Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.
Was this forwarded to you? Subscribe below.
I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.
Bell and Saskatchewan signed a deal on September 14 to grow Bell’s Regina-area AI data centre to 1.2 gigawatts, with an investment Bell says could exceed $50 billion. It came on the first day of the Canada Investment Summit in Toronto, where RBC, CIBC, BMO and Teachers’ also made new investment announcements. Five startups raised, including Calgary’s Ultimarii, where Mistral invested. Intrepid Growth Partners closed a US$525M AI growth fund. Cohere shipped a translation model, and its CEO called the AI slowdown pitch a cartel. Ottawa launched a national AI literacy program out of Edmonton, and a Mila team built a benchmark that makes frontier models catch falling knives. Let’s get into it.
💰 Deals & Milestones
Micrologic (Québec City) raised $45M for its sovereign cloud business: $25M from Fonds de solidarité FTQ and $20M from Investissement Québec.
The company hosts government and enterprise data and AI workloads on Canadian infrastructure. CEO Stéphane Garneau told La Presse the financing includes clauses limiting a foreign takeover.
Ultimarii (Calgary) raised more than $13M in Series A and related financing led by BDC Capital’s Industrial Innovation Venture Fund and yours truly. (Disclosure: Now a Mistral portco.)
Ultimarii builds AI for regulatory and permitting work on energy, utility and mining projects. Founders Doug Schweitzer, Alberta’s former justice and innovation minister, and Josh Malate, who co-founded Athennian, started the company in January 2024.
Onix (Montréal) raised US$5M in a pre-seed convertible note led by Los Angeles-based AlphaEdison.
The consumer app offers health and wellness advice from AI versions of practitioners. Garage Capital, Ride Home Fund and Real Ventures co-founder JS Cournoyer participated.
Galaxia (Halifax) closed an oversubscribed $4.5M round co-led by Amiral Ventures and Invest Nova Scotia.
Galaxia builds satellites with onboard compute and AI so data gets processed in orbit rather than after downlink, for defence and commercial customers.
Vivid Machines (Toronto) received up to $1.69M from Agriculture Canada’s AgriScience Program to adapt its computer-vision crop camera from apple orchards to vineyards.
The camera mounts on tractors and sprayers and scans each vine for disease, fruit quality and yield.
Intrepid Growth Partners (Toronto and London) closed its first fund at US$525M to back growth-stage AI companies in Canada, the UK, the US and Europe.
The founders are Mark Machin, former CEO of CPPIB, Mark Shulgan, former head of growth equity at OMERS, and Ajay Agrawal, co-founder of Creative Destruction Lab. More than 80 LPs signed on, including Temasek, EDC, BDC, CIBC and Scotiabank, and 45% of the capital is Canadian.
The fund writes cheques up to US$50M and has made nine investments, including StackAdapt, Beacon and Blue J. (Disclosure: Blue J is a Mistral Venture Partners portfolio company.)
Most Canadian Series C money still comes from US funds, so a Canadian-run growth fund of this size is really, really welcome.
Signal 1 (Toronto) signed NL Health Services, Newfoundland and Labrador’s provincial health authority, for its AI management system.
Artificial Agency (Edmonton) landed its first commercial game: a custom AI model that drives non-player characters in a single-player mode of Owlchemy Labs’ VR title Dimensional Double Shift, targeted for 2027.
Versos AI (Saint John) launched an agentic system that lets AI labs describe the video footage they need in plain language and assembles a licensed training dataset from it, demoed at IBC2026 in Amsterdam.
Cortico (Vancouver) released MedSafe-Dx, a free, open benchmark that tests whether language models can safely support clinical decisions rather than pass medical exams.
Across 11 frontier models, the safest, GPT-5.2, passed 97.6% of cases but escalated 71% of routine ones. The weakest missed 26 of 156 urgent cases.
🏢 Large Companies
Bell and the Government of Saskatchewan signed a non-binding MOU on September 14 to add up to 900 megawatts to Bell’s AI data centre near Regina, taking it to 1.2 gigawatts. Bell says total capital investment could exceed $50 billion.
Premier Scott Moe and CEO Mirko Bibic announced it with Prime Minister Carney on day one of the Canada Investment Summit. Bell will move its AI Fabric headquarters to the Regina area. The province approved it under its August Data Centre Framework, which requires Canadian ownership, data sovereignty and bring-your-own power.
Per The Logic, Bell will start new buildings only once customers sign, and the $50B counts construction, power generation and the compute hardware. The existing 300 MW site, announced in March, is pre-sold to Cerebras and CoreWeave.
Ottawa’s Canada Investment Summit runs September 13 to 15 in Toronto with roughly 100 institutional investors, and the pledges started before it opened.
RBC launched the RBCx Growth Fund I, a US$1B fund for direct equity in growth-stage Canadian tech, with up to US$300M of RBC’s own capital and applied AI listed first among its sectors. CIBC committed $2B over five years to defence and dual-use SMBs, BMO said it would mobilize up to $70B over a decade, and Teachers’ committed $10B more in Canada by the end of 2027.
A leaked 167-project pitchbook includes a $14.5B data centre near Medicine Hat and a chance to invest in Toronto’s Xanadu.
Cohere (Toronto) released North Small Translate, its first translation model in the North family, on September 10. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)
The mixture-of-experts model has 218B parameters with 25B active, covers 50-plus languages, and ships open weights under a non-commercial licence. Cohere reports a WMT26 all-languages score of 83.6 against 81.4 for DeepL, 81.6 for Qwen 3.5 397B and 68.2 for Google Translate, judged by GPT-5.6.
CEO Aidan Gomez published a post on September 13 calling proposals from Anthropic, OpenAI and xAI for a coordinated AI slowdown a “state-sanctioned cartel.” At the Canadian Global Growth Forum on September 14, he said Canada could become “a superpower in data centres” given its climate, land and clean power.
Vention (Montréal) opened a physical AI lab on September 9 focused on robotic manipulation for manufacturing, led by Jimmy Li and advised by Joelle Pineau, Cohere’s chief AI officer.
The lab has eight researchers and more than 16 open roles, and post-trains robotics models on production data from Vention’s customers. Vention says physical-AI revenue grew 400% over the past year.
Canada’s big banks put fresh numbers on AI at the Scotiabank Financials Summit on September 9.
Scotiabank CEO Scott Thomson said AI saved roughly 24,000 days of work in about four and a half months. TD CEO Raymond Chun said mortgage pre-processing fell from about 15 hours to three minutes.
Export Development Canada told The Logic it has put more than $700M into the AI sector and is positioning itself as an AI investor. EDC is also an LP in Intrepid’s new fund.
Meta made its case for the $13B Sturgeon County data centre at the Alberta Industrial Heartland Association conference in Edmonton on September 10, and disputed a study that said the project could raise Alberta power prices.
🔬 Research
Cohere Labs trained a 3.35B-parameter model, Tiny Aya L2-Thinker, to reason in the language the user writes in, instead of switching to English internally the way most reasoning models do.
Main finding: the model reasons in-language more than 93% of the time across 60 languages and six benchmarks, with little accuracy cost. The recipe is a large English reasoning dataset, translated reasoning traces for some languages, and a small share of ordinary multilingual data. The behaviour transfers to languages that never saw reasoning examples.
So what: reasoning ability and reasoning language are separable, so multilingual reasoning does not have to be built one language at a time. Cohere released the weights and the data.
UdeM/Mila (Bang Liu’s group) built ReactHuman, a benchmark where a multimodal model acts as the brain of a simulated humanoid that has to react to sudden household hazards: a slipping plate, a falling knife, a foam anvil that looks heavy and a steel apple that does not.
Main finding: across seven models, including Claude Opus 4.8, GPT-5.5 and Gemini 2.5 Flash, roughly one hazard in three is mishandled. Models trust appearance over motion and miss the catch point by a metre or more even when they pick the right action. None of this improves with model size.
So what: the paper measures whether a model can turn physics knowledge into a safe action in time, which is the job a household robot’s brain has to do.
That makes two Canadian benchmarks in two weeks showing the models behind the humanoid demos are not there yet.
Yoshua Bengio (Mila, LawZero) and CSIRO’s Qinghua Lu published “AI Safety: Not Optional, Not Later,” an invited column for IEEE Software’s January 2027 issue.
Main finding: an architecture rather than a result. It lays out a safety-by-design stack: Scientist AI-style supervision at the model level, controls on the scaffolds around it, independent verification and monitoring.
So what: it is the most detailed public sketch of how LawZero expects Scientist AI to sit inside a deployed system.
Nicolas Papernot (U of T, Vector, Schwartz Reisman Institute) made the case for independent pre-release testing of frontier models, in a U of T interview and an SRI talk on September 10, building on his lab’s June demonstration of an AI worm assembled from free models.
Main finding: the June worm adapts its attack to each device it reaches, so patching a single vulnerability no longer stops its spread. Papernot told CBC AI is exposing how weak existing cybersecurity already is.
So what: he wants universities to test new models for security risks before release.
🏛️ Policy
Ottawa launched Canada’s National AI Literacy Initiative on September 9 in Edmonton, delivered by Amii under the AI for All strategy.
Three free streams: a three-hour course for up to one million post-secondary students, classroom materials for 50,000-plus K-12 teachers, and a general course for Canadians rolling out through community partners later this year. The first modules open September 21. The release does not state a dollar figure.
Alberta’s data centre town hall tour ended in Grande Prairie on September 11 with the same hostility that met Lacombe and Redwater.
Speakers got 45 seconds each and about 20 people were still in line when the session closed. Minister Nate Glubish said Kevin O’Leary’s Wonder Valley project has not been approved. It faces a court challenge from Sturgeon Lake Cree Nation.
📰 In brief
Mila signed a 12-month research partnership with UK startup Worldmodeldata to test whether world models follow a data scaling law, using video-game-generated training data. Results are due next year.
Amber Simpson (Queen’s, Amii Fellow, Canada CIFAR AI Chair) was elected to the Royal Society of Canada’s College of New Scholars.
Waterloo researchers explained “Claudish,” the terse, imperative style developers pick up from talking to coding agents.
ALL IN 2026 runs September 16 to 17 in Montréal; I will be there, so come say hello. NeurIPS author notifications land September 24. Ontario’s Data Centre Playbook consultation closed September 12.
Thank you for reading! Subscribe, and please share any feedback 🙏🏽
— Raif