Toronto Is 315 MW: What CBRE’s Data Changed in Our Data-Centre Research

When we published Canada’s Data Centre Race, the weakest part was not the argument. It was the Canadian capacity numbers underneath it.

The U.S. figures were easy. American metro capacity is tracked market by market by commercial brokers who count buildings for a living. The Canadian figures were assembled the hard way: operator press releases, trade press, a chart read off a broker’s PDF, and in Montreal’s case an aggregate we built ourselves by adding up the campuses we knew about. We flagged them as estimates, which is the right thing to do and also not much of a defence.

Then CBRE published its H2 2025 numbers, and the problem partly solved itself.

The table that changed the comparison

In March 2026, CBRE published its top ten North American data-centre markets by inventory. Toronto and Montreal are in it. Same broker, same methodology, same table as Northern Virginia and Dallas.

At the end of 2025:

• Northern Virginia: 4,039.6 MW of inventory, 21.5 MW available, 0.5% vacancy
• Atlanta: 1,459.2 MW  ·  Dallas-Fort Worth: 1,067.3 MW
• Chicago: 904.6 MW  ·  Phoenix: 807.3 MW
• Toronto: 315.0 MW, 29.6 MW available
• Montreal: 229.5 MW, 6.4 MW available

Toronto and Montreal together come to 544.5 MW. That is about 13 percent of one American metro — and Northern Virginia grew by more than a gigawatt during 2025 alone, roughly twice everything Canada’s two real markets have built in total.

We had been making that comparison all along. We just could not previously make it without mixing measurement systems, which is exactly the sort of thing that lets a wrong number survive.

What we got wrong

Six corrections came out of the review. One of them is an actual error and the rest are precision problems, but they are worth listing individually, because the pattern is more interesting than any single item.

Calgary.We wrote that real operating capacity sits in three metros: Toronto, Montreal, and Calgary “at roughly 125 MW.” But 90 of those megawatts are eStruxture’s CAL-3, which does not open until the second half of 2026. The operating footprint today is about 35 MW. Our own dataset said so, and a different chapter of the same series stated it correctly. This is the error I mind most, because counting an unbuilt building as operating capacity is precisely the thing that chapter was written to warn against.

Northern Virginia.We quoted “roughly 2,930 to 4,040 MW,” which reads as an uncertainty band. It was never a range. Those are two annual readings, H2 2024 and H2 2025, describing a market that grew 37 percent in a year. Presenting growth as uncertainty hid the more interesting fact.

Chicago. We said Dallas, Atlanta and Chicago had each cleared roughly 1,000 MW of commissioned capacity. Chicago was at 904.6 MW. Only three North American markets had passed a gigawatt.

Montreal.Our ~200 MW estimate quietly included Vantage’s Quebec City campus, which is a separate market about 250 km away. The correct tracked figure, Montreal only, is 229.5 MW.

The project count. Our prose said we tracked 18 projects. The dataset had grown to 20 rows, one of which was the same TELUS B.C. cluster entered twice at two different capacity horizons. The data moved and the writing did not follow it.

The 337 MW problem. This is the one that matters most, and it gets its own section.

Five different numbers, all called “capacity”

Our map chapter had these two sentences next to each other: operating capacity is concentrated in three metros at roughly 312, 200 and 125 MW; and a federal pitch deck put Canada’s current AI data-centre capacity at about 337 MW.

Both statements were sourced. Both were accurate. And placed side by side with nothing between them, they invite a reader to assume the 337 MW is the national roll-up of the metro figures. It is not. On CBRE’s definition, Toronto and Montreal alone exceed 337 MW — which would be incoherent if the two measured the same thing.

They do not. Sorting this out is most of what this update was actually about:

• Commissioned colocation inventory — multi-tenant megawatts a broker counts building by building. Excludes owner-occupied hyperscaler campuses. This is CBRE’s 315.0 MW for Toronto.
• AI-designated capacity — what governments mean when they talk about AI compute. Includes self-built facilities that never appear in a colocation table. This is the federal 337 MW.
• Critical IT load — what the servers can actually draw, usually lower than the headline facility number.
• Secured grid service — megawatts a utility has agreed to deliver, which may be years from energizing anything. TELUS’s 85 MW from BC Hydro is this.
• Announced campus capacity — the press-release number. Wonder Valley’s 7.5 GW is this, and Beacon Indus’s 1,494 MW is not even that: it is the size of the proposed gas plant, not the computing load.

Every chapter that carries more than one of these now says which is which. The dataset has phased columns — operating, preleased, under construction, secured, ultimate — so a project like QScale can record a delivered 14 MW phase against a reserved 142 MW campus instead of one flat, misleading number.

What CBRE knows that nobody else publishes

The corrections are the boring half. The genuinely new material is what a broker sees that a utility filing does not: who is leasing, at what price, and how little is left.

Toronto has almost nothing available.Immediately available, built-out colocation space in the 3-to-6 MW range exists at only three locations in the entire market. CBRE describes it as a bottleneck creating price pressure. Meanwhile, greenfield projects of 50 to 400 MW are being planned for 2027 and 2028, with developers paying for power studies and application fees at both Toronto Hydro and Alectra — Ontario’s quieter version of Alberta’s interconnection queue.

The largest AI lease in Canada is in Cambridge, Ontario.CoreWeave secured roughly 52 MW of preleasing in January 2025, with initial infrastructure operational by that June — the largest wholesale colocation deal in the GTA and the country. The anchor customer is Cohere, and up to C$240 million of federal money is attached under the Sovereign AI Compute Strategy. It was missing from our project dataset entirely, which is why our map read Ontario as “cloud, not campus.” It is now in, and it is the sharpest illustration of the argument in our sovereignty chapter: Canadian soil, Canadian grid, Canadian anchor tenant, Canadian public money, and an American operator within reach of the CLOUD Act.

Quebec is delivering in phases, not campuses.QScale delivered 14 MW at Lévis in under six months, aimed at AI tenants. Enovum expanded at two sites, including deals with Cerebras. Nordik and Accelsius launched a facility built around direct-to-chip liquid cooling. What they have in common is that they already held Hydro-Québec allocations — with new power procurement restricted, essentially no large greenfield project broke ground, and blocks larger than 1 MW stayed rare.

Density is repricing everything. AI deployments in Toronto are arriving at 60 to 132 kW per rack, forcing liquid cooling and pushing build-out costs above C$13 million per project. Continent-wide, the old volume discount has inverted: large deployments used to get a price break, and now AI tenants pay premiums for contiguous power. Land with a grid connection is being bid up over ordinary industrial land in Mississauga, and past $8 million per acre in the hottest American markets.

And the American market is hitting the same wall.U.S. primary-market supply grew 36 percent in 2025 to 9,432 MW, absorption set a record at 2,497.6 MW, and vacancy fell to 1.4 percent — while capacity under construction fell for the first time since 2020, on permitting, zoning and power-procurement delays. Grid capacity is largely booked through 2030 in most markets, and interconnection can run 24, 36, or 48-plus months.

That last point forced a change of emphasis rather than a correction. We had framed behind-the-meter gas as something Alberta developers invented to route around AESO’s cap. It is now routine across North America. What remains distinctly Canadian is not the workaround — it is the consequence: spending a genuinely clean grid’s reputation to make the workaround necessary.

Where we don’t follow CBRE

CBRE sells data-centre services. That does not make its measurements wrong — leasing and vacancy data is exactly the thing a brokerage knows better than anyone, and there is no public substitute for it. But it does mean the reading has to be sorted by type, so our source records now tag every CBRE citation as measured research, forecast, or commentary.

Measured research we use as evidence. Forecasts we date and attribute. Commentary — including CBRE’s view that Canada’s grid “could reach a breaking point,” and its published position that the benefits of data-centre growth outweigh the drawbacks — is named as CBRE’s analysis. Our jobs chapter now states that position at its strongest, with CBRE’s real finding that data centres will supply nearly half of Loudoun County’s property tax revenue in 2026, and then answers it on jobs per megawatt.

We also do not cite CBRE for numbers CBRE is repeating. Its Canadian article carries a market valuation from a third-party research firm and a global electricity figure from the IEA. Those belong to their original publishers, and the second one is four years stale. Nothing in the series was built on either.

And where a utility, regulator, government filing, or operator disclosure already carries a claim, that source stays. AESO still carries the queue. Hydro-Québec still carries the rate. Pembina still carries the gas plant. CBRE corroborates; it does not displace.

What we still don’t know

Six questions came out of this that we deliberately did not answer in prose, because answering them would have meant guessing:

• Who owns the Cambridge building, and whether CoreWeave’s 52 MW of preleasing and CPP Investments’ separately financed 54 MW expansion describe the same megawatts.
• QScale’s cumulative operating capacity at Lévis. We know one 14 MW phase landed. We do not know the total, and 142 MW is a reservation, not a running campus.
• eStruxture’s current load at CAL-1 and CAL-2, behind our ~35 MW Calgary figure.
• Which AESO queue snapshot is current: 21,085 MW, or the earlier 29 projects requesting more than 16 GW.
• Whether Microsoft’s 48 MW GTA campus sits inside or outside CBRE’s 315.0 MW Toronto inventory, which determines whether Toronto’s real number is 315 MW or meaningfully higher.
• CBRE’s claim that nearly half of Canada’s data centres are hyperscalers, which is unsupported in the source and hard to square with a 300-plus facility count. We did not use it.

The part worth keeping

None of the corrections changed a conclusion. Power interconnection is still the binding constraint. The clean grid is still being undercut by gas. Foreign operators still control most Canadian capacity, including the flagship sovereign-AI deployment. The same megawatts would still employ far more people almost anywhere else.

What changed is the precision of the evidence, and one embarrassment about an unbuilt building. Both were worth the trip. If a series argues that the whole Canadian data-centre conversation confuses announced capacity with operating capacity, then the least it can do is not make that mistake itself — and say so in public when it does.

The revised chapters carry a dated update note explaining what changed and why. The underlying dataset, including the phased capacity fields and the open questions above, is part of the same public research record.

Frequently asked questions

Sources

CBRE Research and CBRE Data Center Solutions: North America Data Center Trends H2 2025 (February 25, 2026), the accompanying market-inventory release including Toronto and Montreal (March 18, 2026), Global Data Center Trends 2026 (June 17, 2026, reporting Q1 2026), the U.S. Real Estate Market Outlook 2026 data-centre chapter (January 14, 2026, a forecast rather than measurement), the Toronto and Montreal market profiles from North America Data Center Trends H1 2025 (September 8, 2025), and Power Surge: Energizing Canada’s Data Centre Expansion(February 11, 2026, CBRE commentary). Canadian project detail from operator disclosures, the Alberta Major Projects registry, the Impact Assessment Agency of Canada, AESO, Hydro-Québec, BC Hydro, IESO, ISED, CPP Investments, and named reporting by The Globe and Mail, DataCenterDynamics, CBC News, Canada’s National Observer, and The Narwhal. Figures are dated and labelled by measure; announced capacity is never presented as operating.