# Who is actually building Canada's AI compute?

Date: 2026-10-22 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/canada-compute-who-builds/

**Answer:** Canada's AI compute is being built mainly by telecoms such as Bell and Telus, American and Canadian specialist cloud firms, and foreign hyperscalers, with Ottawa adding money, a planned public supercomputer and non-binding agreements. The chips in nearly all these projects still come from American suppliers, and many of the largest numbers are plans, not running machines.

A press release can say gigawatts, but only a finished building can train a model. So the useful way to read Canada's AI buildout is to ask three questions about each project: is it running, under construction or just planned; who operates it; and how firm are the numbers?

Canada's *AI for All* strategy cites an estimate that Canadian commercial players will need 5.5 gigawatts of AI compute by 2030, with much of it delivered by foreign giant cloud companies, known as hyperscalers. Against that, it reports proposed partnerships for 850 megawatts of sovereign capacity by 2030, scalable to 2.3 gigawatts. Here is who is behind the main projects.

## The main projects

| Project | Who | Size announced | Status | Worth noting |
|---|---|---|---|---|
| Rimouski, Que. | Telus | Not stated in megawatts | Open since September 2025, sold out | Nvidia GPUs, HPE infrastructure |
| B.C. cluster: Kamloops and two Vancouver sites | Telus, with Westbank | 85 MW initially, over 150 MW and 60,000 GPUs by 2032 | Kamloops due this year; M3 Vancouver end of 2026; 150 West Georgia 2029 | In talks under a federal initiative |
| Cambridge, Ont. | CoreWeave (U.S.), Cohere as anchor tenant | $725 million project | Federal contribution up to $240 million | CoreWeave builds and operates it |
| Merritt, B.C. expansion | Bell site, Buzz HPC operates, Hypertec hardware | Not stated | Announced June 2026 | Runs Cohere tools; Nvidia chips |
| B.C. supercluster | Bell AI Fabric | Upwards of 500 MW across six sites | Announced May 2025; the plan names Kamloops and Merritt | Hydroelectric power |
| Sherwood, near Regina | Bell | 300 MW, $1.7 billion | Under construction; first two halls due in the first half of 2027 | Anchor tenants CoreWeave and Cerebras |
| Saskatchewan expansion | Bell | Up to 900 MW more, path to 1.2 GW, over $50 billion | Approved by the province; would need its own power | Dedicated natural gas plant |
| Public supercomputer | Ottawa | Target by 2031 | Applications opened April 2026 | Up to $1 billion envelope for public compute |

A **neocloud** such as CoreWeave or Buzz HPC is a newer cloud company focused on renting AI chips, as opposed to a general cloud giant. An **anchor tenant** signs on for much of a site's capacity early, which helps the builder finance it.

## Reading the numbers with care

Three cautions apply to this table.

**Announced capacity is not capacity in use.** Sherwood's first halls start next year. The $50 billion figure from Saskatchewan's government is an estimate for a site that needs up to 900 megawatts of new power. Ottawa's large-project route is also tentative: Budget 2025 said it would work with the best proposals through non-binding agreements called memoranda of understanding, and the Telus talks are described as work on "the details". So 150 megawatts by 2032 is a company target.

**Sovereign does not always mean a Canadian tenant.** Bell's Regina campus is widely described as sovereign infrastructure, and it is owned by a Canadian company. But its anchor tenants are American firms, CoreWeave and Cerebras. Whether Canadian users get first call on that power is a contract matter that public documents do not settle.

**One word, several structures.** Cohere's two sites show the range. In Cambridge the federal statement says Cohere "will act as an anchor tenant" of a facility "built and operated by CoreWeave". In Merritt, The Logic reports, Canadian firms control facilities, network, infrastructure, software and models, and the chips are from Nvidia. Both are described as sovereign compute. Critics such as Joshua van Es argue only the second comes close, and Cohere's commitment to at least $1.7 billion of further Canadian infrastructure by 2030 will show which model it prefers.

Other Canadian names fill in the supply chain. ThinkOn owns eleven data centres in the country, according to *The Walrus*, and belongs to Bell's Canadian Sovereign AI Alliance along with Celestica and Coveo. Montréal's Hypertec supplies hardware.

## A worked example: what a megawatt means for a buyer

Companies quote megawatts, but you may think in chips. Telus gives a useful ratio: over 60,000 GPUs and 150 megawatts at full scale.

```text
150,000,000 watts / 60,000 GPUs = 2,500 watts per GPU
```

That 2,500 watts covers the chip plus its share of the servers, networking and cooling, assuming Telus's two figures describe the same end state. The same rough rule suggests a 300-megawatt site could hold on the order of 120,000 such chips. Treat this as a sense of scale only, since real density depends on the chip generation and the cooling design.

## Who objects

Local opposition is part of this story. In Saskatchewan, critics worry about the effects on the environment and the grid and say the project moved faster than consultation, reports the *Journal of Commerce*. The Opposition NDP has called for a moratorium on large centres, while Crown Investments Minister Jeremy Harrison says elected representatives reflect broad support. The province's framework says environmental assessment may apply, and the expansion's gas plant makes the "clean compute" pitch harder there than in hydro-powered British Columbia or Quebec.

Federal documents raise such concerns too. Ottawa's own climate assessment of the Cambridge project says it may indirectly add to greenhouse gas emissions, depending on the third-party data centres' energy sources.

> **Try it:** Pick the project closest to you. Find the most recent primary document, such as a company release or municipal filing, and write down three things: megawatts running today, megawatts planned, and who operates it. If you cannot find the first number, assume it is low.

## Can a small Canadian company actually use this capacity?

Often yes, though not usually by renting a whole site. Telus and others sell access to shared chips, and Telus says its platform supports training, fine-tuning and running models. The federal AI Compute Access Fund, with up to $300 million, helps eligible businesses pay for compute, and *AI for All* promises another $700 million for affordable sovereign compute for small and medium-sized firms.

Three checks save time. Confirm availability first, since Rimouski's capacity sold out. Compare the price per GPU-hour with what you pay now, because sovereign compute may cost more. And ask who the operator is and what law governs the contract, so you know what you are paying extra for. If your data has no legal or contractual reason to stay in Canada, a foreign cloud may be the sensible choice, and that decision is yours to make with open eyes.

## Sources

- Single Largest Private Sector Capital Investment in History of Canada Through Bell AI's Saskatchewan Expansion: https://www.saskatchewan.ca/government/news-and-media/2026/september/14/single-largest-private-sector-capital-investment-in-history-of-canada-through-bell-ais-saskatchewan
- Feds and TELUS advance work to build sovereign AI infrastructure: https://www.renewcanada.net/feds-and-telus-advance-work-to-build-sovereign-ai-infrastructure/
- Strategic Response Fund: Cohere compute expansion project public statement: https://ised-isde.canada.ca/site/ised/en/strategic-response-fund-cohere-compute-expansion-project-public-statement

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/canada-compute-who-builds/, no commercial use without permission, and keep the same licence.