's own data file confirms it assigns 23.8 grams to hydro, 47.6 to solar, 11.7 to wind and 5.3 to nuclear, and uses a gas factor of 593 grams globally and 722 for Canada, well above direct combustion. Use Ember for country comparisons and ECCC for Canada alone. Never put them in the same column.

2. Price comparisons are sensitive to four things at once: taxes, exchange rate, consumption band and year. Eurostat band IC (500 to 1,999 MWh, including non-recoverable taxes) gives an EU average of 18.37 euro cents. Band IF for the same half-year excluding taxes gives 10.95, and band IG gives 9.98. That is a 45 per cent range on the same continent in the same six months, driven entirely by definition. Add exchange rate volatility (the Canadian dollar moved 5.4 per cent against the United States dollar in the year to April 2025 by Hydro-Quebec's own reckoning; the euro moved further) and the year of the data, and a careless comparison can be wrong by a factor of two in either direction. Every price in this brief carries its band, its tax treatment, its year and its conversion rate.

3. The IEA's "United States" industrial price is Texas. The IEA's chart of estimated final electricity prices for large industrial customers in energy-intensive industries uses Texas as its United States reference point and Eurostat's above-150 GWh band for Europe. Texas was 6.65 United States cents year to date through June 2026, 23 per cent below the national industrial average of 8.89 and 64 per cent below New England's 18.31. The IEA's headline that European prices are "more than double" United States prices is true against Texas and would be a smaller multiple against the United States average.

4. Ember's installed-capacity totals are constructed, not registry data. Ember's total is the sum of its clean and fossil capacity series. It agrees with Statistics Canada for Canada to within one per cent (156.0 against 156.4 gigawatts) which is reassuring, but it should not be treated as a definitive nameplate figure for every country, and 2025 capacity data are available for only about 90 countries against 195 for generation.

5. Data-centre capacity by country is the weakest data in this brief. Synergy, JLL, Cushman and Wakefield and IDCA report global totals of 122.2, 103, 43.4 (Americas only) and 67.7 gigawatts respectively, on four different definitions. A figure of about 5.5 gigawatts for Germany and about 2.0 for the United Kingdom circulates in aggregated coverage but could not be verified and is inconsistent with JLL's 3.8 gigawatt total for all five FLAP-D markets combined. No credible free national total exists for Korea, and Canada's roughly 1.4 gigawatts comes from DC Byte via a single law-firm summary. The United States ranges from 17.2 to 50 gigawatts depending on whose definition is used.

6. Statistics Canada's implied industrial price is average revenue, not a tariff. It is value of sales divided by quantity sold for the mining and manufacturing class. It blends contract and posted rates, includes some self-generation arrangements differently by province, and is not a price a new customer could sign. It is nevertheless the only official, consistent, province-by-province Canadian series, and its EIA counterpart is constructed the same way, which is why the two are compared here.

7. Quebec's proposed 13 cent data-centre rate is a proposal. It has been reported and is discussed in the companion dimensions of this pack, but its final level, scope and approval status are not settled. If approved at that level it would put Montreal's data-centre price above Toronto's current large-power rate and roughly at the level of the European Union band IF average, which would materially change the Canadian case.

8. Japan, Korea and China industrial prices in this brief are indicative, not official tariffs. The Japanese figure of about 17.5 yen per kilowatt hour and the Korean figure of about 182.7 won come from press and price-aggregator sources, not METI or KEPCO tariff sheets. The Chinese 36-city average of 0.62 yuan is an official NDRC monitoring series but covers supply at 35 kilovolts and above generally, not a data-centre tariff, and Chinese provincial rates vary from about 0.49 to 0.68 yuan. Treat all three as within roughly 20 per cent.

9. Sovereign programme money is not comparable across countries. Stargate's 500 billion United States dollars is a private consortium pledge. France's 109 billion euros is pledged foreign and private investment. The European Union's 200 billion euros is a mobilisation ambition. India's 1.2 billion United States dollars is an appropriated budget that is being under-spent. The United Kingdom's 2 billion pounds is appropriated. Adding these together produces a meaningless number.

10. Reserve margins are forecasts made by system operators with incentives. NERC's Anticipated Reserve Margin counts Tier 1 resources only and reflects each operator's own demand forecast. Since data-centre load forecasts are themselves contested (the IEA finds only about 20 per cent of United States connection requests materialise), the margins for areas with large speculative queues, notably Alberta, may be more comfortable than they look, or less, depending on which load actually shows up.

Open questions students could investigate

1. At what data-centre tariff does Quebec stop winning? Build the model: take Hydro-Quebec's 50,000 kW large-power rate of 5.52 Canadian cents, apply the proposed 13 cent data-centre rate, and find the power-usage-effectiveness and utilisation combination at which Montreal loses to Texas at 6.65 United States cents, to Alberta at 10.01 Canadian cents plus behind-the-meter gas, and to Norway at 4.61 United States cents.
2. Canada added 1.1 gigawatts of non-emitting capacity in 2025 and has a 20 gigawatt data-centre pipeline. How many years of Canada's current build rate would that pipeline consume, and what would the build rate have to be to serve even one quarter of it by 2032?
3. Does the Ember lifecycle carbon intensity or the ECCC operational intensity better reflect what a hyperscaler's Scope 2 accounting will actually recognise, and which one drives siting decisions in practice? Check the greenhouse gas protocol location-based and market-based rules against both.
4. Reserve margin is a peak-demand metric. Data centres are baseload. Construct an energy-adequacy measure rather than a capacity-adequacy one for Quebec, Manitoba and British Columbia, and test whether the NERC warning about Quebec's winter shortfalls actually binds on a 24/7 industrial load.
5. Ireland's large-industrial price is the most expensive in the European Union and it also has the highest data-centre share of national electricity in the world at about 22 per cent. Is that causal, and if so what is the elasticity? What does the Irish experience predict for Alberta at 20 gigawatts?
6. The competitive set of cheap and clean jurisdictions is Norway, Sweden, Finland, France and four Canadian provinces. How much unallocated capacity does each actually have, and what is the combined ceiling on genuinely cheap clean AI capacity worldwide? Is it 5 gigawatts or 50?
7. Announced sovereign AI capacity is roughly 25 to 30 gigawatts against under 1 gigawatt live. Build a base-rate model from historical megaproject delivery in energy and infrastructure, and estimate what fraction of announced capacity is likely to exist by 2030.
8. Alberta's industrial price of 10.01 Canadian cents is set in a competitive pool market while Quebec's 4.92 is set by a regulated monopoly. Which structure serves a 500 megawatt data centre better over a 15-year contract, and what does the 2023 Alberta spike to 23.77 cents and subsequent collapse to 8.02 say about price risk?
9. Canada exports electricity to the United States. If Canadian power is genuinely cheaper and cleaner, why is the marginal Canadian electron worth more exported than sold to a domestic data centre, and what does that imply about the value Canada captures from AI load versus exports?
10. China added 434 gigawatts of solar and wind in 2025 while its grid intensity remained at 526 grams. Model how long China's additions must continue at that rate before Chinese AI data centres are cleaner than Alberta's, and what that means for the durability of Canada's clean-power pitch.

Sources

Narrative

Start with the unit. A data centre is sized in megawatts because everything inside it, from the graphics processing units (GPUs) to the pumps and fans, turns electricity into heat, continuously. A conventional cloud facility considered by the International Energy Agency (IEA) is about 25 MW; a hyperscale site is 100 MW; the largest campuses under construction are around 2,000 MW and the largest planned about 5,000 MW. One MW running all year is 8,760 MWh, which is the electricity of roughly 750 average Canadian homes, or 450 Quebec homes that heat with electricity, or 1,160 Ontario homes that mostly heat with gas (NRCan, 2022 data). That is why a single 1 GW campus, at a 90% load factor, would use about 7.9 TWh a year, 1.3% of everything Canada generated in 2024 and about 9% of Alberta's 2025 output.

The AI difference is density and duty cycle. Traditional data centres serve web traffic, cluster near users to keep latency low, and run at moderate utilization. AI training clusters pack tens of thousands of GPUs into tightly coupled racks and run flat out for weeks; the IEA notes that training "is less latency-sensitive than traditional data centre workloads," which is why training sites can sit far from cities, next to power. Inference, the serving of trained models, is more mixed: a chatbot query tolerates tens of milliseconds and can be served from a distant region, while autonomous driving or trading cannot. In practice, the industry is separating remote, power-led training campuses from smaller, metro-located inference and colocation capacity, with edge computing on devices at the extreme.

Density drives cooling. For fifteen years a rack drew 4 to 10 kW and air cooling was enough. Uptime Institute's 2025 survey finds the typical rack still at about 9 kW and more than 80% of operators with nothing above 30 kW. An Nvidia GB200 NVL72 rack draws around 120 kW, and hyperscale halls now support over 135 kW per rack. Air cannot move that heat, so liquid is piped directly to the chips and rejected outdoors through dry coolers, cooling towers or evaporative systems. The choice sets the water bill: closed-loop designs use almost nothing (Meta's Alberta campus is designed to consume no water in normal operation), while evaporative cooling can consume 1.9 to 2.5 litres per kWh of IT load. Lawrence Berkeley National Laboratory (LBNL) puts the US average at 0.36 L/kWh in 2023, rising as liquid-cooled AI grows. A 100 MW IT load at 85% utilization therefore uses anywhere from near zero to about 1.4 billion litres a year depending on design, a fact that matters in Alberta's dry south and much less in Quebec or coastal BC.

Efficiency is summarized by power usage effectiveness (PUE), total facility energy divided by IT energy. The industry average of 1.54 has not moved in six years, but the frontier has: Google reports 1.09, hyperscale facilities average 1.14 worldwide, and new high-latitude builds routinely reach 1.3 or better. Cold climates help because economizers can use outside air or cold water for most of the year; Calgary has ten months with mean temperatures at or below 15 degrees C, Winnipeg, Edmonton and Quebec City nine, Montreal, Toronto and Vancouver eight. The gain is real but bounded: moving from PUE 1.5 to 1.2 cuts overhead by 30 percentage points of IT load, whereas the IT load itself is fixed by the chips. Climate improves the operating bill; it does not shrink the megawatts the utility must deliver.

Those megawatts are the hard part. Transformers, switchgear and substations are the bottleneck of the decade: Wood Mackenzie's mid-2025 survey shows power transformers averaging 128 weeks and generator step-up units 144 weeks, with prices up 77% since 2019. Grid connection queues run one to three years in the US on average, up to seven in Northern Virginia, up to ten in the Netherlands, and Dublin has paused new connections until 2030 (IEA). The IEA's 2026 update states plainly that grid waits of five to ten years and turbine deliveries of about five years are now the pacing items. Developers respond with behind-the-meter gas plants, batteries and long-dated nuclear deals, but each has its own queue: GE Vernova ended 2025 with an 80 GW turbine backlog and expects to be sold out through 2030; reliable onsite gas must be overbuilt by 30 to 70%; the first Canadian small modular reactor at Darlington costs $7.7 billion for 300 MW and arrives in 2030. Capital intensity follows: about US$11 million per MW to build the shell and power systems, up to US$25 million per MW more for AI compute, and in Canadian terms C$30 to 40 million per MW fully loaded.

Against that backdrop Canada's resource position is strong on paper. The grid is 78% non-emitting, hydro alone supplies 55% of generation, and the country exported 36 TWh to the United States in 2024. Industrial power in Winnipeg (5.09 cents per kWh) and Montreal (5.52) is the cheapest of 22 North American cities benchmarked by Hydro-Quebec, roughly a quarter of Boston or San Francisco and less than half of Toronto or Seattle. Canada has gas (17.9 Bcf/d, the Montney alone 8.1 Bcf/d), water, land, a critical minerals strategy covering copper and rare earths, a trans-Pacific fibre landing at Port Alberni and Vancouver (Topaz, 240 Tbps), and a construction workforce of 1.6 million.

The catch is that surplus has become scarce almost everywhere at once. Quebec, long the continent's exporter, was a net importer in 2023 after drought, exported only 5.9 TWh to the US in 2024, and plans 60 TWh and 8,000 to 9,000 MW of new supply by 2035 at a cost of $155 to 185 billion; it now requires ministerial authorization for any load above 5 MW and proposes a data centre rate of 13 cents per kWh, about double today's large-power rate. British Columbia imported 14.5 TWh from the US in 2024, has just absorbed Site C's 1,100 MW, and is offering a total of 400 MW to AI and data centres over two years through a competitive call that was roughly twice oversubscribed. Manitoba's hydro is spoken for by export contracts and needs new capacity by 2029; it screens every request above 5 MW and has already rejected one 100 MW proposal. Ontario plans for demand to grow 65% to 250 TWh by 2050, with data centres 13% of new demand by 2035, and is betting on refurbished and new nuclear while legislating priority for data centres that bring jobs and domestic hosting. New Brunswick faces a capacity shortfall by 2028 before any data centre connects; Nova Scotia still burns coal and says supply comes first.

Alberta is where the numbers are largest and the fuel is different. Its deregulated market delivered the lowest pool price in eight years in 2025 ($43.68/MWh) and has a supply cushion, but 20.7 GW of data centre requests against a 12.8 GW peak cannot be served by the grid; only 1,200 MW was allocated for 2027/28, and the 2026 Data Centre Regulation now gives priority to "tethered" projects that bring their own generation. Meta's Sturgeon County campus is the template: 1 GW growing to 1.8 GW, a dedicated 932 MW combined-cycle gas plant, closed-loop cooling with no water draw, and a 2030 in-service date for the power plant. The province can host gigawatts, but on gas, which is what the Canada Energy Regulator's scenarios, Ottawa's emissions rules and the tech companies' own clean-energy commitments will all have to reconcile.

For managers the arithmetic is simple. Every 100 MW of AI capacity needs a 110 to 150 MW grid connection, a substation whose transformers take two to three years, water or dry-cooling choices that fix the local footprint, and either a utility with firm surplus or a power plant of its own. Canada has the cheapest clean electrons in North America in Manitoba and Quebec, the largest gas-backed pipeline of projects in Alberta, the deepest nuclear build-out in Ontario, and almost no uncommitted capacity anywhere before 2028 to 2030. The opportunity is real; the constraint is time, iron and firm megawatts, and the provinces that move fastest on new supply and connection rules will capture it.

Physics for managers

1. Power, energy, time. A megawatt (MW) is a rate; a megawatt-hour (MWh) is a quantity. 1 MW for one hour = 1 MWh; 1 MW for a year = 8,760 MWh = 8.76 GWh; 1 TWh = 1,000 GWh = 1,000,000 MWh. (Definitions; Canada generated 622.2 TWh in 2024, Statistics Canada.)
2. Data centres are near-baseload. Modelled load factors are about 50% for the fleet and higher for AI training clusters; a large training run was estimated at an 84% load factor (IEA Energy and AI, 2025, citing Shehabi et al.). Utilities must plan for the nameplate MW around the clock, unlike a factory with shifts.
3. Servers eat most of the power. Servers with CPUs and GPUs take about 60% of a modern data centre's electricity, storage about 5%, networking up to 5%, cooling 7% to over 30% depending on efficiency (IEA, 2025).
4. PUE = total facility energy divided by IT energy. Industry average 1.54 (Uptime, 2025); hyperscale world average 1.14 in 2024 falling to about 1.10 by 2030 (IEA, 2025); Google 1.09 (2025); US fleet 1.4 in 2023 heading to 1.15 to 1.35 by 2028 (LBNL, December 2024). A 100 MW IT load needs 154 MW at PUE 1.54 and 110 MW at 1.10.
5. Water usage effectiveness (WUE) is litres per IT kWh. US average site WUE about 0.36 L/kWh (360 L/MWh) in 2023, rising to 0.45 to 0.48 by 2028 as liquid-cooled AI grows (LBNL, 2024); evaporative cooling can reach 1.9 to 2.5 L/kWh (EESI, June 2025); closed-loop hyperscale designs report 0.1 to 0.3 L/kWh (LBNL) and Meta's Alberta campus is designed to use no water in normal operation (Data Center Frontier, July 2026).
6. Racks went from 4 to 10 kW to 100 kW plus. Typical racks are 10 to 30 kW today; AI training is the main driver only above 60 kW per rack (Uptime, 2025). A GB200 NVL72 rack is about 120 kW and hyperscale halls now support over 135 kW per rack (Nvidia, 2025). Above roughly 30 to 50 kW air cannot carry the heat away; direct-to-chip liquid cooling becomes mandatory.
7. Cold air is free cooling. Air-side or water-side economizers replace compressor cooling whenever outdoor air is cool and dry enough. New facilities at high latitudes in North America and Europe often achieve PUE 1.3 or better (Uptime, 2025). Monthly mean temperatures in Canadian cities are below 15 degrees C for 8 to 10 months a year (ECCC normals, 1991-2020).
8. Capacity factor is output divided by nameplate over a year. Hydro and nuclear can run 80 to 95%; wind counts for only about 15% of its nameplate toward Quebec's winter capacity need (Hydro-Quebec Action Plan 2035) and 20% falling to 5% in Nova Scotia at high penetration (IESO Nova Scotia 2026 outlook). Data centres need firm capacity, so wind must be paired with hydro, gas, storage or nuclear.
9. Interconnection takes years because iron takes years. Power transformers 128 weeks, generator step-up transformers 144 weeks, switchgear 44 weeks, with power-transformer prices up 77% since 2019 (Wood Mackenzie Q2 2025 via POWER, January 2026). Grid queues: 1 to 3 years in the US on average, up to 7 in Northern Virginia, up to 10 in the Netherlands, paused in Dublin until 2030 (IEA, 2025, Table 2.4).
10. Behind-the-meter gas is a bridge with its own queue. GE Vernova expected an 80 GW turbine backlog at end 2025 stretching into 2029 and sold out through 2030 (Utility Dive, December 2025). Reliable onsite gas must be overbuilt by 30 to 70% relative to the data centre load (IEA, 2026). Diesel generators remain the standard backup for outages, not a source of daily energy; 20 to 25 GW of batteries could sit inside data centres globally by 2030 (IEA, 2026).
11. Small modular reactors are real but slow and costly. Darlington's four BWRX-300 units (1,200 MW) are budgeted at $20.9 billion, $7.7 billion for the first 300 MW unit in service in 2030, about 14.9 cents per kWh projected (World Nuclear News, May 2025). Global conditional SMR offtake agreements rose from 25 GW to 45 GW in about a year (IEA, April 2026).
12. Capital cost scales with MW. Construction runs about US$10.7 million per MW in 2025 and US$11.3 million in 2026, excluding tenant fit-out that can reach US$25 million per MW for AI (JLL, January 2026); Canadian shell-and-core C$7 to 12 million per MW and C$30 to 40 million per MW fully loaded with compute (Encor Advisors, June 2026); Meta's Alberta campus is over C$13 billion for 1 GW plus C$4 to 4.6 billion for its 932 MW gas plant (Data Center Frontier, July 2026).

The picture in numbers

Table 1. Data centre fundamentals

Table 2. Canada's electricity system, 2024

Table 3. Province by province (generation mix 2024 from StatCan 25-10-0015; industrial price 2025 from Hydro-Quebec for a 50,000 kW, 85% load factor customer; policy status as of August 2026)

Non-emitting share here counts hydro, nuclear, wind, solar and tidal only; biomass is excluded. The StatCan monthly table totals 609.5 TWh for 2024 against 622.2 TWh in the annual survey; shares are internally consistent within each table.

Table 4. Industrial electricity prices, 50,000 kW customer, 85% load factor, 30.6 GWh per month, rates in effect 1 April 2025 (Hydro-Quebec, https://www.hydroquebec.com/data/documents-donnees/pdf/comparaison-prix-2025-en.pdf)

For a 5,000 kW customer at 85% load factor the same ranking holds: Montreal 5.83, Winnipeg 6.00, Calgary 8.02, Vancouver 8.42, Regina 9.57, Edmonton 10.17, Ottawa 12.75, Toronto 12.80, Halifax 13.43 cents per kWh. Alberta's 2025 average pool price of $43.68/MWh (4.37 cents) is energy only; the Hydro-Quebec figures for Calgary and Edmonton are estimated all-in bills.

Table 5. Climate and free cooling, monthly mean temperature, ECCC normals 1991-2020 (via Wikipedia city climate tables and the Calgary weatherbox template)

The 15 degree threshold is an analyst simplification for comparing cities; actual economizer hours depend on humidity, design supply temperature and cooling architecture. A Montreal operator claims outside-air economization up to 80% of the year (Vantage Data Centers).

Table 6. Alberta market snapshot, 2025 (AESO Annual Market Statistics, March 2026, https://www.aeso.ca/assets/Annual-Market-Stats-2025.pdf)

Contested or uncertain figures

• Land per MW. No primary source with a defensible acres-per-MW figure could be retrieved; industry rules of thumb vary widely with density and the power plant footprint. Only building-floor anchors were found (Vantage Montreal I: 11 MW of IT in 60,000 square feet). Treat any acreage figure as unverified.
• Alberta's queue. AESO reported 16,229 MW (29 projects, June 2025) and 20.7 GW (September 2025); a secondary tracker counts 19.6 GW (July 2026) and 9,762 MW "announced." These are requests, not commitments; the queue median slip is reported at 416 days. DeSmog's emissions (33 to 63 Mt CO2e for 15.5 GW) and water (200 billion litres for 20 GW) figures are third-party estimates.
• CER data centre range. The 1.5 to 12 GW by 2050 figure comes from a Canadian Press report on the April 2026 release; the CER results page itself gives about 100 TWh (a quarter of growth) in the Higher scenario. Both are scenario outputs, not forecasts.
• Generation totals. StatCan's monthly table (25-10-0015) sums to 609.5 TWh for 2024 while the annual survey (25-10-0020/21) reports 622.2 TWh; shares in Table 3 come from the monthly table. CER provincial profiles use 2021 data and split gas from coal; StatCan does not.
• Household conversions. NEUD household electricity (2022) ranges from 7,300 kWh in Alberta to 19,600 kWh in Quebec; "homes per MW" changes by a factor of 2.7 depending on the province chosen, and assumes 100% load factor.
• PUE and WUE. Survey PUE is self-reported and skewed toward larger operators; Google's 1.09 is a fleet average across mild and cold sites. WUE conventions differ (site versus source water; litres per IT kWh versus per facility kWh).
• Capital cost per MW. JLL (US$10.7 to 11.3 million construction), Encor (C$7 to 12 million shell, C$30 to 40 million fully loaded) and project anchors (Meta C$13 million per MW campus; Darlington SMR C$17 million per MW average, C$26 million first unit) define scope differently; compare only like with like.
• Free cooling hours. The month counts in Table 5 use a 15 degree mean-temperature proxy; the Vantage "80% of the year" claim is a vendor statement.
• Churchill Falls values. The 2024 MOU (5.9 cents per kWh average; $225 billion nominal) and the August 2026 agreement ($49 billion 2026 NPV; $273 billion nominal) use different bases and are politically contested in Newfoundland and Labrador.
• Labour gap. BuildForce's July 2026 outlook shows a 34,300-worker shortfall by 2034; the prior-year outlook cited in search results showed 108,300. Neither isolates electricians.
• Nova Scotia and New Brunswick project sizes. The Lorneville data centre is reported at 190 MW (CBC, November 2025) and 390 MW (secondary tracker, August 2026).

Open questions students could investigate

1. If Quebec's proposed 13 cents per kWh data centre rate is approved, at what PUE and utilization does Montreal still beat a gas-backed Alberta site at 8 cents plus the capital cost of a dedicated plant?
2. What is the realistic firm surplus, in MW by year from 2027 to 2032, in each of BC, Manitoba, Quebec and Ontario once export contracts, electrification and committed industrial loads are netted out?
3. How much of Alberta's 20 GW queue would survive a federal clean electricity standard or a carbon price on behind-the-meter gas, and what would carbon capture add per MWh?
4. Which Canadian cities can support 100 kW liquid-cooled racks with closed-loop dry cooling year-round, and how many hours per year would compressor cooling still be required?
5. What would a 300 MW Darlington-type SMR cost per MWh for a data centre offtaker compared with the 14.9 cents per kWh OPG projection, and who carries construction risk?
6. Can Ontario's Industrial Conservation Initiative be redesigned so that flat AI loads provide peak flexibility (workload shifting, batteries) rather than simply paying more Global Adjustment?
7. How many electricians, power engineers and pipefitters does a 1 GW campus plus its substation and power plant require at peak, and can Canada's 306,200-worker hiring requirement absorb three or four such projects at once?
8. Is Churchill Falls and Gull Island power (2,350 MW retained in Newfoundland and Labrador) better used for data centres in Labrador, for export through Quebec, or for green industry, on a dollars-per-MWh and jobs-per-MW basis?
9. What fibre latency, in milliseconds, separates Montreal, Toronto and Calgary from New York, Chicago and Seattle, and which AI inference workloads can tolerate it?
10. How should provinces rank competing 100 MW applicants: jobs per MW, tax per MW, domestic data sovereignty, or willingness to be curtailed at peak?

Sources

Narrative

Start with one chip and one number. An accelerator is a specialised processor that does the arithmetic behind AI. The best known is NVIDIA's H100 graphics processing unit (GPU), and NVIDIA's own specification says its maximum thermal design power, the heat its cooling system must be built to remove, is up to 700 watts. That single figure is the seed from which every other number in this brief grows. A domestic kettle is about 1,500 watts, so two H100s running flat out are roughly one kettle, permanently on, for five years.

The escalation is the story, not the level. In 2016 NVIDIA's flagship data centre chip, the Tesla P100, drew 300 watts. The A100 in 2020 drew 400. The H100 in 2022 drew 700. In 2025 AMD published a maximum total board power of 1,400 watts for its Instinct MI355X, and NVIDIA's Blackwell Ultra is reported at the same figure. AMD is the cleanest illustration because it publishes one comparable number per part in a datasheet: 750 watts for the MI300X in 2023, 1,000 for the MI350X and 1,400 for the MI355X in 2025. That is a near doubling in two years, and it happened because performance per chip rose faster than efficiency per watt. Chips did become more efficient; they just got bigger faster.

Above roughly forty kilowatts a rack, air stops working. A rack is a standard cabinet about the size of a fridge. For twenty years the typical rack drew under ten kilowatts and was cooled by blowing cold air through it. Uptime Institute's July 2025 survey found more than 80% of operators still have no rack above thirty kilowatts and only 22% use direct liquid cooling. Meanwhile NVIDIA's GB200 NVL72, which packs 72 Blackwell GPUs and 36 Grace central processing units into one cabinet, draws about 120 kilowatts, and NVIDIA's own reference architecture states its successor, the GB300 NVL72, requires up to 142 kilowatts, fed by eight power shelves of 33 kilowatts each. Air physically cannot move that much heat. So the industry has moved to direct-to-chip liquid cooling, in which water or a coolant runs through a metal cold plate bolted onto the chip, capturing roughly 80% of the heat in a closed loop. Immersion cooling, in which whole servers sit in a bath of non-conductive fluid, handles even higher densities but is still rare. The practical threshold where liquid becomes mandatory is somewhere between thirty and fifty kilowatts a rack, and NVIDIA and AMD have settled the question for their customers by shipping their top parts as liquid-only. NVIDIA now says today's racks exceed 200 kilowatts and projects one-megawatt racks in 2027.

From one chip to one building. The manager's question is how many accelerators fit in a given amount of power. Build it in three steps. First, chip to server: NVIDIA's own DGX H200 puts eight 700-watt GPUs in a chassis rated at 10.2 kilowatts, a multiplier of about 1.8 once the host processors, memory, storage, fans and power supply losses are counted. In a rack-scale system that overhead is already inside the published rack figure, so 120 kilowatts divided by 72 GPUs is about 1,670 watts per GPU. Second, add the network switches and storage that sit outside the compute rack, roughly 8 to 10% more. Third, multiply by power usage effectiveness (PUE), the ratio of total building power to information technology (IT) power, which is about 1.15 for a new cold-climate liquid-cooled hall and 1.30 for a merely competent one. The answer is that one accelerator needs roughly 1.5 to 2.5 kilowatts of facility power, and a modern AI data centre holds roughly 400 to 500 accelerators per megawatt, about half a million per gigawatt.

Two real installations confirm it. xAI's Colossus in Memphis ran 100,000 NVIDIA Hopper GPUs, a figure NVIDIA itself published in October 2024, against 150 megawatts of grid supply approved by the local utility and the Tennessee Valley Authority. That is 1.5 kilowatts per GPU, within 10% of the bottom-up estimate. TELUS announced in May 2026 that its British Columbia cluster will hold over 60,000 NVIDIA GPUs against more than 150 megawatts by 2032: exactly 400 accelerators per megawatt, which is what the arithmetic gives for a Blackwell Ultra or Rubin class part. Larry Ellison's claim of more than 450,000 GB200 GPUs at Stargate Abilene against a 1.2 gigawatt grid interconnect implies 2.67 kilowatts per GPU, at the high end of the same band. The rule holds across three companies, two countries and three chip generations, which is the best evidence that it is not a coincidence.

What that costs. At Epoch AI's transparent estimate of US$25,000 per H100 and US$37,000 to US$43,000 per Blackwell part, a 100 megawatt campus contains roughly 43,000 to 48,000 Blackwell GPUs, which is about 725 GB200 NVL72 racks. At HSBC's estimated US$3 million per rack, that is about US$2.2 billion of silicon and racks before a single wall is built. Two independent checks agree: JLL puts AI technology fit-out at up to US$25 million per megawatt, that is US$2.5 billion for 100 megawatts, and Jensen Huang's stated US$50 to 60 billion per gigawatt with about US$35 billion of NVIDIA content implies US$3.5 billion of NVIDIA content per 100 megawatts. Against that, the electricity is almost cheap: C$43 million a year at Manitoba or Quebec industrial rates. The capital is the constraint, and the capital sits in a five-year asset.

Which is why depreciation is the real argument. Two haircuts sit under every headline GPU count. The first is operational. Meta's own Llama 3 paper reports model FLOPs utilisation, the share of a cluster's theoretical arithmetic actually used, of 38% to 43% on up to 16,000 H100s, with effective training time above 90%. So the compute actually delivered is roughly 35 to 39% of the nameplate number in the press release. The same paper records 466 job interruptions in 54 days, with about 78% traced to hardware and 220 of them to failed GPUs or failed high-bandwidth memory: a failure roughly every three hours. The second haircut is financial. In the same twelve months, Amazon shortened the assumed useful life of a subset of its servers from six years to five, explicitly citing the pace of AI, which added about US$1.4 billion of depreciation and cut fiscal 2025 net income by about US$1.0 billion. Meta went the other way, lengthening most servers to 5.5 years, which reduced depreciation by about US$2.92 billion and lifted net income by about US$2.59 billion. Alphabet uses six years, Microsoft two to six, CoreWeave six. Michael Burry's November 2025 argument was that the real economic life is two to three years and that the industry is therefore understating depreciation by roughly US$176 billion between 2026 and 2028. The counter-argument is that older accelerators cascade to cheaper work rather than becoming scrap, and that A100s from 2020 still rent. Both sides are arguing about the same unknown, which is how fast a 700-watt chip becomes uneconomic when a 1,400-watt chip delivers five times the work. For anyone financing a Canadian data centre, this single assumption swings the return more than the electricity price does.

The units, one last time. A megawatt is a rate; a megawatt-hour is an amount. A gigawatt data centre in Canada would consume 7.88 terawatt-hours a year at a 90% load factor, which is 1.26% of the 625.2 terawatt-hours Canada generated in 2025 and about 1.4% of the 71 gigawatts the entire country produces on average. It would hold roughly half a million accelerators, about six thousand liquid-cooled racks, and about US$20 billion of silicon. It would cost between C$434 million and C$1.02 billion a year to power depending on where in Canada it sits. Those five numbers, held together, are enough to interrogate almost any claim made about this industry.

The units, in plain language

Everything in this brief reduces to one distinction. Power is a rate: how fast energy is being used, measured in watts. Energy is an amount: how much was used, measured in watt-hours. A grid operator sizes wires, transformers and generators against the rate. A finance team pays for the amount. Confusing the two is the most common error in public discussion of data centres.

Power: the rate

Nameplate, or capacity, is the maximum rate the equipment is designed to draw or produce. It is what the utility must be able to deliver at any instant.

Energy: the amount

The bridge between rate and amount is time, and it is exact:

• 1 MW sustained for 1 hour = 1 MWh.
• 1 MW sustained for a full year (8,760 hours) = 8,760 MWh = 8.76 GWh.
• 1 GW sustained for a full year = 8,760 GWh = 8.76 TWh.

Data centres run close to baseload, near their maximum almost all the time, unlike a factory with shifts. At a 90% load factor, a realistic planning figure for an AI campus, one MW delivers 7.884 GWh a year rather than 8.76.

Worked conversions a manager can use

All rows assume a 90% load factor. Homes use the Natural Resources Canada figure of about 11,650 kWh per Canadian household per year. The Canadian share uses Statistics Canada's 625.2 TWh of total generation in 2025. Costs are Canadian cents per kilowatt-hour at the two ends of the Canadian industrial range: 5.5 cents, roughly Winnipeg (5.09) or Montreal (5.52) for a 50 MW customer, and 13 cents, roughly Toronto (12.66) or Halifax (13.43), from Hydro-Quebec's price comparison at rates in effect 1 April 2025.

Three things fall straight out of that table. First, the electricity bill for a gigawatt campus is roughly half a billion Canadian dollars a year at cheap Canadian industrial rates and about a billion at expensive ones, so a 7.5 cent spread is worth about C$590 million a year. That is why siting is a board-level decision. Second, one gigawatt campus is 1.26% of everything Canada generates, and Canada's whole system averages only 71 GW, so the 20.7 GW of Alberta data centre applications reported in September 2025 is not a rounding error. Third, at a 100% load factor the homes figure is 752 per MW, so the common shorthand "a megawatt powers 750 homes" is right for a facility running flat out and about 10% too high at a realistic load factor.

Capacity, energy and peak demand, and why the two audiences differ

• Capacity (MW) is what the utility must be able to deliver at any instant. It sets the transformer, the substation, the line and the reserve margin. It is the thing that takes five to ten years to get in a congested interconnection queue.
• Energy (MWh or TWh) is what is metered and billed. It sets fuel burn, emissions and the operating cost line in the model.
• Peak demand is the highest instantaneous draw over a period. For a normal industrial customer, peak is much higher than average. For a data centre the two are almost identical, which is exactly why utilities find them hard: a data centre uses the expensive capacity all year rather than for a few hours.

A grid operator cares about the rate because a power system must match supply and demand instant by instant, and because building the rate is slow and lumpy. A finance team cares about the amount because the amount is what appears in the cash flow statement. The bridge between them is the load factor, and for AI the load factor is high, which is good for a utility's revenue per dollar of asset and bad for its ability to say yes.

Nine terms, each in a sentence or two

• Power factor. In an alternating-current system, real power (kW, which does useful work) divided by apparent power (kVA, which is what the wires and transformer must actually carry). A power factor of 0.9 means equipment must be sized about 11% larger than the kW figure suggests. Modern data centre power supplies run near 0.99, so kW and kVA are nearly interchangeable for AI halls, but the concept still governs equipment sizing.
• Load factor. Average load divided by peak load over a period. An office runs 0.3 to 0.5; an AI training campus runs 0.85 to 0.95.
• Capacity factor. The generator-side twin: energy actually produced divided by energy that would have been produced at continuous full power (US Energy Information Administration definition). Mid-latitude solar runs roughly 0.15 to 0.20. This matters in the frontier section, because the entire case for orbital compute rests on pushing solar capacity factor toward 1.0.
• N+1 redundancy. Install the number of units needed to carry the load (N) plus one spare, so any single unit can fail or be serviced without dropping the load.
• 2N redundancy. Two complete, independent systems, each able to carry the whole load. Twice the capital, and the basis of fault tolerance.
• Tier ratings. Uptime Institute's four-level classification of facility topology: Tier I basic, Tier II redundant components, Tier III concurrently maintainable (dual paths, so anything can be serviced without a shutdown), Tier IV fault tolerant (2N, survives any single failure). AI training halls are often deliberately built to a lower tier than a bank's data centre, because a training job can restart from a checkpoint whereas a payment system cannot.
• IT load versus facility load. IT load is the power drawn by servers, storage and network equipment. Facility load adds cooling, power conversion losses, lighting and everything else in the building. The utility bills facility load; the vendor quotes IT load.
• PUE, power usage effectiveness. Facility load divided by IT load, standardised as ISO/IEC 30134-2. A PUE of 1.0 would be a building that spends nothing on cooling. The global average has sat near 1.5 for years; best-in-class hyperscale runs 1.1 or better; a cold-climate liquid-cooled AI hall should reach 1.15 to 1.25.
• WUE, water usage effectiveness. Litres of water per kWh of IT energy. Closed-loop liquid cooling can drive it near zero; evaporative cooling in a hot climate cannot.

The picture in numbers

Table 1. Accelerator thermal design power by generation

Thermal design power (TDP) is the maximum sustained heat, and therefore power, the cooling solution must remove for that part alone. It excludes the host processor, storage, the network card and everything in the building. AMD publishes the equivalent as total board power (TBP).

Table 2. Rack and system power

Table 3. Where the power goes, and the units of a data centre

Table 4. Accelerators per facility, built up from watts

Method: take the accelerator TDP, apply the system overhead (NVIDIA's own DGX H200 is 10.2 kW for eight 700 W GPUs, a multiplier of about 1.8; in a rack-scale system the overhead is already inside the published rack figure), add 8 to 10% for scale-out networking and storage outside the compute rack, then multiply by PUE.

Table 5. Real installations, as a check on the arithmetic

Table 6. What the silicon costs, and what it is worth

The frontier, including space

Nothing in this section is a product. The most useful discipline is to sort every claim into three tenses: what has flown and worked, what has been filed with a regulator, and what has been said on a stage. The gap between the three is about three orders of magnitude.

What has actually flown

The first operational orbital AI computing constellation is Chinese, launched about six months before either Starcloud-1 or Google's paper. That is the most useful geopolitical fact in this section. Note also the honest scale: nanoGPT is a teaching-scale model, and Axiom's operating optical links run at 2.5 Gbps against Google's 1.6 Tbps bench demo.

What has been filed with a regulator

Those three filings total roughly 1.14 million satellites. There are about 11,000 operational payloads in orbit today and about 40,000 tracked objects (European Space Agency, Annual Space Environment Report, data to end 2024, published April 2025). A filing is an option bought cheaply; treat it as such.

What has been said

• Jeff Bezos, Italian Tech Week, Turin, 3 October 2025: "One of the things that's going to happen next is we're going to start building these giant gigawatt data centers in space. These giant training clusters will be better built in space because we have solar power there 24/7. There are no clouds, no rain, no weather. We'll be able to beat the cost of terrestrial data centers in the next couple of decades." His timeline: "10-plus years but I bet it's no more than 20 years." He said this five months before Blue Origin filed.
• Elon Musk, on X, about 30 October 2025: "Simply scaling up Starlink V3 satellites, which have high-speed laser links, would work. SpaceX will be doing this."
• Elon Musk, on X, 19 November 2025: "Starship should be able to deliver around 300 GW per year of solar-powered AI satellites to orbit, maybe 500 GW. The 'per year' part is what makes this such a big deal. Average US electricity consumption is around 500 GW, so at 300 GW/year, AI in space would exceed the entire US..."
• Elon Musk, June 2026, in a SpaceX-produced interview: the AI1 satellite is 20 m tall with a 70 m wingspan, 150 kW peak power, about 120 kW of average compute, in a 600 to 800 km orbit, at about 3 ms latency, with solar at 250 W per square metre and radiators at 1,400 W per square metre. A scaling ladder of about 1 GW a year by the end of 2027, then 10, then 100, then eventually 1 TW a year, roughly an order of magnitude a year.
• SpaceX and NVIDIA, 4 August 2026: "Starmind", putting NVIDIA's Vera Rubin NVL72 platform on the first AI1 satellites, prototype testing early 2027, first launches reported for the fourth quarter of 2027.

A precision point that matters. Starlink V2 mini downlink is about 100 Gbps and V3 is expected at about 1 Tbps downlink; the inter-satellite laser links are about 100 Gbps. Google's own paper independently states Starlink's laser system operates at about 100 Gbps over up to about 5,400 km. Quoting "1 Tbps lasers" conflates the downlink with the laser link, and is the most common error in coverage of this topic.

Google Project Suncatcher, the most rigorous public work

Announced 4 November 2025, with a paper submitted to arXiv on 22 November 2025 (arXiv 2511.19468) and a partnership with Planet Labs to launch two prototype satellites by early 2027. The measured results are genuinely new science.

Google's economic argument, and it is worth quoting precisely because it is so often over-read: at its own assumption of US$3,600 per kg for Falcon 9 the launched power price is about US$14,700 per kW-year; at US$200 per kg it falls to about US$810 per kW-year, against US terrestrial data centre power costs of about US$570 to US$3,000 per kW-year. Google concludes that at US$200 per kg "the cost of launch amortized over spacecraft lifetime could be roughly comparable to data center energy costs, on a per-kW basis". That sentence compares launch cost only against energy cost only. It excludes the satellite, the radiator, the chips and all ground infrastructure. Getting to US$200 per kg requires a sustained 20% learning rate and about 180 Starship launches a year, and Google notes this is "still far below stated launch targets for Starship".

The physics that decides it

The honest synthesis is more interesting than either the hype or the dismissal: inference may well move to orbit; frontier training almost certainly will not this decade. Van Berkel puts it directly: cost-effective frontier training in space "within the next 2-3 years is not credible and even a 10-year horizon appears optimistic with current technology trends".

One more signal is worth more than any forecast. In August 2026 SpaceX, the company proposing a million compute satellites, was building a natural gas power plant in Texas, hiring a natural gas trading team, and starting in-house turbine blade manufacturing to cut generator delivery delays by 18 months, all for terrestrial AI data centres. Watch what the capital does, not what the founder says.

The terrestrial frontier, which is where Canada actually competes

The pattern is that Canada's real frontier options are boring, slow and already on the books. None of them arrive faster than a gas turbine, and all of them arrive faster than a gigawatt in orbit.

Why this matters for Canada

Three hardware facts bear directly on where Canadian data centres get built.

Liquid cooling changes what the climate is worth, and mostly reduces it. The old Canadian pitch was free cooling: cold outside air means the chillers run less, so PUE falls and the power bill falls with it. That advantage is real and quantifiable. Cooling is 7% of load in an efficient hyperscale facility and over 30% in an inefficient one (IEA, April 2025), and the climate is worth roughly 0.2 to 0.3 on PUE. On a fixed 80 MW IT load, moving from PUE 1.45 to 1.20 removes 20 MW of facility draw, which at a 90% load factor is 158 GWh a year, worth about C$8.7 million at 5.5 cents per kWh and about C$20.5 million at 13 cents. But the AI racks arriving now are liquid cooled by vendor mandate: NVIDIA ships GB200 and GB300 NVL72 only as liquid-cooled racks, and AMD's MI355X is liquid only. Direct-to-chip cooling captures most of the heat in a closed loop and is far less sensitive to outside air temperature than an air-cooled hall. So the marginal value of Canadian cold falls just as the absolute value of Canadian electricity price and carbon intensity rises. The Canadian argument has to shift from "our air is cold" to "our electricity is cheap, clean and, where there is capacity, available", plus the one cold-climate bonus that liquid cooling creates rather than destroys: usable waste heat. Bell's Thompson Rivers University site feeds its heat into a campus district energy system, and TELUS claims its BC cluster will produce waste heat equivalent to heating 150,000 homes, a claim this research set has not verified. Warm water out of a cold-plate loop is worth something in Kamloops in February and nothing in Phoenix in July.

The chip supply chain is a tariff exposure, not just a procurement problem. No AI accelerator is made in Canada. The United States applies a 25% Section 232 tariff to a subset of semiconductors and derivative products, with exemptions for goods used in US domestic production including US data centres, and there is no CUSMA-compliant exemption (Government of Canada, Trade Commissioner Service, updated 27 August 2026). That asymmetry runs against Canadian siting: the same rack can be cheaper on the American side of the border. Add NVIDIA's own figure of up to 200 kg of copper busbar per megawatt-class rack and up to 200,000 kg of copper per gigawatt facility, and the equipment supply chain, rather than the chip, becomes the part with Canadian content: transformers, switchgear, busbar, cooling distribution units and copper. Canada mines copper; it does not make GPUs.

Orbital compute is not a threat to a Canadian siting decision this decade. Google's own analysis says orbital compute becomes cost-comparable to terrestrial energy costs only if launch reaches about US$200 per kilogram to low Earth orbit, which its learning curve puts in the mid-2030s and only if roughly 180 Starship launches a year are sustained. The only accelerator-class hardware demonstrated in orbit is a single NVIDIA H100 on a 60 kg satellite, and the strongest quantified objection, a roughly 1,000-fold bisection bandwidth deficit for distributed training, does not improve with cheaper launch. Every larger constellation is a regulatory filing or a statement. The right way for a Canadian decision maker to treat space is as a terminal-value question, not a siting question: it belongs in the residual value assumption for a 15 or 20 year asset, not in the choice between Kamloops and Sturgeon County.

Contested or uncertain figures

• The GB200 NVL72 rack does not have an NVIDIA-published power figure. NVIDIA's public specification table for GB200 NVL72 lists GPUs, FLOPS, memory and NVLink bandwidth, but no kilowatts. The widely used 120 kW comes from NVIDIA presentations, partner documentation and press coverage. Deployed racks are reported at 130 to 132 kW. NVIDIA's own enterprise reference architecture does state a figure for the successor, GB300 NVL72: up to 142 kW. Use the GB300 number when a primary citation matters.
• Per-GPU TDP for Blackwell is reported, not published. NVIDIA does not put a "1,200 W" or "1,400 W" line in a public datasheet the way AMD does. The B200 at 1,000 W air and 1,200 W liquid, and the B300 at 1,400 W, come from reporting and from partner documentation. AMD's MI355X at 1,400 W is the one directly citable vendor figure at that level.
• Rubin power is a supply-chain estimate. The figures of about 1.8 kW per GPU and 190 kW per rack (Max Q), rising to 2.3 kW and 230 kW (Max P), come from analyst Ming-Chi Kuo, not from NVIDIA. NVIDIA also renamed VR200 NVL144, counting GPU dies, to VR200 NVL72, counting packages, which makes cross-source comparison genuinely confusing.
• Google and AWS do not publish per-chip wattage. The Ironwood figure of about 1,080 W per chip in this brief is derived from Google's own "9,216 liquid cooled chips" and "nearly 10 MW" per pod. The commonly quoted 450 W for TPU v5p and 600 W for Ironwood are not Google figures. AWS has never stated a Trainium wattage; the inference that Trainium3 is above 1,000 W rests on AWS's own statement that chips above 1,000 W require liquid cooling and that Trainium3 does.
• xAI Colossus figures beyond 2025 are unreliable. The widely circulated 555,000 GPUs, US$18 billion and 2 GW for January 2026, and roughly 770,000 GPUs for mid-2026, trace only to aggregator blogs with no primary source. This brief uses NVIDIA's own 100,000 and 200,000 Hopper figures and Epoch AI's estimate of 946 MW and about 1.11 million H100-equivalents, with Epoch's stated bands of plus or minus 40% on power and 50% on compute.
• "H100-equivalents" is not a chip count. Meta's "nearly 600,000 H100s" and Epoch's H100-equivalent figures are compute-normalised measures. A facility with 400,000 Blackwell parts may report over a million H100-equivalents. Always ask which is being quoted.
• Stargate Abilene is an Oracle site, not a Microsoft site. The Stargate partners are OpenAI, Oracle, SoftBank and MGX, with Crusoe building on Lancium's campus. Microsoft's comparable asset is Fairwater in Wisconsin and Atlanta. The "450,000+ GB200" figure is a keynote claim by Larry Ellison, reported by Data Center Dynamics, not a filed disclosure. The "50,000 GB200 per building" figure that circulates is unsourced.
• TELUS has not published a Kamloops GPU count. The only published unit figure is "over 60,000" NVIDIA GPUs by 2032 across the whole British Columbia cluster (TELUS, 11 May 2026). A separate BetaKit report gives 75,500 GPUs across the three sites, which is cited in brief 07. The two do not reconcile and neither is a site-level Kamloops figure. TELUS's claims of 98% renewable power, 80% cooling energy reduction and waste heat equal to heating 150,000 homes are company claims; the last of these could not be verified in this research set.
• Bell's Merritt site is 7 MW, not 26 MW. The 26 MW figures are the Thompson Rivers University site (2026) and its Kamloops twin (2027). Bell's "upwards of 500 MW" is a six-site total, most of it not yet sited.
• Chip prices are estimates, not list prices. Epoch AI's US$25,000 per H100 is derived from revenue divided by units. Jensen Huang's US$30,000 to US$40,000 per Blackwell GPU is an executive statement in an interview. The US$3 million GB200 NVL72 rack is an HSBC analyst estimate. Street-price ranges of US$35,000 to US$40,000 for an H100 come from vendor blogs and should be treated as indicative only.
• Useful life is the largest unresolved number in the industry. Book lives run from two to six years across Microsoft, Amazon, Meta, Alphabet and CoreWeave, and moved in opposite directions at Amazon and Meta in the same year. Michael Burry's US$176 billion of understated depreciation and his 2 to 3 year economic life are estimates by an investor with a position, not audited figures. The counter-argument, that older parts cascade to cheaper work, is plausible but this research found no clean primary disclosure of A100 utilisation or rental economics from Amazon or Google to settle it.
• The GB200 NVL72 rack mass of about 1,360 kg could not be verified. NVIDIA's own DGX GB200 documentation gives the power but no mass, and the Supermicro datasheet gives dimensions and power but no weight. Any launch-mass calculation built on 1,360 kg per rack should be discarded. This brief uses the Starlink V2 mini figure of 20.5 kg per kW instead.
• SpaceX's 1,400 W per square metre radiator claim is unverified and pivotal. It is roughly 10 times the planform figure of NASA's operating ISS radiators and about 4 times NASA's Photovoltaic Radiator. If it is right, orbital compute is far more feasible than the skeptics allow. If it is wrong, the mass penalty is decisive. No independent verification exists.
• Solar capacity factors here are arithmetic, not citations. The roughly 19% mid-latitude terrestrial figure in this brief is derived from about 1,700 kWh per square metre per year divided by 8,766 hours at standard test conditions. The commonly quoted 15 to 25% range could not be pinned to a single primary source in this research.
• The Duke University "Rethinking Load Growth" curtailment headroom figure (76 to 98 GW in the United States at 0.25 to 1% annual curtailment) is widely cited in this debate but was not independently verified in this research. Treat it as unconfirmed.

Open questions students could investigate

1. Build the bottom-up model. Take a 100 MW Canadian AI campus and compute, from vendor datasheets only, the number of GB300 NVL72 racks, the accelerator count, the chip capital, the annual electricity bill in Manitoba versus Ontario, and the depreciation charge at 3, 5 and 6 year lives. Which assumption moves the internal rate of return most?
2. Value the useful-life assumption. Amazon shortened server lives to five years and lost US$1.0 billion of FY2025 net income; Meta lengthened them to 5.5 years and gained US$2.59 billion. Model both policies on the same 100 MW asset. At what assumed life does a Canadian project stop clearing its cost of capital, and who bears that risk in a build-to-suit lease?
3. Test the 400-accelerators-per-megawatt rule against every publicly disclosed Canadian project. Where it fails, is the discrepancy PUE, chip generation, or a company inflating a GPU count for a press release?
4. Quantify the Canadian cold-climate advantage now that liquid cooling is mandatory. Model PUE 1.45 versus 1.20 on a fixed 80 MW IT load and price the difference at Winnipeg, Montreal, Toronto and Halifax rates. Then ask how much of that advantage survives when direct-to-chip cooling makes outside air temperature nearly irrelevant.
5. Price the waste heat. A 26 MW Kamloops hall rejects roughly 26 MW of low-grade heat. What is that worth to a district energy system, a greenhouse or an aquaculture operation in a Canadian winter, and what contract structure would make a data centre operator care?
6. Model the tariff wedge. Given the 25% US Section 232 semiconductor measure, its exemption for goods used in US data centres, and the absence of a CUSMA carve-out, estimate the landed cost difference for an identical 100 MW rack order in Ontario versus Ohio. What would Canada have to offer to close it?
7. Interrogate the orbital case with Google's own numbers. At US$200 per kg and 20.5 kg per kW, orbital compute is roughly cost-comparable to terrestrial energy on a per-kW basis. Rebuild that comparison including the satellite, the radiator at NASA's measured 96 kg per kW, the chips and the ground segment. What launch price does the full system need?
8. Settle the inference-versus-training split. Using van Berkel's bisection bandwidth figures and Meta's measured 41% model FLOPs utilisation, estimate what an orbital training run would actually cost per token compared with a terrestrial one. Then do the same for inference, where the satellite can be self-contained.
9. Follow the copper, not the silicon. NVIDIA says a 1 MW rack needs up to 200 kg of copper busbar and a gigawatt data centre up to 200 tonnes of copper. Map the Canadian content of an AI data centre build: copper, transformers, switchgear, cooling distribution units, construction trades. Is the industrial opportunity in the building rather than the chips?
10. Ask what "sovereign" means physically. Canada funded Cohere's compute at a facility operated by CoreWeave, an American company, and TELUS's cluster runs NVIDIA silicon under US export control. Define a measurable test of compute sovereignty that a Canadian government could actually apply at procurement.

Sources

Narrative

Canada's AI data-centre boom is really two stories that are easy to confuse. The first is an announcement wave, concentrated in Alberta, that adds up to more electricity than the province has ever consumed at once. The second is a much shorter list of projects that have secured land, power, a tenant and a construction contract. The distance between the two lists is where the business questions live.

Start with what is real. Meta's Sturgeon County campus is the anchor. Announced and broken ground on July 8, 2026, it is a 1 GW, AI-optimized facility with more than C$13 billion of private capital, about 3,000 construction jobs and 300-plus permanent roles. Its power plan shows how Alberta now works: a 970 MW grid position won under AESO's one-time Phase 1 process, a 250 MW supply contract with Capital Power from the second half of 2028, and a dedicated 932 MW combined-cycle gas plant, Greenlight, owned by Pembina, Morgan Stanley Infrastructure Partners and Kineticor, that took final investment decision on July 2, 2026 for service in the second half of 2030. Meta expects to be running two years before its own plant is finished, so the grid carries the campus in between. The second Alberta project with a grid contract is TransAlta's Keephills site, where CPP Investments and Brookfield signed a memorandum of understanding on February 27, 2026 for an initial 230 MW that could grow to 1 GW.

Outside Alberta, the projects under construction are smaller and mostly hydro-fed. Bell's 300 MW campus outside Regina (C$1.7 billion, tenants Cerebras and CoreWeave, first data halls in the first half of 2027) is the largest purpose-built AI site being built in Canada today and the only one where a Crown utility, SaskPower, agreed to supply the full load. Microsoft is spending C$7.5 billion over two years on its Quebec City and Toronto regions, QScale is adding 60 MW in Lévis for C$700 million under new owner Goldman Sachs, Vantage is finishing a 32 MW building in Quebec City, Keel (formerly Bitfarms) is consolidating 96 MW of bitcoin capacity in Sherbrooke into an AI campus, eStruxture's 90 MW CAL-3 near Calgary opens in fall 2026, and TELUS opened a sold-out Rimouski facility in September 2025 with 85 MW secured from BC Hydro for a Kamloops and Vancouver cluster that scales past 150 MW by 2032.

Now the announcement wave. AESO counted 29 projects and 16 GW of requests in June 2025, "over 20,000 MW" by September 2025, and the province reported 19,565 MW on July 30, 2026, against a record peak of about 12,800 MW. Alberta Data Centre Watch lists 30 active projects with a published maximum of 13.1 GW. The biggest names, Wonder Valley (7.5 GW, C$70 billion), Beacon's six campuses (roughly 2,000 MW), Malachite One (1 GW), Travers (up to 2,000 MW), Wild Rose (1.3 GW) and Newell (1.2 GW), are applications, not projects. Wonder Valley illustrates the gap: 20 months after its announcement it has a land deal, a water certificate held by the municipality, an exemption from provincial environmental assessment, no announced tenant, no gas-plant partner, and a construction start that has slipped from 2026 to about late 2028.

Alberta's response to the queue is the most important policy development in the country. AESO's interim limit of 1,200 MW through 2028 was allocated pro rata among 15 qualified projects holding 4,800 MW; after a trading window and withdrawals, two projects took everything. The province then passed Bill 8 and Bill 12 (December 11, 2025) and filed the Data Centre Regulation on June 9, 2026. The model is "bring your own power": a data centre that pairs its demand with new dedicated generation or storage goes to the front of the connection queue, may use up to three years of bridging grid supply inside an initial 1,600 MW pool, and pays a levy of 1 to 2 per cent of the value of its computing equipment, lower if it draws less from the grid. Developers, not existing customers, pay for transmission upgrades. In practice this means every large Alberta project is also a gas-plant project, which pulls in a second set of constraints: turbine delivery slots that now stretch beyond five years, capital costs that have doubled in two years, Alberta Utilities Commission siting approval, water licences that rank behind existing users, and carbon exposure.

The other provinces are rationing. British Columbia replaced first-come, first-served with a competitive call, capped at 400 MW per two years and 145 MW per project, and permanently banned crypto connections. Quebec requires ministerial authorization for any load of 5 MW or more, evaluates projects comparatively, and is doubling the data-centre rate to about 13 cents per kWh from November 1, 2026. Ontario now requires approval under section 28.1 of the Electricity Act, and its draft Playbook makes data centres pay full grid costs, screens them for jobs, sovereignty and community benefit, and creates a higher Class C rate. Saskatchewan will only consider Canadian-owned projects that supply their own power. Manitoba passed a law letting Manitoba Hydro refuse large loads and its premier walked away from a gas-fired hyperscale proposal. Nova Scotia has no active file. New Brunswick is consulting on a framework while a 390 MW gas-and-grid project at Lorneville sits in environmental assessment. Ottawa's 100 MW-plus sovereign data-centre call closed February 15, 2026 and had named no proponents by September 2, 2026, although TELUS says it is advancing work under it.

Which roadblock binds where? In Alberta it is no longer policy but execution: siting and social licence (Rocky View, Olds, Sturgeon County), generation lead times, water, and above all the absence of tenants and financing for speculative land. In British Columbia and Quebec the binding constraint is administrative allocation of scarce hydro, and the price of it. In Ontario it is process time and cost. In Saskatchewan and Manitoba it is Crown-utility capacity and politics. In New Brunswick it is emissions and environmental assessment. Nationally, the financing market has turned selective: lenders want investment-grade tenants, and the tenants (Meta, Microsoft, CoreWeave, Cerebras, Cohere) are few.

For students, the useful frame is a funnel. Roughly 20 GW is requested in Alberta; 4.8 GW qualified; 1.2 GW is contracted; about 1.3 GW (Meta plus Keephills Phase I) is credibly funded; and, across all of Canada, less than 1 GW of new AI-grade capacity is actually under construction in 2026. Watch the events that move projects down the funnel: BC Hydro's call results (fall 2026), Ontario's final Playbook (fall 2026), the Régie's decision on Quebec's tariff, AESO's first Phase 2 bring-your-own-generation intake, federal MOUs, a gas-plant partner for Wonder Valley, the Rocky View moratorium, Synapse's appeal, and first power at Meta's campus in 2028.