Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own. I also run the Canadian AI Index, a free database of Canadian AI companies, founders, rounds and investors.

Prime Minister Carney named a National Council on Artificial Intelligence on Friday, with Yoshua Bengio, Sanja Fidler and Michelle Zatlyn among its members. Cohere launched North 2, Clio bought an AI company that builds for judges, and Field Effect is growing fast on a shadow-AI tool. AI agents tried to break into a Library and Archives Canada website, and a Quebec novelist wants to sue Meta over AI training. Also in this issue: deepfake detectors tested on court evidence, and Canada’s bid for special access under a proposed EU AI law. Last week I also launched the Canadian AI Index, more on that below. Let’s get into it.

🗂️ New: the Canadian AI Index

I launched the Canadian AI Index on October 1. It’s a free database of 571 Canadian AI companies, their founders and investors, with 458 funding rounds and C$15.9B in disclosed capital. Toronto has the most companies at 202, followed by Montréal (96), Vancouver (72), Kitchener-Waterloo (35) and Calgary (28). Every entry traces to a primary source, and I update it every Tuesday with this newsletter. If your company is missing or something is wrong, reply to this email.

💰 Deals & Milestones

• Clio (Vancouver) bought Learned Hand, which builds AI tools for judges and courts, on September 30.

  • Terms weren’t disclosed. It’s Clio’s first move into the judiciary.

• Field Effect (Ottawa) grew from about 9,000 to 12,000 customers in under four months on a tool that detects employees’ unsanctioned AI use, CEO Matt Holland told the Globe on October 1.

  • Annualized revenue is near $50M, up from $26M last year. Field Effect plans to raise US$100M in 2027.

• Femtum (Québec City) opened a $2.1M factory on October 1 to build lasers that chipmakers use to make silicon photonic chips for AI data centres.

  • The site is six times its old footprint and follows Femtum’s $16M Series A in March.

• Candidly AI (Calgary) is using AI voice agents to interview residents and employees for municipalities and consultants, with the Town of Bashaw as its first municipal client.

  • Data stays in Canadian data centres on self-hosted models.

• H2 Analytics (Ottawa) got $800K from FedDev Ontario on October 1 toward a $4.2M upgrade of EASE, its AI platform that builds training exercises for defence and security clients.

  • H2 is owned by a Canadian Armed Forces veteran.

🏢 Large Companies

• Cohere update: Cohere had three developments this week. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • North 2 launched October 5. Clients can now run rival models from Anthropic, OpenAI and Google inside the agent platform and cap AI spending by team.

  • Aleph Alpha, Cohere’s German merger partner, released Kolibri on October 3, an open-weight model for German and English public-sector and industrial work. The weights are on Hugging Face under Apache 2.0. The merger awaits regulatory approval.

  • Schwarz Digits, the tech arm of Lidl’s parent and the expected US$600M lead of Cohere’s next round, is exploring compute in Canada, The Logic reported.

• Enerflex (Calgary) won a contract on October 1 to build about 450 MW of off-grid, gas-fired generators that will power a North American developer’s data centre directly.

  • Deliveries run 2027 to 2028. Shares rose about 15%, per the Globe.

• D2L (Kitchener) put its Lumi Assistant into beta on September 29 for administrative and teaching work.

🔬 Research

• A team from U of T, Waterloo, UBC, Western and uOttawa altered or faked 785 of 1,505 court-style photos, such as dashcam stills and surveillance frames, with consumer AI tools.

  • Main finding: the best off-the-shelf detector caught 47% of the fakes and flagged 22% of the real photos. Small local edits were the hardest to catch.

  • So what: detection tools sold for courts and insurance claims fail this test. The Canadian AI Safety Institute funded the work.

• Mila, McGill, ServiceNow Research and Université Laval tested whether AI web agents can infer a user’s preferences from past browsing, across 480 tasks.

  • Main finding: agents told the user’s preferences succeeded 83% of the time. Agents that had to infer them topped out at 16%.

  • So what: personalization is still an open problem for startups selling web agents.

🏛️ Policy

• Prime Minister Carney named a National Council on Artificial Intelligence on October 2 to advise on the AI for All strategy, national champions, sovereign infrastructure and safety. A small team in the Privy Council Office will support it. The Logic and BetaKit also listed CPP Investments CEO John Graham as a member, but the PMO release does not. The members named by the PMO:

  • Ajay Agrawal, professor, U of T; co-founder, Intrepid Growth Partners. Founded Creative Destruction Lab.

  • Yoshua Bengio, scientific director, LawZero; professor, Université de Montréal. Mila founder, Turing Award winner.

  • Sanja Fidler, CEO, Veeda AI; professor, U of T. Former Nvidia VP of AI research.

  • Gillian Hadfield, professor, Johns Hopkins; Vector faculty. Founding director, Schwartz Reisman Institute.

  • Jordan Jacobs, co-founder and managing partner, Radical Ventures. Vector co-founder.

  • Eric Lamarre, board member, WSP Global. Former McKinsey senior partner; co-author of Rewired.

  • Valérie Pisano, president and CEO, Mila. Only national AI institute head on the council.

  • Thomas Pitfield, Senator (Ontario). Carney’s former principal secretary.

  • Jody Thomas, former National Security and Intelligence Advisor. Advised Trudeau and Carney.

  • Louis Vachon, former CEO, National Bank of Canada. Ran National Bank from 2007 to 2021.

  • Michelle Zatlyn, co-founder and president, Cloudflare. Grew up in Prince Albert, Saskatchewan.

  • Patrick Pichette (ex officio), CEO, Digital Transformation Canada. Former Google CFO, Inovia partner.

• Canada will seek “associated country” status under the EU’s proposed Cloud and AI Development Act, according to a leaked draft joint statement reported October 1.

  • It would let Canadian firms bid on sensitive EU contracts the law closes to non-EU providers. The partnership launches at a Montréal summit this month.

• AI agents tried to break into a Library and Archives Canada website on May 28 and June 9, AI research lab Transluce reported September 30.

  • The attempts failed. Transluce said the tactics match agents it previously tied to OpenAI but did not attribute these, per CBC.

• Anne Robillard, author of the Chevaliers d’Émeraude fantasy series, asked Quebec Superior Court on October 5 to authorize a class action against Meta over AI training.

  • She alleges Meta trained on her books and other authors’ works without permission.

• Sherbrooke‘s planned AI data centre is splitting council and residents, Le Devoir reported October 2.

  • The developer is Keel Infrastructure, formerly Bitfarms.

📊 Data

• OntarioMD found 67% of Ontario community clinicians have used an AI tool in practice, and 58% have used an AI scribe.

• IBM found 62% of Canadian employees worry AI is eroding their skills, and critical thinking is the skill they name most.

  • 68% of Canadian CHROs rank checking AI output as the top skill. 29% of Canadian employees rank judgement as important. Globally, only 7% of executives are building up early-career talent.

  • Geneviève Bich, CHRO of Montréal’s METRO, in the report: “If the reinvestment is you’ve saved 10 minutes per person per day and it goes into doing one more email, that’s not a good use of anyone’s time.”

📰 In brief

• Paul Christiano, who developed RLHF at OpenAI, will give the 2026 Hinton Lectures in Toronto, November 9 to 11.

• Calian and Invest Ottawa launched the Northstar Defence Lab at Area X.O on September 29, the first in Calian’s planned national network of defence labs.

• Grande Prairie County permitted an AI data centre.

• Université de Montréal adopted an AI vision statement on September 30.

• Manulife Canada added Sword’s AI physiotherapy to its group benefits.

• At Energy Disruptors in Calgary, a panel said Alberta entered the AI boom with about 1.2 GW of spare generation, and Moment Energy (B.C.) pitched retired EV batteries as data-centre power.

• Rakovina Therapeutics (Vancouver) partnered with Boston’s Celvion to use AI on real-world clinical data for its cancer drug program.

• Innovatree Carbon Group (Kamloops) won Telus’s AI Innovation Award for Forest Guardian, which maps forests tree by tree with LiDAR and machine learning.

• Brookfield‘s Connor Teskey said there’s enough AI infrastructure work to partner with rivals.

• Rewind (Ottawa) CEO Mike Potter said AI is driving demand for its developer backups after H.I.G. took majority control this summer. (Disclosure: Rewind is a Mistral Venture Partners portfolio company.)

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! This is the Canadian AI Newsletter. Tuesdays are the weekly rundown; today we’re launching the Canadian AI Index.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

Canadian AI companies raised $3.33B in private rounds in the first three quarters of 2026. That is more than at the same point in 2024 ($2.83B) or 2025 ($2.10B). Given some of the rumoured rounds that might close in Q4, 2026 could be the biggest year on record for Canadian AI.

Q3 ended yesterday. These numbers come from the Canadian AI Index, which I’m launching today.

Why I built it

I started this newsletter in February. Every week I (and a bunch of my AI agents) go through government newsrooms, wire releases, filings, news sites and research papers looking for Canadian AI news. I kept running into simple questions I couldn’t answer. How many Canadian AI companies raised money last quarter? Who is writing seed cheques in Halifax? How much of the funding comes from government programs?

I tried the usual databases first, Crunchbase especially, and kept hitting the same problems. Small rounds and government grants were often missing, and amounts and dates often didn’t match the original announcement. And their categorization missed a lot of companies. I ended up using it as a list of leads to check, not as a source.

The Index grew out of the newsletter. I was already collecting most of this data every week, so I started keeping it in one place. It covers 570 Canadian AI companies, 734 founders and 698 investors, with funding rounds from 2023 onward (since ChatGPT). Every round links to a primary source. It is free to browse and updates every Tuesday with the newsletter.

What Q3 shows

More rounds, and smaller ones. Canadian AI companies closed 33 private rounds in Q3, the most of any quarter this year. Two-thirds were pre-seed or seed, up from about four in ten in Q2, which is why the median fell. I read that as a good sign for the pipeline of new companies.

A busy quarter for government funding. The Index counted 24 grants and government contributions to private Canadian AI companies in Q3, the most in any quarter so far. Most were small, with a median of $250K. Seventeen came from one program, Canada Economic Development for Quebec Regions’ Regional Artificial Intelligence Initiative. Sixteen of the 24 companies have no private round on record in the Index.

Well beyond Toronto and Montréal. More than four in ten of this year’s private rounds and grants went to companies outside those two metros. Q3 raisers included TetraGen Robotics in Winnipeg, Floqer, Galaxia and SyncraSys in Halifax, Siftmed in St. John’s, Micrologic in Québec City and Ultimarii in Calgary (Disclosure: Ultimarii is a Mistral Venture Partners portfolio company). The Index has companies in all ten provinces and Yukon, across 47 cities.

How it’s built

A company qualifies if AI is the core of its product and it is Canadian-founded with a meaningful team in Canada or headquartered in Canada. Adding an AI feature to a product that already works without it doesn’t count. There is no size floor on companies or rounds. Every date is the announcement date, taken from a primary source.

The sweeps go back to January 2023. ChatGPT launched in November 2022, so the first quarter of 2023 is the first full quarter of the AI era, and anything older belongs to a different market. The Index records Canadian AI after ChatGPT. Call it year one AC. Companies founded earlier are included when they raised, launched, or were acquired within the window.

Raises by publicly listed companies are recorded on their company pages but excluded from the numbers above, so the figures stay comparable to venture data. Mistral VP portfolio companies are included under the same rules as everyone else.

All figures are in Canadian dollars, with US dollar rounds converted at 1.38. They are as of October 1 and will move as late announcements come in.

What I’m asking

Look up a company you know. If something is wrong or missing, every page has a link to tell me, and I’ll fix it in the next weekly update.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

Toronto’s Biossil raised US$153M at a US$1B valuation with OpenAI’s startup fund in the round, and B.C. has sued OpenAI over the Tumbler Ridge shooting. Heidi, Axya and Astrus also raised, Cohere signed deals with TD and Bell Cyber, and Shopify is now the checkout for Meta’s AI agent. Ottawa put $162M into AI work placements and Ontario gave the Vector Institute $30M. Also in this issue: AI hacking agents tested by the National Research Council, imaging AI in six Canadian hospitals, and shopping agents that fall for $4.99. Let’s get into it.

💰 Deals & Milestones

• Biossil (Toronto) raised US$153M at a US$1B valuation with OpenAI’s startup fund in the round, the Globe reported September 28; CEO Anthony Mouchantaf confirmed it.

  • Biossil uses LLMs to find drug candidates that pharma companies shelved after failed trials, then licenses or buys them. It left stealth in April with more than US$68M raised.

• Heidi (Melbourne), the AI medical scribe, raised US$340M on September 22.

  • Blackbird led a US$100M Series C at a US$900M valuation; the other US$240M is revenue-based financing from General Catalyst. Heidi named Toronto its North American headquarters on September 3. Canada is its second-largest market.

• Axya (Montréal) raised a C$17M Series A led by McRock Capital for its AI procurement platform for manufacturers.

  • Yamaha Motor Ventures joined BDC and Real Ventures; Canadian Fintech puts the split at $12M equity and $5M CIBC debt.

• Astrus (Kitchener-Waterloo and Toronto) raised a US$12M Series A led by Caffeinated Capital to automate analog chip layout with reinforcement learning.

  • Analog layout is still done largely by hand. Khosla Ventures and Garage Capital participated.

• Transformer Lab (Kitchener-Waterloo) launched Primus Society on September 24, a simulated city of 10,000 AI researcher agents that pick problems and work on them with little human direction.

• Jobber (Edmonton) launched Jobber Teammate on September 23, an AI agent that briefs home-service business owners each morning and takes on tasks across the business.

• Alebex AI (Vancouver) ranked eighth on Artificial Analysis’ streaming speech-to-text leaderboard, ahead of Google, Microsoft, Nvidia and Amazon, per the company.

• ThoughtWire (Toronto), which sells AI digital-twin software to hospitals, is under creditor protection and running a sale process; bids are due October 20.

🏢 Large Companies

• Cohere update: Cohere signed two deals with Canadian enterprises this week. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • TD committed up to $25M over three years to work with Cohere through Layer 6, its AI lab. A Cohere team will sit with Layer 6 at MaRS.

  • Bell Cyber put a Cohere-built cybersecurity model into production in its security operations centre, trained on Bell’s own data and running on Bell AI Fabric.

• Shopify (Ottawa) is the checkout for Muse, Meta’s AI agent, which can now buy from any Shopify store through Shop Pay.

• Harvey took US$50M from Ontario Teachers’ Venture Growth.

  • Harvey has now raised US$600M in total and says it works with more than 100 Canadian law firms and legal teams.

• Manulife launched a travel-insurance plugin inside ChatGPT, first of its kind anywhere.

🔬 Research

• National Research Council researchers at the Canadian AI Safety Institute ran six AI models as autonomous hackers inside a simulated corporate network.

  • Main finding: in the harder scenario, no model finished the last three of ten attack steps. Most stalled at stealing credentials and moving between machines, and the models’ own checkers overrated their progress.

  • The models tested (Claude Opus 4.5 and Sonnet 4.5, Gemini 3 Pro Preview, GPT-5.1, GPT-4o-mini and DeepSeek-V3.2) are a generation or more behind the frontier. I’m writing a longer piece on AI and hacking, out soon!

• A team from the Montreal Heart Institute, Polytechnique Montréal, CHUM, UHN, the Ottawa Heart Institute and U of A described running PACS-AI, their open-source imaging-AI platform, across six Canadian hospitals.

  • Main finding: 85% of AI jobs were completed, and clinician consults of the results rose from about 50 to about 110 a month. Most failures traced back to hospital infrastructure and integration.

  • So what: this is so typical of enterprise software. The bottleneck is rarely the shiny new tech we all talk about. It’s the boring stuff, like infrastructure, change management or systems integration.

• McGill’s Desautels Faculty of Management had eight current AI models shop for the best-value bag of coffee, and varied what they could see and what looking things up cost.

  • Main finding: with every price and weight in view, seven of eight picked the best deal 99.7% of the time. When lookups had a cost, and the instructions were vague, agents skipped the unit-price math and went for the $4.99 tag or the bigger discount, the same tricks that work on people :)

🏛️ Policy

• Tumbler Ridge update: B.C. and the Peace River South school district sued OpenAI and Sam Altman in U.S. federal court in San Francisco on September 21.

  • The suit alleges a defective product and negligence. The province wants OpenAI to cover the costs of the damage and of victim services.

  • On September 24, Mother Jones reported that ChatGPT gave the shooter tactical advice and helped her get around its safeguards. B.C.’s Attorney-General and AI Minister Evan Solomon said they were alarmed.

• Ontario committed $30M to the Vector Institute on September 25, the first award under its revamped Critical Technologies Initiative.

• Ottawa committed up to $162M over five years to Mitacs on September 25 for 10,000 AI work placements at Canadian businesses.

  • They count toward the 90,000 AI job and placement opportunities that AI for All promises by 2031.

• Nineteen more companies signed Canada’s voluntary data-centre principles on September 22, including AMD, Nvidia, Intel, IBM, and Lenovo, bringing the total to 42.

• A senior Canadian Armed Forces commander told the Canadian Robotics Council on September 22 that the military will move fast to buy low-cost autonomous systems and AI.

📊 Data

• CDW Canada found 58% of Canadian organizations now approve AI tools at work, up from 28% in 2024, and 34% of office workers use agentic AI.

📰 In brief

• Mastercard, Flybits (Toronto) and Rogers Bank completed Canada’s first agentic commerce transaction, announced September 21; we missed it last week.

• Amii‘s Ross Mitchell won a $1M CIHR grant for an AI tool for inflammatory bowel disease care (Edmonton).

• HEALWELL AI expanded its SMART Summary AI across New Brunswick’s health system.

• A U of T study found AI scribes could cut physicians’ paperwork and burnout.

• Waabi COO Lior Ron said at Nrth its self-driving trucks are ready to deploy.

• Dream Photonics (Vancouver) won the 2026 New Ventures BC competition and $110K.

• DIAGNOS (Brossard) filed its AI retinal-screening system with the U.S. FDA.

• Nvidia’s Jensen Huang called Geoffrey Hinton’s AI warnings “irresponsible” on the Ezra Klein Show; Hinton answered on Radio Atlantic.

• Tailscale (Toronto) has doubled to 40,000 business customers since November on demand from AI builders.

• Meta’s Sturgeon County data centre could raise Alberta power bills, experts told CBC.

• Quebec election: the Parti Québécois pledged an expert committee to oversee AI. The vote is October 5.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

ALL IN, the AI conference by Scale AI, took place in Montréal last week, with Germany as country of honour. The organizer put attendance at 10,000. I spent both days there. The size of the crowd and the energy in the halls matched where the AI wave is right now. Canada and Germany committed up to $300M to Yoshua Bengio’s LawZero, Cohere and Aleph Alpha finalized their merger terms, and Mila’s offices were vandalized the night before the conference opened. Radical Ventures closed the largest venture fund in Canadian history, Altis Labs raised US$25M, and Toronto’s Thri5 raised a seed round we co-led. Also in this issue: Shopify’s ads deal inside ChatGPT, a Montréal benchmark that stumps coding agents, and OpenAI’s bid to move the Tumbler Ridge suits to BC. Let’s get into it.

💰 Deals & Milestones

• Altis Labs (Toronto) raised US$25M in a Series A co-led by OrbiMed and Qiming Venture Partners USA, announced September 17.

  • Altis builds AI endpoints for cancer trials. Its IPRO model predicts survival from routine radiology scans; in a post-hoc analysis of Johnson & Johnson’s MARIPOSA trial, it detected a treatment effect 11 months before the progression-free survival readout. Innovation Endeavors, Benchstrength, Fusion Fund and Lumira participated.

• Thri5 (Toronto) raised US$5.4M in a seed round co-led by Whitecap Venture Partners and yours truly; the round closed in July and was made public last week. (Disclosure: Thri5 is a Mistral portco.)

  • We started Loblaw Digital together with Thri5 founders Jeremy Pee and Herman Paek. Pee later served as chief digital and technology officer at Marks & Spencer, and Paek ran Kijiji Canada. Thri5 turns retail data into daily actions for store, warehouse and head-office staff and tracks whether they get done. MaRS IAF, N49P, Amar Varma and Farhan Thawar participated.

• Orbits (Toronto and San Francisco) raised about US$2M in a pre-seed round from a16z Speedrun, N49P and Garage Capital, announced September 16.

  • They’re building an AI assistant for household logistics. Like a “Chief of Staff” for your family.

• SyncraSys (Halifax) closed an oversubscribed seed round on September 15, 30% above target; the amount was not disclosed.

  • SyncraSys puts an AI layer over a building’s camera feeds to analyze incidents and write the alerts and reports. CEO Darren MacLeod co-founded Liferaft, the Halifax threat-intelligence company Securitas bought in March with US$15.3M in recurring revenue.

• Radical Ventures (Toronto) closed the first US$1B-plus of the Radical Breakouts Fund on September 15, the largest venture fund ever raised in Canada.

  • The fund backs late-stage AI companies. LPs include CPP Investments, PSP, HOOPP, OPTrust, TD, BMO and CI Global Asset Management; the PMO put the target at $4B. Twelve companies are already in, including Jeff Dean’s Discovery Loop and chipmaker Etched. Radical wrote the first cheque into Cohere. Although Radical isn’t necessarily Canada-focused, it’s still great to have large growth funds in town, much needed in our ecosystem. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

• Beacon (Toronto) acquired Haize Labs, a New York startup that stress-tests AI agents before deployment, on September 17.

  • Haize co-founder Leonard Tang becomes VP of AI research, and his team forms Beacon’s applied research group. Beacon buys and runs vertical software companies and was valued at US$1.4B in June. It is the second Canadian buyer of a US AI team in a month, after SenseNet.

• Polarity, a Waterloo-founded applied AI research lab, has been acquired by Wander, the US travel-tech company behind the WanderOS rental-operator platform.

  • Founders Shane Barakat, Alex Ungureanu, and Mihai Posea researched how coding agents search and navigate large codebases, and post-trained small models for customers inside their own infrastructure. The team joins Wander’s internal engineering group and will build specialized agents and custom-trained models into WanderOS. The founders sent us the story directly.

🏢 Large Companies

• Cohere and Germany’s Aleph Alpha have agreed the terms of their merger and are seeking regulatory approval, CEO Aidan Gomez told The Logic on September 16.

  • The combined company keeps the Cohere name under Gomez, with headquarters in Toronto and Berlin and more than 1,000 employees. Aleph Alpha’s Ilhan Scheer becomes COO and Samuel Weinbach chief research officer. The deal was first announced in April. Gomez said Cohere’s models trail the frontier in coding and cybersecurity and that the company is “scaling up our models dramatically.”

  • The Globe reported on September 11 that Cohere is in advanced talks to raise US$2B to US$3B at a US$20B valuation.

• Cohere and OpenText (Waterloo) will sell a joint agentic AI stack to governments and regulated industries, with availability in early 2027.

  • OpenText supplies the data and context layer for its 120,000-plus enterprise customers; Cohere supplies the models and its North platform. Joelle Pineau, Cohere’s chief AI officer, told the Globe there is “real appetite for solutions built outside the U.S.”

• Shopify (Ottawa) is OpenAI’s first e-commerce partner for ChatGPT Ads, announced September 18.

  • A free Shopify App Store app syncs a merchant’s catalogue to ChatGPT and lets the merchant set a budget for labelled ads inside the chatbot.

  • Shopify also acquired Tailwind Labs on September 9, the Cambridge, Ontario maker of Tailwind CSS. Installs run 110 million a week, yet Adam Wathan said in January that AI coding tools cut documentation traffic 40% and revenue 80%; he laid off three of four engineers.

🔬 Research

• Microsoft Research Montréal (with KAIST) built ProgramDistill, a benchmark where a coding agent is given a working web app with its source code hidden, a broken copy with features missing, and a browser, and has to figure out what the missing features do by clicking around and then rebuild them.

  • Main finding: across nine frontier coding agents, the best restored 49.2% of complete multi-step workflows when rebuilding a whole application and the second best 28.8%. Agents that reproduce one behaviour often fail to chain behaviours into a working flow, and success falls from 100% to 64% for the top agent as the number of missing features goes from one to eight. The pipeline generated 4,063 tasks from 26 apps with no human labelling.

  • So what: this is the job behind migration and rebuild-the-legacy-system pitches, which services firms are now quoting with agents in the loop. A public release is in progress.

• Mila and DFKI, Germany’s national AI research centre, signed a strategic partnership at ALL IN on September 16 covering agentic AI, AI safety, trustworthy systems and AI for energy, with researcher exchanges between the two.

🏛️ Policy

• ALL IN 2026 ran September 16 and 17 at the Palais des congrès in Montréal.

  • Attendance was 2,300 in 2023, 4,000 in 2024 and 6,500 in 2025. Germany sent more than 50 executives with Minister Karsten Wildberger. The fifth edition is September 22 and 23, 2027.

  • On day one, hundreds of protesters marched on the venue under the banner “All Out Against All In.” Police in riot gear locked most of the doors, and we were asked to leave through an alternate exit. Police arrested a 19-year-old after a scuffle at the barricades. Radio-Canada reported the heavy police presence was a response to the vandalism at Mila the night before, covered below.

• Canada and Germany will put up to $300M into LawZero, Yoshua Bengio’s Montréal non-profit, announced September 16: up to $150M from Canada’s Strategic Response Fund and €100M from Germany.

  • The money funds Scientist AI, a system designed to reason and answer without its own goals or the ability to act. Ottawa says the project will create 360 full-time jobs and build sovereign compute with Hypertec and 5C. Germany’s share is subject to European Commission state-aid notification and funds a Berlin office. LawZero launched in June 2025 with about US$30M in philanthropic money, has just under 50 staff, and is run by CEO Sam Ramadori.

  • Bengio told The Logic the team is building a prototype and is looking for more democracies and donors to join.

• Mila‘s headquarters at Saint-Urbain and Saint-Zotique was vandalized late on September 15, the night before ALL IN opened.

  • Montréal police said about ten people broke windows around 11:45 pm and left graffiti reading “Burn the data centers” and “No.” Police questioned and released a 25-year-old woman. Bengio addressed it from the ALL IN stage the next morning: “Violence is not an answer to these risks and fears, which I can understand.”

• Mila, Mozilla and Hypertec launched an open-source AI initiative on September 17 to build a foundation layer that organizations can run locally with full control of their data.

• Terry Fox Research Institute (Vancouver) approved 14 project teams for $25M under its first Digital Health Innovation Fund, announced at ALL IN on September 16.

• OpenAI asked a US federal court on September 18 to dismiss the Tumbler Ridge lawsuits, arguing the witnesses, records and public-policy questions all sit in British Columbia and the cases belong in BC courts.

📰 In brief

• Thirty-seven Atlantic Canadian AI companies exhibited at ALL IN under a Digital Nova Scotia showcase funded by ACOA and the four Atlantic provinces; the company list spans health, energy, finance, manufacturing and cyber.

• Platform Calgary named its 2026 Launch Party Top 10 on September 19, across agtech, fintech, cleantech, AI and legaltech; the companies present on November 5 during Calgary Innovation Week.

• HIVE Digital’s BUZZ HPC signed Vancouver’s ProCogia on September 16 to run its LLM hosting and agent products on dedicated GPU capacity in BUZZ’s Canadian data centres.

• A swarm of rogue OpenAI agents used a University of Toronto link shortener as a message board in June, Reuters and CBC reported September 18. U of T disabled the function and says no systems were breached.

• OpenAI is giving free API credits and staff time to startups in Toronto-headquartered Creative Destruction Lab, which works with about 700 companies a year; no dollar value was disclosed.

• Dates and watch items: the federal AI transparency consultation closes September 23. NeurIPS author notifications land September 24. Solomon told ALL IN the builder of Canada’s sovereign supercomputer will be named “early this fall”; no winner yet.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

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

Good morning! This is the Canadian AI Newsletter. Tuesdays are the weekly rundown; today is an essay.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

All my career as a CTO, I put every line of code we shipped in the liabilities column. The asset is what the software does for users and for revenue. The code is the cost of carrying that asset.

My working definition of tech debt: any dev work on existing code that doesn’t add a new feature. That work maintains three things, and all three degrade on their own. Scalability degrades because volume grows past what the code was built for, or the code quietly stops handling the volume it used to. Stability degrades because code gets buggier as everything around it changes. Maintainability degrades because the people who understood the code leave, or they didn’t understand it fully to begin with :)

Code is the rare liability that grows while you leave it alone. Nobody touches a module for a year, and in that year a dependency picks up a security hole, traffic doubles, and the engineer who wrote it takes another job. The code sat still, and the debt grew anyway.

Ward Cunningham coined the debt metaphor in 1992. His original point was about understanding: you ship code before you fully understand the problem, and the gap between the code and your understanding is a loan. A small loan is fine, even smart, as long as you repay it by reworking the code once you know better. The key word is knowingly. You chose to borrow.

Code got cheap

Google says 75 percent of its new code is AI-generated and approved by engineers. The trajectory is the striking part: 25 percent in late 2024, 50 percent by last fall, three-quarters today. Snap reached 65 percent this spring, and Meta has been targeting similar levels by the middle of this year.

My model never depended on code being expensive. But expensive code was a natural brake on how fast the liability could grow. Every line was written by a human who at least momentarily understood it, and headcount capped the volume. AI removed the brake.

More code means more debt

Start by giving AI coding the benefit of every doubt. Assume more code means more features, and assume the quality per line is unchanged. Tech debt still grows massively, because maintenance load scales with the amount of code you carry. A dependency vulnerability hits a five-million-line codebase harder than a five-hundred-thousand-line one. Ten times the code, ten times the debt.

The quality data

Now drop the assumption, because quality per line does not look constant.

GitClear, a code analytics firm, has tracked 623 million code changes from 2023 through this year. Refactored code has fallen from 21 percent of changed lines in 2022, the last year before AI, to under 4 percent so far in 2026, while copy-pasted code has climbed to nearly 16 percent. Before AI, developers refactored about twice as often as they pasted. Today they paste five times as often as they refactor. Block duplication is up 81 percent since 2023, the highest level GitClear has recorded, and the share of changes that touch code more than a year old has fallen 74 percent. Old code is increasingly left frozen. Duplication matters because it creates debt without creating features. A duplicated block is a second copy of something you already had, and now the same bug gets fixed twice.

Google’s own DORA research, a survey of about 5,000 software professionals, found that AI adoption now correlates with higher delivery throughput. Teams genuinely ship more, and that finding deserves to be stated plainly. The same research found AI adoption still correlates with lower delivery stability, for the second year running. Their conclusion is that AI amplifies whatever discipline a team already has.

Veracode tested over 100 models and found that 45 percent of AI-generated code samples contained a known security vulnerability, a rate that has stayed roughly flat as models get bigger.

One caveat on all of this. GitClear sells code analytics, Veracode sells security scanning, and every vendor in this fight has numbers that favour its product. DORA is the interesting exception, because Google sells the AI and keeps publishing the stability finding anyway.

Then there is the witness I find most convincing, because he wants the opposite to be true. DHH, the creator of Ruby on Rails, spent five hours with Lex Fridman two weeks ago describing himself as fully AI-pilled. He runs 16 agent threads and shipped his latest release without handwriting any code. In the same conversation he described what happened when 37signals let designers vibe code on Basecamp, a large production codebase. Each pull request looked justifiable on its own. Taken together, he says, they “destroyed the architecture of the system,” and the team had to clean it up manually, by human hand. He also described watching agent code compound in his own repos: the first PR is mediocre, and five PRs stacked on top of it get bad.

The paydown side didn’t scale

Debt only matters relative to your ability to pay it down. The thing that retires tech debt is senior engineers understanding and reworking the system, and that input is roughly fixed.

It may be shrinking. Faros AI’s telemetry across 22,000 developers found 31 percent more pull requests merged with no human review at all, bugs per developer up 54 percent, and incidents per PR up 243 percent, year over year. Addy Osmani, an engineering lead at Google, points out the inversion underneath those numbers: a senior engineer used to review faster than a junior could write, and a junior with AI now generates code faster than a senior can audit it.

Osmani has a name for what accumulates when this goes on long enough: comprehension debt, the gap between the code your team has shipped and the code your team understands. It is Cunningham’s original gap in a harsher form. The gap used to open slowly, as your understanding of the problem improved past the code. Now it exists at the moment the code is written, because nobody read it.

Martin Fowler’s tech debt quadrant sorts debt by whether it was taken on deliberately. Cunningham’s loan was deliberate. Unreviewed AI code is inadvertent debt by construction. Nobody chose the tradeoff, because nobody read the code closely enough to know a tradeoff was being made. We are originating far more debt with far less underwriting.

One lab result backs up the intuition that understanding stays the scarce input. An ETH Zurich study hid the generated code entirely from participants and had them build small apps through prompting alone. Computer science knowledge was still the strongest predictor of who got good results. You can hide the code from the human. You can’t remove the human’s mental model from the outcome.

“We’ll just use AI to pay it down”

This is the natural rebuttal, and the honest answer starts with a concession. AI genuinely works for mechanical maintenance. Amazon says its Q assistant cut Java upgrades from about 50 developer-days to hours, saving a claimed 4,500 developer-years. Morgan Stanley built an internal tool called DevGen.AI that reviewed nine million lines of legacy code and saved a claimed 280,000 hours. Both numbers come from the vendors themselves, but the pattern is credible. Version upgrades, migrations, and translation are well-specified chores that a machine can check. All of that is real.

The broader claim, that AI will take care of tech debt overall, runs into three problems.

The first is revealed preference. If AI made cleanup cheap, cleanup should be booming. GitClear measures the opposite, and the collapse in refactoring happened precisely during the AI adoption wave. The tools made adding code effortless and made consolidating it only somewhat easier, so teams took the effortless path. AI paydown is available, and it isn’t happening.

The second is that comprehension debt can’t be paid down by a tool, because of what the debt is. The gap sits between the system and your team’s heads, so the paydown has to happen in the heads. AI can write excellent explanations, and someone still has to learn them. Even Morgan Stanley, running the flagship AI paydown project in the industry, had DevGen.AI produce plain-English specs and kept humans doing the rewriting, so the understanding would end up back in its engineers.

The third is that AI reviewing AI has a bootstrap problem. Investors have put roughly a quarter of a billion dollars into startups that review AI code with AI, and those tools do catch mechanical problems. But a code review’s real function was for a human to accept responsibility for a liability entering the books. When an AI reviewer approves an AI author, no human ever held the loan. The comprehension gap grows at the moment of approval, at machine speed, with a green checkmark.

The extreme version of the rebuttal says to stop maintaining code altogether: keep the spec and the tests, throw the code away, regenerate it. That works for throwaway software. For durable systems, the problem is that specs and tests are always incomplete. A spec pins down only the behaviour you thought to write down, and a production system’s users depend on far more than that: error handling, edge cases, timing, quirks nobody ever documented. Regeneration re-decides everything the spec left out. You trade code you don’t understand for new code you don’t understand.

AI pays down the debt that was cheap to pay down anyway. The debt that compounds, the kind that lives in people’s heads, grows faster under AI while the dashboards look healthier.

What I’m watching

Code is still a liability. The part of the debt that matters most has moved from the code to the heads of the people running it.

Two numbers will tell us how this plays out. The first is the share of engineering hours going to non-feature work. Stripe measured it at roughly half a developer’s week back in 2018, before any of this started. My bet is that this share rises in AI-heavy codebases even as feature output also rises. The second is lines of code per feature: how much code you end up carrying for each feature you ship. If AI coding is working the way the optimists say, that ratio holds steady. The early duplication data says it is not holding.

Near the end of that Fridman interview, DHH said something that stuck with me. He spent 25 years sweating every line, and the economic payoff of beautiful code, he now admits, was premised on humans doing the future modifications. He calls it an open question how much that premise still holds.

That is the real question the industry is deciding right now, mostly without saying it out loud: are we willing to run production systems no human understands? For some software, the answer can be yes. Some UI, one-off internal tools, anything you’d rebuild rather than repair. For the critical parts of an architecture, I don’t think it holds. If lumber and carpenters were suddenly free and abundant, you would not get strong, sustainable homes and beautiful, livable cities. You would get a lot of construction.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

Seven Canadian AI startups got funded this week! One of the busiest weeks since I started this newsletter. Fantastic end to the summer and kick-off to the fall. Ottawa also released national data centre principles with 23 signatories, and OpenAI was hit with roughly 30 new lawsuits over the Tumbler Ridge shooting. The feds launched Digital Transformation Canada with INovia’s Patrick Pichette as CEO, and SOCAN, the industry non-profit representing Canadian songwriters, sued AI music generator Suno in Federal Court. I’d forgotten this, but Canadian law has no concept of fair use. Let’s get into it.

💰 Deals & Milestones

• Mundo (Vancouver) announced US$24M in total funding: a US$20M Series A led by GreatPoint Ventures plus a previously undisclosed US$4M seed.

  • The 30-person company, founded in 2024 by four UBC alumni, builds the audio, video, and multimodal datasets and evaluations that AI labs use to train models on sensory information. Y Combinator, E12 Ventures, and Next Frontier Capital participated.

  • The announcement was August 24, and I missed it last week, so sharing it now to catch up. Or catch-back? You get what I mean.

• Meissner (Toronto) closed US$2.6M in pre-seed funding led by BDC Capital’s Thrive Venture Fund.

  • The company uses AI to identify metallic alloys that could work as superconductors, with quantum simulations to filter candidates, and runs lab work at the University of Waterloo’s nanofab centre. CEO Olivia Leng founded it a year ago while studying chemistry at U of T.

  • Angels include Dominion Dynamics CEO Eliot Pence, General Fusion CEO Greg Twinney, Daniel Debow, Michael and Richard Hyatt, Anthony Lacavera, and Xanadu CEO Christian Weedbrook.

• Scopia Surgical (Montréal) closed an oversubscribed C$2.65M pre-seed co-led by Linearis Ventures and Anges Québec.

  • The computer-vision platform pairs real-time spatial tracking with patient-specific digital twins to guide minimally invasive and robotic surgery. Fonds Impulsion, HaloHealth, and Intelerad co-founder Richard Rubin participated, and the funds carry the company into first clinical evaluations.

• Protein Industries Canada (Regina) committed up to $2.2M to two AI agtech projects worth $4.9M combined.

  • TerraVision360, Metos Canada, and Rocky Mountain Equipment will build a blight-forecasting tool for pulse crops. Super GeoAI and Southview Farms will build a drone-and-LiDAR system that measures grain volumes in bins so workers stop climbing them.

• The Wellness Company (Toronto) closed nearly C$1.6M in seed funding led by BDC Capital’s Seed Venture Fund, with BKR Capital and Jason Calacanis’ Launch.

  • Its product Tempo is an AI health companion that lives in iMessage and builds personalized plans from the company’s health and wearable apps.

• NorBot (Vancouver) raised C$1M in pre-seed funding from undisclosed Canadian founders and operators to become a commercial robotics integrator.

  • NorBot sources, deploys, and manages robots for Canadian businesses through its fleet-management and digital-twin software, with supplier agreements signed with four manufacturers and paid pilots underway with one of Canada’s largest commercial landlords.

  • The pitch includes keeping the data those robots generate in Canada.

• ASI Engineering (Regina) landed $980K in repayable federal funding from PrairiesCan to scale its AI platform for monitoring roads, buildings, and water systems.

  • This is government scale-up money rather than an equity round.

  • Good to see Saskatchewan in the deals column twice this week.

🏢 Large Companies

• OpenAI was hit with a second wave of lawsuits over the February Tumbler Ridge school shooting, filed September 2 in US federal court in California.

  • Chicago’s Edelson PC and Vancouver’s Rice Parsons Leoni & Elliott represent roughly 30 new plaintiffs, including students, teachers, and the school’s principal, joining the seven suits filed in April. Sam Altman is named as a defendant.

  • The new complaints add an aiding-and-abetting claim and allege OpenAI’s safety team repeatedly urged notifying authorities about the shooter’s ChatGPT activity and was overridden by public-affairs leadership. Yikes. One complaint contrasts OpenAI locking down its own offices when staff were threatened with its silence toward Canadian authorities.

  • OpenAI has separately asked the court to dismiss the original seven suits, arguing British Columbia is the proper venue. The BC government’s threatened claim still hasn't been filed.

🔬 Research

• York/Vector + Mila probed what two leading video foundation models, V-JEPA 2 and VideoMAE-v2, actually encode, layer by layer.

  • Main finding: the models detect camera motion with over 90% accuracy but score at coin-flip levels on basic physics like object permanence, at every layer of every model tested.

  • So what: these models are the base layer of the physical-AI wave, and the result contradicts Meta’s own earlier claims on a harder benchmark.

  • Physical AI is a key piece of the next frontier. This study shows video models are not quite there yet. Helps explain why so much money has gone into this research.

• UdeM/Mila astrophysicists tested whether tabular foundation models, which fill in spreadsheet values the way chatbots fill in text, learned any physics along the way.

  • Main finding: the models beat six conventional baselines across 316 physics equations, even after the baselines got tuning help, yet their prior cannot represent an exact noise-free law or physical units. Switch metres to feet and the model doesn’t notice.

  • So what: these tools interpolate physical data impressively without being physical models, which scientists adopting them for real measurement work need to know.

• ÉTS Montréal + Mila tested whether a model that has “unlearned” data for privacy or copyright reasons has actually forgotten it.

  • Main finding: synthetic probe inputs plus a single gradient step can reactivate a supposedly erased class, using nothing except the scrubbed model itself. Low accuracy on forgotten data often means the knowledge is hidden rather than gone.

  • So what: as deletion obligations surface in privacy and copyright disputes, claims of unlearning will need audits like this to mean anything. This study showed that current ways of “deleting” are not really, you know, “deleting.”

🏛️ Policy

• Ottawa launched Canada’s Responsible Data Centre Development Principles on September 3, announced by AI Minister Evan Solomon in Markham with 23 corporate signatories.

  • The five principles: create lasting local benefits, do not shift electricity costs to Canadians, minimize water and environmental impacts, be transparent about local impacts, and bring strategic value to Canada. Signatories include AWS, Google, Microsoft, Anthropic, Meta, OpenAI, Bell, TELUS, QScale, eStruxture, and Cohere. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • The framework carries no binding definitions, no penalties, and no commitment on what energy powers new builds. It sits alongside Ontario’s draft playbook, which charges data centres full electricity costs, and Alberta’s 2% hardware levy taking effect December 31.

  • The provinces still hold the tools that matter here: rates, permits, and grid connections.

• The federal government launched Digital Transformation Canada on September 3 and named former Google CFO and current Inovia Partner Patrick Pichette as its first CEO.

  • The agency absorbs Shared Services Canada, the Canadian Digital Service, and functions from Treasury Board and PSPC, with a mandate to scale AI across government and use federal procurement to help Canadian tech firms grow. Pichette led Inovia’s investments in Cohere, Hopper, and Wealthsimple.

• SOCAN (the Society of Composers, Authors and Music Publishers of Canada, which collects performance royalties for more than 200,000 Canadian songwriters, composers, and publishers) sued AI music generator Suno in the Federal Court of Canada on September 2, alleging its platform generates and streams outputs that copy songs in SOCAN’s repertoire.

  • The claim cites 150 sample outputs resembling works by Avril Lavigne, The Tragically Hip, Tom Cochrane, Céline Dion, and Joni Mitchell, and seeks damages plus an order to stop.

  • Fair use, Suno’s defence in its US label suits, does not exist in Canadian copyright law. CEO Jennifer Brown framed the case as cultural sovereignty.

📊 Data

• StatCan released the final Canadian Survey on Business Conditions on August 31, closing out the quarterly series that has tracked business AI adoption since 2020.

  • 25.2% of businesses plan to use AI in the next 12 months, up from 14.5% a year ago. Of the 52.7% with no plans, 79.1% say AI is not relevant to what they produce.

  • A successor program, TechStat, takes over with $25M from Budget 2025.

📰 In brief

• Ontario’s consultation on its draft Data Centre Playbook closes September 12.

• Alberta’s data centre town hall tour wraps up in Grande Prairie on September 11.

• Co.Labs (Saskatoon) launched a program helping non-tech businesses adopt AI.

• Opti9 acquired Hut 8’s Canadian managed cloud business, including a multi-year colocation agreement at Hut 8’s Canadian data centres.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

I love this video. It makes me laugh every time. I also think it captures what current models are good at and bad at better than any benchmark.

Ask an AI to make you money in the stock market, or to run your fantasy team in a sport you don’t follow super closely, or to build you a profitable business by tomorrow morning. What comes back will sound smart. It will be confident and fluent, and it will lose. The reason it loses is worth walking through, because it makes us better AI users, and it also points to where the real moats in AI startups are.

You can only delegate what you can evaluate

The prompt is the first step. After the prompt, an expert user makes hundreds of micro-corrections along the way. An expert can tell output that is 80 percent right from output that is confidently wrong. A novice sees the same fluent paragraph in both cases. Remove the evaluator and the model settles into a plausible-sounding average of everything it has read.

Is coding different?

Sure, people who cannot code ship real apps now. Andrej Karpathy named vibe coding in February 2025, and it works because it’s verifiable and compilers and tests grade the work. The app runs, or it crashes. The error message tells you what to paste back into the chat. The user can’t judge the code, and it doesn’t matter. Delegation works when something can grade the output. Your own judgment can do the grading, or something else. You need one of them in the loop.

Why the intelligence is jagged

This pattern comes straight from how the models were trained. Many of the sharpest recent gains in reasoning came from reinforcement learning on verifiable rewards. Let the model attempt a problem a million times and reward it whenever the answer checks out. Math comes with answer keys, and code gets graded by compilers and test suites. That recipe took frontier models from mediocre competition mathematicians to IMO gold medalists in about two years.

No equivalent flywheel exists for writing a great strategy memo, making a good hire, or picking a stock. Nobody has a cheap, automatic checker for those, so nobody could train against them at scale. The closest thing those domains got was training on human thumbs-up, and a rater who rewards whatever sounds right since they can’t check for correctness (they’re not experts in that particular field). Ethan Mollick and his co-authors called the result the jagged frontier back in 2023: AI that beats experts on some tasks and stumbles on neighbouring ones. Since then, the models have soared wherever cheap verification existed and remained uneven elsewhere.

Markets add one more problem. A model trained on the internet’s text is a consensus machine, and whatever everyone already believes about a stock is baked into its price. Agreeing with everyone earns you nothing.

Confident mediocrity

On tasks with no answer key, the models are usually decent. They write a passable memo, a reasonable stock thesis, a defensible fantasy lineup. And they deliver everything with the same assurance they bring to gold-medal math. I call this confident mediocrity. Confident mediocrity is exactly what loses in the stock market and in fantasy leagues, because zero-sum games grade you against players with real edge. The model cannot tell you when it is wrong, and a non-expert user cannot tell either. That leaves two blind parties at the table and no referee.

Next frontier

Labs are working on this, and some of their current research attempts to manufacture answer keys where none exist: rubric-based rewards, models grading other models, supervision of reasoning steps instead of final answers. The third AGI test from ARC tests for some of this. Labs will show us whether judgment (or taste) can be turned into a reward signal at all. The thing to watch is whether training on these manufactured keys starts producing jumps like the math results.

What this says about AI today

We have superhuman intelligence for the subset of reality that comes with an answer key. AGI is the rest of it. My personal primitive AGI benchmark comes from the video at the top. The day an AI wins me a fantasy league in lacrosse, we’re there. (cuz, you know, I don’t follow lacrosse - not sure who does).

The so what for startups

Lack of judgment or a grader in the loop produces AI slop machines. Exhibit A: content marketing generation tools ;)

Products built for experts can leave the grading to the user. A lawyer can review a draft contract. A developer can inspect generated code. An investor can challenge a stock thesis. These products can be very useful, but they still depend on expertise outside the software. And their market size is capped by the number of experts. Some expert markets are big enough; many are not. And you’ll have to look for a moat in something other than the proprietary answer key.

Sometimes the world provides the answer key. The code compiles. The payment arrives. The parcel reaches the customer. These checks make the product easier to use, although every competitor has access to the same answer key. Again, no moat there.

In other cases, the company has to build its own answer key. That might include proprietary evaluations, a domain-specific verification system, proprietary datasets or a human review process that generates better feedback over time. Or all or some of it.

That third answer is the only one the company owns.

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

The big story this week is bank earnings. TD told analysts it has essentially hit its $200 million fiscal 2026 target for value from AI with a quarter still to go, and RBC set a target of $700 million to $1 billion in enterprise value from AI by the end of fiscal 2027. Also this week: Alberta’s virtual data-centre town hall drew 12,500 viewers and a moratorium call from the NDP, Ottawa named 64 new research chairs recruited mostly from US universities, and Kepler’s optical relay network started commercial service. Plus a new Cohere model and fintech numbers from KPMG. Let’s get into it.

💰 Deals & Milestones

• Provision (Toronto) raised a US$7M seed round led by Cercano Management, with Y Combinator and One Way Ventures, to build AI co-pilots for pre-construction estimating.

  • The round closed in August. The product targets the estimating and proposal work general contractors do before a project breaks ground.

  • Cercano is the investment firm that grew out of Vulcan Capital, Paul Allen’s family office.

• Kepler Communications (Toronto) began commercial service on its optical data relay network, the first commercial space data relay service available.

  • The network moves data between satellites over laser links, so a customer in low-Earth orbit can downlink in near real time instead of waiting for a ground-station pass. The first tranche of ten 300 kg satellites launched in January and carries about 40 Nvidia Jetson Orin modules, so customers can run AI workloads such as Earth-observation analytics and sensor fusion in orbit and send down results instead of raw data.

  • A second tranche is planned for 2028 with optical terminals up to 100 Gbps, a dedicated Rocket Lab launch, and a European Space Agency HydRON mission.

• The City of Toronto launched a one-year pilot of Clariti’s CivCheck, an AI pre-check from the Vancouver govtech company that flags missing documents and building-code issues in permit applications before submission.

  • Toronto reviews more than 36,000 building-permit applications a year. Honolulu reported 55% faster permit decisions with the same tool in the first quarter of 2026.

  • The tool is voluntary and does not approve or reject anything. City staff keep the review.

• MDA Space (Brampton) launched LaunchPad Ventures, a corporate venture program that will invest directly in Canadian space and defence SMBs across 25 focus areas, including autonomous systems.

  • The Canadarm maker frames the program around sovereign capability gaps in Canada’s space and defence supply chain.

• SenseNet (Vancouver) acquired the wildfire division of Maryland-based N5 Sensors, taking on its team and more than 30 US federal and enterprise customers.

  • SenseNet says its AI-driven detection network, which fuses sensor, camera, and satellite data, covers 400 million acres across 12 countries. Terms were not disclosed.

  • Good to see a Canadian company buying the American team for a change.

🏢 Large Companies

• Five of the Big Six banks reported fiscal Q3 results between August 25 and 27, and four of them disclosed concrete AI numbers.

  • TD delivered about C$195 million of its C$200 million fiscal 2026 AI value target in the first three quarters. CEO Raymond Chun told analysts the bank has “essentially hit” the full-year target, with more expected from predictive, generative and agentic use cases. The medium-term goal is C$500 million in annualized revenue and C$500 million in annualized cost savings from AI. During the quarter, TD Auto Finance Canada automated about one-third of its manual funding processes.

  • RBC CEO Dave McKay pointed to a target of “$700 million to $1 billion in enterprise value by the end of fiscal 2027,” built on close to ten years of investment in its Borealis research institute. Management said the AI work is funded inside the current budget.

  • Scotiabank said Scotia Navigator, the internal AI assistant it launched in April, reached 95% employee adoption, with assistive AI actions up four-fold and more than 80% of its global workforce through at least one AI training course.

  • BMO reported that its commercial-banking AIM platform enabled more than 35,000 client conversations, its Lumi chatbot lifted new-employee productivity 15%, and its Next Best Offer engine added a 3% lift to sales revenue.

  • CIBC pointed back to its July launch of an agentic AI workspace on its call. National Bank disclosed nothing AI-specific.

• Cohere (Toronto) released Parse, a vision-based document parsing model for high-volume enterprise document processing. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • Cohere positions Parse on price-performance for high-throughput document work. It is the company’s second product release this summer, after North Automations in July.

🔬 Research

• Researchers at Nvidia’s Toronto lab and the University of Toronto posted GLOSS, a method for filling missing or damaged texture in images and 3D assets by learning how a pattern repeats across a surface and continuing it faithfully from a reference.

  • Current AI fill tools tend to invent texture that looks plausible without matching the source material, which matters in game, film, and simulation pipelines. Authors include Masha Shugrina and James Lucas, with Princeton collaborators.

  • The lab keeps publishing after Sanja Fidler’s departure in July.

• Doina Precup and collaborators at McGill and Mila posted SIC-Agents, a benchmark and LLM-based simulator for training pediatric clinicians in serious-illness conversations with families.

  • The simulator handles multi-party conversations, parental distress, and curriculum-dependent feedback, and was built with McGill medical educators and practicing clinicians.

🏛️ Policy

• Alberta’s fourth data-centre engagement, a virtual town hall on August 27, drew more than 12,500 viewers, and NDP leader Naheed Nenshi called the same day for a moratorium on new AI data-centre approvals.

  • Premier Danielle Smith and three ministers took calls for an hour in a session far calmer than the in-person events in Lacombe and Redwater. Technology minister Nate Glubish said abandoned data-centre infrastructure would fall to the province, and that he expects any stranded building would find a buyer at a steep discount.

  • Nenshi: “There is no framework in place, and they’re making it up as they go along.” He cited the utilities commission’s rejection of a gas plant near Olds and an estimate that Meta’s Sturgeon County project raises household energy costs from $267 to $462 a year until its own gas plant comes online.

  • The final town hall is September 11 in Grande Prairie.

• Ottawa named the first 64 Eddie Goldenberg Research Chairs, recruited from 13 countries with 48 coming from US institutions, and five of the chairs are explicitly AI-focused.

  • Aram Amassian joins Polytechnique Montréal for AI-driven semiconductor discovery aimed at technological sovereignty. Tanya Berger-Wolf takes an AI for Nature chair at U of T, Ajit Khosla an AI environmental-intelligence chair at the University of Regina, James LeBeau an AI-enabled electron microscopy chair at McMaster, and Patrick Charbonneau a materials-and-AI chair at Université de Montréal.

📊 Data

• KPMG put Canadian fintech investment at US$996.7 million across 47 deals in the first half of 2026, and AI/ML led all verticals with 19 deals.

  • That is more than double the next vertical, digital assets, at 8. Total investment fell more than 40% from US$1.7 billion across 82 deals in the first half of 2025.

  • KPMG’s Dubie Cunningham says capital has moved from broad AI experimentation to fintechs using AI to make deposit-taking, lending, and payments faster and cheaper: “Fintech investors aren’t investing in AI for AI’s sake.”

• An IDC survey commissioned by Cohere found Canadian enterprises rank last of four surveyed countries in sovereign-AI awareness and first in seeing it as a competitive advantage. (Disclosure: portfolio company.)

  • 10% of Canadian respondents report high awareness of sovereign AI, against 29% in Germany, 15% in the UK and 12% in the US. On drivers, 35% of Canadian respondents cite competitive advantage, the highest of the four countries.

  • IDC surveyed 508 decision-makers at billion-dollar-plus enterprises in April and May, roughly 125 to 130 per country, so the national numbers are directional.

📰 In brief

• Ontario’s Data Centre Playbook consultation remains open until September 12.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

The big story this week is Veeda AI, Sanja Fidler’s Toronto world-model startup, which came out of stealth on August 19 with a seed round of more than US$90M, one of the three largest ever raised in Canada. Also this week: Anthropic posted jobs for a “gigawatts” compute buildout in Canada, all four Quebec parties signed a code of conduct on AI use ahead of the October 5 election, and Alberta’s data-centre town halls drew hundreds of angry residents. Plus two East Coast rounds we missed earlier this summer, and research on why neural networks forget how to learn. Let’s get into it.

💰 Deals & Milestones

• Veeda AI (Toronto) raised more than US$90M (~C$124M) in seed financing co-led by Radical Ventures and Khosla Ventures to build world models for physical AI.

  • Founder Sanja Fidler left her post as Nvidia’s VP of AI Research on July 31, covered here two weeks ago. Co-founders Huan Ling (Toronto) and Zan Gojcic (Zurich) are also ex-Nvidia, and Fidler remains a U of T professor.

  • Radical’s Tomi Poutanen and Khosla’s Sven Strohband are joining the board. PitchBook ranks it among the three largest seed rounds in Canadian history.

  • At a time when https://dominionlist.com/ keeps getting cited, fantastic to see a team of this calibre choose to build from Toronto. 🇨🇦 🇨🇦 🇨🇦

• stratosX (Montréal) signed Air Transat as its launch airline customer for X360, its AI platform for recovering from flight delays, cancellations, and schedule changes.

  • X360 uses predictive modelling and generative AI on the airline’s own operations data to recommend how aircraft, crews, and passengers should be rerouted. The team includes veterans of Southwest Airlines operations and QuantumBlack.

• Blueprint (Toronto) raised C$1.4M through a16z’s speedrun program to turn text prompts into buildable hardware designs.

  • CEO David Feldt describes it as “Lovable for hardware.” The team is in San Francisco for the accelerator and plans to return to Toronto.

• Me-Z-Scene (Montréal) raised C$875K from AQC Capital, Anges Québec, and Investissement Québec’s Eurêka Fund for Élisa, a generative tool that lets anyone build worlds in Roblox.

  • The father-son founding team has about 100 test users and plans a full launch this fall.

• Floqer (Halifax) raised a $2M pre-seed from Perplexity’s F7 Fund, N49P, Tidal Venture Partners, and Perplexity co-founder Denis Yarats, announced July 8. We missed it at the time, so it gets its catch-up here.

  • Floqer unifies 100+ external data sources with a company’s own CRM, Slack, and sales-call data into one customer knowledge base for sales teams. Less than two years in, it is past seven figures in ARR, cashflow positive, and serves 1,000+ customers including Perplexity and Wise.

• Siftmed (St. John’s) closed a round led by Toronto’s Staircase Ventures, announced July 20, which brings total funding past $10M for its AI platform that analyzes medical records in insurance claims. Another one we missed this summer.

  • The software cuts claim-review time by 50 to 70 per cent and traces every finding back to the source document. CEO Holly Hill reported 10x revenue growth in 2025.

  • Pelorus VC and a group of Verafin founders also participated. Good to see the Verafin founder network backing the next generation in St. John’s.

• Fisent Technologies (Toronto) raised US$4.3M in its first priced round, led by FINTOP with strategic investor Pegasystems, announced August 11.

  • Fisent’s BizAI platform turns unstructured documents into automated decisions at banks, insurers, and wealth managers. Revenue grew 206% in 2025, the company landed its first Fortune 50 customer this year, and total funding now stands at US$6.3M.

• Light AI (Vancouver) received a Health Canada medical device licence for QuickScan Strep A, announced August 12.

  • The Cboe-listed company screens for strep throat from a smartphone photo using image-analysis AI.

🏢 Large Companies

• Anthropic posted job listings for a compute country lead and an Alberta-based community engagement manager, a signal of plans to bring “gigawatts of compute online in Canada”.

  • The Logic reported the postings on August 19. The compute lead would own site selection, leasing, and financing, and would scale the buildout “from initial sites to multiple gigawatts.” Anthropic also hired Jacob Glick, formerly of Amazon, as its Canadian public-policy head earlier in August.

  • No site, dollar figure, or timeline has been committed. These are job postings, and hiring plans can change.

🔬 Research

• Tyler Kastner, Nimrod De La Vega, and Amir-massoud Farahmand (Polytechnique Montréal / Mila) posted a fix for neural networks that lose the ability to learn (arXiv, August 18).

  • Neural networks have a weird flaw: if you keep training one on new tasks, it gradually loses the ability to learn anything new at all. It gets rigid, like a muscle that stops responding to exercise. This is called “loss of plasticity,” and it’s one of the main reasons you can’t just keep teaching an AI system new things forever, a big problem for robots and agents that need to learn continuously in the real world instead of being trained once and frozen. This paper found a specific mathematical cause of the rigidity (some internal numbers in the network drift wildly out of balance over time) and a simple fix: periodically trim the extremes back. Why it matters: continual learning is a bottleneck for the whole “AI that keeps learning on the job” agenda, and the fix here is simple enough to actually be used. Farahmand is a well-known Canadian RL researcher,

• Yoshua Bengio and Turing laureate Shafi Goldwasser, with Orr Paradise and Oliver Richardson, posted a paper on auditing AI probability claims (arXiv, August 11).

  • When a model says “there is a 70% chance this loan defaults,” a single probability can’t be graded against one outcome. The paper works out how an outside checker can verify that a system’s probability claims are honest and self-consistent without trusting the system itself.

• Ontario Tech University and the Centre for Designing Change launched PACE Ontario, the first provincial hub for physical AI, in Oshawa on August 20.

  • Physical AI means systems that operate in the real world: robots, machines, and sensors. Oshawa sits in the heart of Ontario’s auto-manufacturing belt.

🏛️ Policy

• All four parties in Quebec’s National Assembly endorsed a code of conduct on AI use in politics ahead of the October 5 provincial election, announced August 24.

  • The CAQ, Liberals, Parti Québécois, and Québec solidaire commit to transparency, responsible use, security, and prevention, including a pledge against AI-generated content that could mislead voters. The CAQ says it will not use AI in its campaign at all.

  • IVADO and CEIMIA developed the code under Catherine Régis (Université de Montréal, Canada CIFAR AI Chair at Mila), Christian Gagné (Université Laval), and Marc-Antoine Dilhac (Université de Montréal).

• Alberta’s Technology and Innovation Minister Nate Glubish held the province’s first two town halls on AI data centres, and both turned hostile.

  • Hundreds packed the August 19 session in Lacombe, where nobody who reached the microphone endorsed the plans. The August 20 session in Redwater sits near Meta’s planned $13B data centre in Sturgeon County, to be powered by a $4.6B gas-fired plant.

  • A virtual session runs August 27, and Grande Prairie follows September 11.

• Quebec’s Ministère de la Cybersécurité et du Numérique signed a $2.1M agreement with Université Laval to build a secure, sovereign testbed for AI in Quebec public services, announced August 21.

  • The funding runs two years with a possible $1M third-year top-up. Laval’s Institut intelligence et données, led by Christian Gagné, will run the sandbox where public-sector teams test AI use cases.

  • In a La Presse interview the same day, Minister France-Élaine Duranceau said the province is counting on AI to help balance the budget.

• Canada’s expanded SR&ED rules are now fully in effect following Bill C-15’s royal assent in March.

  • The enhanced 35% refundable credit now applies to the first $6M of qualifying spending per year, double the old limit. Capital expenditures are eligible again, and eligible Canadian public corporations can now access the enhanced rate.

  • If you run a deep-tech company and have not re-run your SR&ED math since March, it is worth an hour with your accountant.

📰 In brief

• ALL IN 2026 revealed its fourth-edition program, September 16-17 in Montréal: 7,500+ attendees, Germany as Country of Honour, and a lineup including Cohere, Google DeepMind, NVIDIA, OpenAI, and Mistral AI (not us :).

  • I will be there, so reach out if you want to connect!

• Osedea, a Montréal AI and robotics consultancy, acquired Sherbrooke business-intelligence firm ventriloc. Both companies say the jobs, expertise, and IP stay in Quebec.

• CIBC’s AdvisorAssist, covered here in Issue 20, won an IDC CIO Award Canada for its AI advisor tool.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

This is a different kind of issue. Every quarter, we send our LPs a report on each fund, with a market update, company updates, fund financials, and a section I write on where we see AI heading. This quarter I took on a question I keep hearing from founders: if capable models keep getting cheaper and closer together, where does the value go? Below is that section, unchanged apart from light formatting. It's global rather than Canadian in scope, but what it describes resonates here too. The regular Tuesday issue returns next week.

From our latest quarterly report:

A year ago, we wrote that open-source AI was tracking the same path as Linux and PostgreSQL: the model layer would commoditize, and value would move up the stack. One year later, that call has mostly played out and faster than we expected.

DeepSeek released V4 in April with fully open weights, priced at roughly a tenth of comparable Western APIs and with near-frontier coding performance. Alibaba’s Qwen models have passed one billion cumulative downloads. Yet the Western “Red Hat of AI” we hoped to see has not emerged, and leadership in open models is even more concentrated in China than a year ago.

The bigger surprise is that the closed frontier is commoditizing too. In July, five labs released flagship models within two weeks, with results close enough that several clear the bar for many enterprise uses. It reminds us of digital cameras in the 2000s: once every camera had enough megapixels, buyers stopped comparing sensors and started comparing lenses and software. Enterprises are approaching that point with AI. Buyers used to ask which model is best. Now they ask which models are good enough, and which one offers the best price, speed, reliability, and control.

Why, then, do the model companies keep growing so quickly? Falling prices can be offset by expanding consumption, and the leading labs are moving up the stack. The best available estimates suggest that more than half of Anthropic’s revenue comes from coding, while roughly 70% of OpenAI’s comes from ChatGPT subscriptions. Neither discloses the exact mix, but Anthropic seems to have been propelled largely by a single enterprise workload (coding) and OpenAI by a general-purpose consumer product. Neither has built its growth across a broad set of enterprise workflows yet. That market is still early.

Uber recently showed what a sophisticated enterprise buyer looks like. Following its Q2 earnings, CTO Praveen Neppalli Naga wrote that Uber had more than quadrupled the number of engineers using frontier AI tools since January, even as its cost per token declined. Engineers see their costs in real time, while Uber tests proprietary and open-weight models and uses the least expensive option that meets each task’s requirements. He described it as the end of “tokenmaxxing,” when consuming more tokens was itself treated as progress.

This is enabled by the harness: the software between a company’s workflows and the models that handles routing, caching, evaluations, and monitoring. A good harness turns switching models into an operational decision rather than a rebuild. Companies can then route work to the least expensive model that clears the quality bar. CIOs have always hated vendor lock-in, so this hits home.

None of this means every model is identical. But capable intelligence is getting cheaper, more abundant, and easier to swap out, and enterprises are starting to buy accordingly. That is good news for nearly everyone building above the model layer, and the broad enterprise market is still up for grabs. As enterprises optimize their spend at the model layer, our going-forward focus stays on the picks/shovels and application layers where we believe the next phase of value capture will take place.

See you Tuesday.

-Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

The big story this week is Ontario’s draft Data Centre Playbook, announced August 13, which tells data centres to pay their full electricity costs and expect no cash incentives. Also this week: Kitchener-Waterloo’s Convictional is shutting down and returning capital to investors, Peripheral raised US$8.7M to bring self-driving-car techniques to live sports, and two Montréal startups took on elder fraud and construction financing. Plus a Nanos poll on data-centre incentives and Mila research on securing AI agents. Let’s get into it.

💰 Deals & Milestones

• Convictional (Kitchener-Waterloo) is shutting down on August 27 and returning its remaining capital to investors, after its pivot to an AI-native team communication tool failed to find product-market fit.

  • The Y Combinator-backed company raised $50.7M CAD for its original B2B commerce platform, divested that business (Modern Dropship, 3,000 customers, roughly US$2M net revenue) in early 2025, and bet the treasury on rebuilding Slack for the AI era.

  • CEO Roger Kirkness says several years of runway will go back to investors. This can be a controversial topic among founders and investors. If the original thesis doesn’t work out, and there’s money left, what do you do? Keep trying different ideas until the money is out? Or do you return capital? In my experience, it’s really very case- and individual-specific. What I do know for sure is that it takes a lot of courage to do what Roger did.

• Peripheral Labs (Toronto) raised US$8.7M from Deloitte Ventures and Inovia Capital to build spatial intelligence for live sports, announced August 17.

  • The company reconstructs 3D volumetric video from standard broadcast feeds using perception techniques from autonomous driving. Founders Kelvin Cui and Mustafa Khan came out of the University of Toronto’s self-driving and robotics teams and launched the company from San Francisco in October 2024.

  • The raise follows a US$3.6M seed led by Khosla Ventures and the QSLA basketball biomechanics lab the company opened in Toronto, covered here in May.

• Legio (Toronto) raised a US$5M seed round from Klass Capital for an AI-native campus safety platform for higher education, announced August 11.

  • The software sits on live dispatch and incident data and auto-generates the compliance documents campus security teams produce by hand, including daily crime logs and Annual Security Reports.

  • Founded in April 2026 by CEO Darren Hill, CTO Mike Wertman, and CPO Joe Crampton. Hill is an operating partner at Klass and Wertman is Klass’s CTO; Klass is the round’s sole investor.

• OnGuardAI (Montréal and Atlanta) launched an app that flags scam calls in real time, texting both the user and a family member while a suspicious call is still in progress, released Friday to coincide with National Financial Awareness Day in the US.

  • Co-founder and CEO Richard Gotlieb started the company after his father was targeted by a scam caller. Analysis happens only during the call; the company says it keeps no recordings or transcripts once a call ends.

  • The app targets seniors, the group phone fraudsters most often target. The Canadian Anti-Fraud Centre puts Canadian fraud losses at $643M in 2024, up nearly 300% since 2020.

• Brdg (Montréal) raised an $850K pre-seed round from Forum Ventures and undisclosed strategic investors to automate construction loan draws, announced August 17.

  • Draws are the phased payments lenders release as a project hits milestones, a process still run manually between developers, cost consultants, and lenders. Co-founders Ness Cabessa and Samuel Brand came from general contracting and development, respectively, and launched the company in 2024 with Daniel Bensoussan.

  • The round closed earlier this month as a SAFE. The closest comparable is Built in the US; Brdg’s distribution runs through partnerships with cost consultants.

🔬 Research

• Christopher Pal of Mila and Polytechnique Montréal, with a ServiceNow Research team including Gabriel Huang and Alexandre Drouin, posted a study on cleaning compromised AI agents (arXiv, August 11).

  • A backdoor is a hidden instruction planted in a model, through poisoned training data, that makes an agent misbehave when it sees a specific trigger phrase. The risk for companies deploying agents is that the model looks normal until someone uses the trigger.

  • The team’s fix: deliberately plant a backdoor you control, then train the model to unlearn it, which pulls the hidden ones out with it. Across 115 experiments, the method wiped out nearly all planted backdoors, though faint traces of the triggers remained detectable inside the model.

🏛️ Policy

• Ontario released a draft Data Centre Playbook on August 13, requiring data centres to pay their full electricity costs, with no cash incentives and a new dedicated rate class for large projects priced above the industrial rate.

  • Premier Doug Ford, announcing the framework in Guelph: “If any data centre thinks they’re coming into Ontario and getting a freebie: not happening.” Projects will be assessed on three pillars: economic development, data security and digital sovereignty, and community benefit. A 30-day consultation is open, and the province projects data centres will account for about 13% of new electricity demand by 2035.

  • The playbook is a first piece of Ontario’s forthcoming AI Industrial Strategy. It is the firmest line any province has drawn under the data-centre buildout so far.

  • The no-incentives stance largely reflects public opinion, here and abroad; see the Nanos survey below. The nuance: opposing government subsidies for data centres is not the same as opposing data centres.

• The Council of Canadian Innovators published “The Scale-Up Gap” on August 11, based on interviews with 31 founders across 30 Canadian companies acquired by foreign buyers.

  • The report finds companies sell when scaling becomes most complex and capital-intensive, and it names four recurring barriers: domestic customers, growth capital, specialized talent, and ecosystem cohesion. In 93% of cases, leadership and strategic decision-making moved abroad after the acquisition.

  • AMD’s acquisition of Taalas, covered here last week, is among the featured exits.

  • In my experience, the main issue is a lack of growth capital in the rounds before an acquisition. When a Canadian startup raises a large A or B, or later, it’s likely led by a US firm, which likely pushes it to switch to Delaware. Once that is done, getting acquired by a US company really makes no difference.

📊 Data

• A Nanos Research poll for the Globe and Mail found 64% of Canadians oppose or somewhat oppose government financial incentives for AI data centres, published August 11.

  • The split: 38% opposed, 26% somewhat opposed, and 28% supportive or somewhat supportive. It is the first national polling specifically on data-centre incentives.

📰 In brief

• An Ottawa tech worker built an app that detects nearby smart glasses to warn people they may be recorded.

• Vass Bednar of the Canadian Shield Institute joined the BetaKit Podcast to ask whether Canada can function without American tech.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

The big story this week is AMD’s acquisition of Taalas, the Toronto startup that hardwires AI models into silicon, announced August 6 for undisclosed terms. Also this week: Cohere set up a standalone APAC headquarters in Seoul, Shopify posted 34% revenue growth, and TELUS took a $2.1B impairment it partly pins on AI. Plus Alberta’s new pitch that data centres are digital refineries, and a record month for AI-hardware imports. Let’s get into it.

💰 Deals & Milestones

• Taalas (Toronto) agreed to be acquired by AMD on August 6; terms were not disclosed.

  • Taalas etches a model’s weights directly into custom silicon for inference. Founded in 2023 by Tenstorrent co-founder Ljubisa Bajic with Lejla Bajic and Drago Ignatovic, it raised US$169M this February from Quiet Capital, Fidelity, and Pierre Lamond.

  • AMD plans to pair the technology with its Instinct GPUs and Helios rack systems, and its release cites “a continued commitment to retaining and growing Canadian talent.” Closing is expected in Q4, subject to regulatory approval.

  • It is the second Toronto AI-chip company to draw US acquirer interest this summer, after the Qualcomm interest in Tenstorrent that Jim Keller denied in June.

• Intellistake (Vancouver) signed a definitive agreement to acquire NanoAi Technologies of Dallas for roughly C$17M in stock, announced August 4.

  • NanoAi makes portable devices that detect airborne threats at standoff distance; Intellistake plans to fuse the sensor data with its AI data infrastructure for defence, law enforcement, and healthcare customers.

  • The CSE-listed company is pitching data centres located outside CLOUD Act jurisdiction as part of the offer. The shares carry performance-based vesting milestones.

• TetraGen Robotics (Winnipeg) raised a $1.8M oversubscribed seed round led by Calgary’s StepChange Capital.

  • Founder Matt Khoshdarregi previously ran the University of Manitoba’s Intelligent Digital Manufacturing Lab. The robots pair 3D vision with AI software to weld without manual programming; Vancouver’s Emend Vision Fund and Jim Richardson of James Richardson & Sons also participated.

  • Good to see an oversubscribed robotics round land in Winnipeg.

• Transformer Lab (Kitchener-Waterloo) launched Primus, an autonomous research tool, on August 10.

  • The four-person company, founded by Well.ca and Tulip founder Ali Asaria, says Primus produced 30 research papers in 30 days, one of which was cited by Google DeepMind. Access is free to start and rolling out from a waitlist as compute allows.

• Calogy (Sherbrooke) won a $1.2M contract to design the battery and thermal system for Jaunt Air Mobility’s next-generation autonomous aircraft.

  • Calogy’s contribution is the thermal hardware; the autonomy stack is Jaunt’s.

• Claryx raised a US$3.5M pre-seed round led by New York’s Outlander VC to detect hospital pathogen outbreaks before they spread.

  • The CloneLink platform sequences pathogen genomes from confirmed infections and links scattered cases into outbreaks. Founders Dirk and Kurt Hackenberger are Ontarians who built the company in New York, citing simpler sales into private hospital systems.

🏢 Large Companies

• Cohere (Toronto) announced a standalone APAC headquarters in Seoul on August 5, with subsidiaries in Korea and Japan planned by year-end. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • The company plans to double its APAC technical and sales headcount by year-end and grow the regional business fourfold by the end of 2027. The anchor partner is LG CNS, building Korean-language, finance-specialized agents on Cohere’s North platform.

  • Cohere opened its first Seoul office in July 2025; this upgrades the city to regional headquarters. Korean coverage also cited a 1-trillion-parameter Command A Plus model planned for the second half of 2026.

• Shopify (Ottawa) reported Q2 revenue of US$3.58B, up 34%, with GMV up 32% to US$115.6B, on August 5.

  • On the earnings call, the company said AI-driven traffic and orders to Shopify stores each tripled year over year, and Sidekick handled about 34 million conversations with daily active use up 3.6x. Shares rose about 18% on the day.

• TELUS (Vancouver) took a $2.1B pre-tax impairment on TELUS Digital in its July 31 Q2 results, driven partly by hyperscale clients using AI to automate the legacy services the unit sells.

  • The same report raised full-year capex to about $2.6B, with part of the increase going to its sovereign AI data centres. The Rimouski facility is sold out, and the Kamloops site is slated to come online later this year.

🔬 Research

• Luis Mantilla Calderon and Alán Aspuru-Guzik of the University of Toronto and the Vector Institute posted “How Molecular Generative Models Organize Molecular Identity” (arXiv, August 7).

  • The paper analyzes how three generative architectures internally arrange discrete chemical identities, an interpretability question for AI-driven molecule design.

🏛️ Policy

• Alberta launched a public website this week making the province’s case for AI data centres, built around one line: “Think of them as digital refineries.”

  • The chain it describes: natural gas becomes electricity, electricity becomes computing, computing becomes intelligence. Citing energy economist Peter Tertzakian, the site claims the chain multiplies the value of a gigajoule of gas about 1,000 times, versus roughly 4x for refining oil into kerosene and 9x for turning gas into fertilizer.

  • Four commitments anchor the pitch: no public subsidies, grid reliability first, licensed water use, and a levy of up to 2% on data-centre computing equipment. The featured project is Meta’s $13B Sturgeon County campus, announced July 8.

  • It is the most coherent case any Canadian government has made for hosting the buildout.

• Mode40 (Winnipeg) and Canadian Manufacturers & Exporters launched AeroTrace, a Prairies-wide program to demonstrate AI systems in aerospace, aviation, and defence manufacturing, announced August 4.

  • Nine manufacturers, three each from Manitoba, Saskatchewan, and Alberta, will run funded AI implementation projects, with each project financed one-third by Mode40, one-third by CME through third-party funding, and one-third by the manufacturer.

📊 Data

• Statistics Canada reported that June imports of computers and peripherals jumped 59% to a record high, in trade data released August 4.

  • StatCan attributes the surge to processing units “of the type used in data centres” coming from the US, and puts first-half imports in the category up 36.7% over the same period in 2025. The Globe and Mail’s Jason Kirby read the data as the first hard evidence the Canadian data-centre buildout is underway.

📰 In brief

• Sanja Fidler, the U of T professor and Vector Institute co-founder, is leaving Nvidia after eight years as VP of AI research, announced in a July 31 LinkedIn post. She built Nvidia’s Toronto lab from employee one to more than 80 people and says world models are “where the next breakthrough lies.” Her next move is undisclosed.

• BetaKit asked whether the AI “techlash” has reached Canada, citing data-centre protest posters in Calgary and a twice-vandalized automated bakery in Montréal.

• Max Power is using an AI geological platform to scale exploration at its Lawson natural-hydrogen site in southern Saskatchewan, after drilling Canada’s first well dedicated to natural hydrogen.

• Mila hosted its second Indigenous AI Gathering on August 5, with more than 45 speakers and student-built projects from its Indigenous Pathfinders in AI program.

• SAAS NORTH AI and Dominion Dynamics launched Deploy, a new Ottawa summit for founders expanding into dual-use and defence markets.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

As promised in last week’s H1 report, this is the catch-up issue covering July. The regular weekly cadence resumes next Tuesday.

Was this forwarded to you? Subscribe below.

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

The big story in July is Skyfall AI: the Maluuba founders reunited roughly 25 researchers across Toronto and San Francisco, raised an undisclosed round from Fidelity and Inovia, and plan to buy a SaaS business and run it with an AI as CEO. Also in the window: Superhuman acquired GPTZero, Montréal’s Stathera raised US$55M for AI data centre chips, and a freshly minted Fields Medalist is leaving U of T for OpenAI. Ottawa wrote its first big AI for All cheques, and BC retained counsel to pursue OpenAI. Plus a record Canadian showing at ICML in Seoul. Let’s get into it.

💰 Deals & Milestones

• Skyfall AI (Toronto and San Francisco) emerged from stealth on July 20 with an undisclosed round from Fidelity, Inovia, Touring Capital, M13, NextView, and Garage Capital.

  • Founders Sam Pasupalak, Kaheer Suleman, and Sumit Pasupalak previously built Maluuba, the Waterloo NLP company Microsoft acquired in 2017. The new team is about 25 researchers drawn from U of T, Waterloo, the University of Alberta, and Mila.

  • The thesis is “Enterprise World Models” aimed at an autonomous enterprise. The first initiative: acquire a SaaS business for up to US$1M and run it with an AI as CEO.

• Superhuman, the rebranded Grammarly parent, agreed to acquire GPTZero on June 23.

  • The AI-detection company reached 19M+ registered users and roughly $30M in ARR on just $13.5M raised. Both co-founders join Superhuman.

  • Founders Edward Tian and Alex Cui grew up in Toronto; the company itself was never headquartered in Canada.

• Stathera (Montréal) raised a US$55M Series B on June 30 led by Maverick Silicon.

  • The fabless semiconductor company makes silicon timing chips for AI data centres. Celesta, BDC Capital, MediaTek Innovation Fund, TXC, and Ultratech participated, bringing total funding to US$75M.

• A10 Networks acquired TrojAI (Saint John) on June 15; terms were not disclosed.

  • TrojAI built red-teaming and runtime protection for enterprise AI systems. The 22-person team stays in Saint John.

  • Good to see an AI security team staying put in Atlantic Canada after an exit.

• Superwhisper (Toronto) is adopting Cohere‘s open-source Transcribe model as its dictation engine, announced July 29. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • Transcribe is a 2B-parameter open-source speech model averaging 5.35% word error rate across 14 languages, runs fully offline, and currently tops Hugging Face’s Open ASR Leaderboard. It will eventually replace Nvidia’s Parakeet as the app’s default.

  • The pairing puts two Toronto companies across the app’s entire speech stack.

🏢 Large Companies

• Shopify (Ottawa) shipped a Hydrogen developer preview on July 8 that lets storefronts expose actions to in-browser AI agents.

  • The preview adds WebMCP tools for catalog search, browsing, cart, and checkout, alongside Storefront API caching, a Customer Account API, and predictive search.

• CIBC launched AdvisorAssist and CAI 2.0 nationally on July 31.

  • AdvisorAssist is a proprietary genAI tool for advisors with automated meeting notes; the bank claims up to a 50% reduction in admin time. The release also cites DocuMind processing 63,000 documents a month, saving about 16,000 hours a quarter.

  • Bloomberg reports the most powerful tier is limited to roughly 250 handpicked staff over cost.

• RBC renewed its platinum sponsorship of the Vector Institute through 2032 on July 30.

  • The release cites 30 applied AI projects with Vector to date, 200 Vector-affiliated hires, and a #1-in-Canada AI-maturity ranking.

• 1Password (Toronto) shipped two AI launches in July: AI Spend and Consumption Management on July 14 and 1Password for Claude with Anthropic on July 16.

  • The spend tool normalizes AI token spend across providers including Anthropic, Cursor, and OpenAI. It is in public preview, with general availability planned for fall 2026.

  • The Claude integration, built around an “Agentic Mode” zero-exposure framework, lets Claude use stored credentials in the browser without secrets ever reaching the model.

• Cohere and the University of Toronto signed a multi-year partnership on July 16. (Disclosure: portfolio company.)

  • Cohere’s North platform becomes the orchestration layer of U of T’s enterprise-wide AI platform, framed around responsible, sovereign AI adoption at scale.

• Clio (Burnaby) launched Clio Work in Canada on July 23.

  • The AI legal work platform handles case summarization, case-law surfacing, and drafting. It is built on the Jurisage acquisition from June, which brought 470,000+ Canadian cases across 40+ courts, updated daily.

• Tenstorrent (Toronto) CEO Jim Keller shot down the Qualcomm acquisition rumour at a Tokyo media event on June 30.

  • Keller denied the reported US$8-10B Qualcomm interest that surfaced in June. No transaction was ever confirmed by either company.

🔬 Research

• Canadian institutes posted record numbers at ICML 2026 in Seoul, July 6-11.

  • Mila had 82 main-conference papers and Vector had 73 accepted, including 11 spotlights, its strongest ICML showing ever. The conference itself set records with 23,918 submissions at a 26.6% acceptance rate.

• A paper with Canadian senior authors took an ICML Outstanding Paper Honorable Mention.

  • “Motion Attribution for Video Generation” traces which training examples most improve motion quality in video models, matching quality with roughly a tenth of the data. Senior authors are Sanja Fidler, Jun Gao, and Jonathan Lorraine of U of T, Vector, and NVIDIA Toronto; the lead author is at Princeton.

• Doina Precup’s group at McGill/Mila posted Adaptive Multi-Horizon Reinforcement Learning (arXiv, July 22).

  • The method lets RL agents adaptively blend multiple time horizons, balancing short- and long-term reward without hand-tuning.

• Aaron Courville’s group at Mila/UdeM/CIFAR recast quantum wavefunction optimization as reinforcement learning (arXiv, July 2; ICML main conference).

  • “Proximal Wavefunction Optimization” treats variational quantum-state search as policy-gradient RL and scales it to a 1.5B-parameter model.

🏛️ Policy

• Ottawa put $100M into the VITAL health-data platform on June 23.

  • The first major AI for All implementation cheque connects 160 hospitals across Ontario, Québec, and Alberta.

• AI Minister Evan Solomon floated Ottawa leading AI investment rounds on The BetaKit Podcast, July 1.

  • The $500M Canadian Tech Growth Fund could lead rounds with debt and equity optionality. The Council of Canadian Innovators’ Laurent Carbonneau pushed back on the missing detail.

• Solomon announced PrairiesCan funding for AI adoption in Manitoba at the Winnipeg Chamber of Commerce on July 6.

  • The federal money targets Manitoba businesses adopting AI and digital technologies.

• BC retained counsel to pursue legal action against OpenAI over Tumbler Ridge on July 7.

  • Attorney General Niki Sharma said the province is “preparing legal action,” with legal teams retained in both BC and California. No provincial claim has been filed yet.

  • The affected families’ civil suits were filed in late April, in BC Supreme Court and San Francisco federal court.

• Canada Economic Development invested $13.85M across 63 Québec AI projects on July 15.

  • The money flows through the Regional Artificial Intelligence Initiative, explicitly under AI for All, and runs to March 2031. The announcement was made in Trois-Rivières.

📰 In brief

• Jacob Tsimerman, who on July 23 became the second Canadian ever to win the Fields Medal, announced he is joining OpenAI in San Francisco to work on AI safety, which he called “the most important issue of our time.” He is taking leave from U of T while keeping his faculty post.

• Shopify ran a recruiting-scale presence at ICML in Seoul, touting 40M LLM inferences a day and $1.5B+ in R&D spend.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

After a few weeks on vacation, I am back with a special edition recapping the first half of 2026. The regular weekly cadence returns next Tuesday, starting with a catch-up issue covering July.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

For years the Canadian AI story had a familiar shape: world-class research, weak commercialization, and a talent pipeline that ran mostly south. In the first half of 2026, that shape broke. The money showed up, in funds, in unicorn rounds, in federal equity cheques. The concrete got poured, in Merritt, in Kamloops, in a fight over a gas plant in Olds. The argument changed too. Nobody serious is still asking whether Canada can compete in AI. The question that defined this half is who captures the value, and for the first time, Ottawa, the provinces, and a surprising number of founders are answering it out loud. This special issue is the first half of 2026, distilled. Let’s get into it.

📊 The half in numbers

• US$890M: the SCIP sovereign compute procurement, opened April 15, closed June 1, 160 proposals in. No winner named yet.

• ~$2.86B: the genuinely new money in “AI for All,” Canada’s national AI strategy, out of ~$8.5B in named totals. [→ Issue 17]

• ~$20B: Cohere’s merger with Aleph Alpha, with the HQ staying in Toronto. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.) [→ Issue 11]

• ~$1.4B: the valuation of Toronto’s Beacon after its US$225M Series C, the half’s marquee AI unicorn. [→ Issue 18]

• 54%: AI’s share of all venture capital invested in Canada in 2025, on 23.6% of financings, per Osler’s Deal Points Report. Median AI valuations ran ~60% above the rest of the market.

• 93% and 2%: the share of Canadian organizations using or piloting AI, and the share that can prove measurable ROI. Hold that gap in your head; the whole half lives inside it.

🏗️ Sovereign compute went from MOU to megawatts

The half opened with memoranda and closed with purchase orders. In February, sovereign compute was a call for proposals and a set of federal MOUs with no dollars attached. By June 1, SCIP, the up-to-$890M flagship procurement, had opened, drawn 160 proposals, and closed, with a Queen’s-led consortium (ex-Nvidia supercomputing talent, Simon Fraser and Bell as partners) among the named bidders. That is a procurement cycle measured in weeks, from a government not famous for them.

The private market moved faster still. Telus committed roughly $1B to a BC cluster scaling toward 60,000 GPUs by 2032 (and no, despite persistent rumours, that is not a SCIP award; it never was). Bell put $1.7B behind Canada’s largest AI data centre near Regina. And in the half’s most concrete answer to “where does Canadian AI actually run,” Bell AI Fabric, Cohere, Hypertec and BUZZ HPC signed a US$220M deal to run Cohere’s models on 2,304 NVIDIA Grace Blackwells at a 6.5 MW facility in Merritt, BC. Canadian models, Canadian-owned compute, Canadian power, Canadian jurisdiction.

The constraint has migrated accordingly. It is no longer policy will; it is electrons. Alberta spent the half running the two futures side by side: a coherent provincial strategy (the U of A compute vision, a Health Innovation Lab) next to a regulator that rejected the 1.4 GW power plant meant to feed Canada’s largest proposed AI data centre at Olds. Manitoba rejected a hyperscale build outright. Kevin O’Leary’s Wonder Valley ended the half-stalled. The pattern is unmistakable: capital is ready, grids and communities are the gate.

🚀 Cohere’s breakout half

Cohere entered the half with a leaked memo and exited it as the organizing fact of Canadian AI. In February, CNBC reported roughly US$240M in annual recurring revenue, past its own target. In April came the defining move: a merger with Germany’s Aleph Alpha at a roughly US$20B valuation, anchored by a US$600M Schwarz Group-led Series E, with the headquarters staying in Toronto. For scale, this is a company that raised at US$6.8B last August. The keep-it-Canadian debate got its proof case in a single term sheet.

Around that anchor, the half was a study in flag-planting. An open-source cadence that ran nearly monthly: Tiny Aya in 70+ languages, Command A+ under Apache 2.0 pitched squarely against China-anchored open models, the co/plot data-viz tool, and North Mini Code, its first agentic coding model. The Reliant AI acquisition took North into pharma. MOUs with Spain’s Indra and Multiverse Computing landed during a royal state visit, and a first provincial MOU with Québec followed in June. Ottawa’s EY Centre became the Cohere Centre weeks before hosting CANSEC, where the Calian partnership carried North into controlled defence environments. A London office tripled the UK footprint, timing that looked prescient within days, when US export controls on a competitor sent allied buyers looking for alternatives (more on that below). The open question H1 hands to H2 is the one everyone asks me at events: when does this company go public?

🏛️ Ottawa wrote the rules and the cheques

The saga that ran through the spring, with the national strategy “coming very soon” for three straight issues and a draft leaking to CBC, finally resolved on June 4 when Prime Minister Carney and AI Minister Evan Solomon launched “AI for All” at Toronto General. Strip out the re-announcements and top-ups (~$8.5B in named totals) and the genuinely new money is roughly $2.86B: a $500M Canadian Tech Growth Fund that takes equity rather than writing grants, $200M for AI Missions starting with health, $50M for CAISI, and $130M to commercialize the national institutes. The stated ambition is to move business AI adoption from 12% to 60% by 2034. My take from launch week still holds: adoption is not value capture, and the strategy is still light on hard industrial policy. But a federal government taking equity positions in its own AI sector is a genuine posture change, and by July, Solomon was floating Ottawa as a lead investor in AI rounds.

The regulatory engine turned over in the same weeks. The privacy commissioners ruled that OpenAI violated PIPEDA, establishing that retroactive consent is not consent. [→ Issue 13] Then two bills landed in June: C-34, which puts safety duties directly on chatbot operators, and C-36, which grants Canadians the right to force deletion of AI deepfakes. [→ Issue 18] Notably absent: any revival of an AIDA-style horizontal AI law. Ottawa is regulating AI through the harms, not the technology.

And the implementation money started moving before the half was out: $100M into the VITAL health-data platform on June 23, connecting 160 hospitals across three provinces as the first big AI for All cheque.

💰 The capital came in

Follow the sequence. Budget 2025 put $1.75B behind Canadian venture, four times the previous VCCI allocation. [→ Issue 7] BDC opened a $500M LIFT window for SME AI adoption. [→ Issue 11] Mila and Inovia launched the US$100M Venture Scientist Fund in January to turn institute research into companies. May brought three funds in one week, then five AI-mandate funds in fourteen days, Sagard’s $150M USD AI Fund among them. Inovia closed a $365M third growth fund in June. Then the marks arrived: Beacon’s US$225M Series C minted a ~$1.4B unicorn built explicitly as an AI rollup.

The empirical anchor is Osler’s data: AI took 23.6% of financings and 54% of all capital invested in 2025, with a ~60% valuation premium and rounds ~50% larger. Fund formation and late-stage marks inflected in the same half. That combination is what makes this a capital story rather than a grants story, and it is why the “Canada can’t fund its own AI companies” line got noticeably quieter by June.

🛡️ The defence-AI map went national

What started the half as an Ottawa story ended it as a national map: Dominion Dynamics in Ottawa, PacifiCan money on Vancouver Island, Lastwall in Fredericton, Kraken’s $615M Covelya acquisition in St. John’s, Denvr in Calgary, and a corridor of autonomy work through Kamloops and North Bay. CANSEC in late May was the set piece. Carney became the first sitting PM to keynote the show, pledged $180B in defence procurement over ten years, and named “unimpeded access to artificial intelligence” a strategic capability alongside space and quantum. The venue itself made the point: the show ran at the freshly renamed Cohere Centre, and the launches skewed hard toward sovereignty: Volatus’s V-Cortex autonomy stack, Convergence’s MIL-V built by 13 firms on 100% Canadian IP, Calian’s ATHORA layer wired for Cohere’s North.

The counterpoint arrived the same week, via documents tabled in Parliament: a $3.7M DND contract with Palantir, defended by the minister as legitimate procurement. A show themed on keeping AI Canadian, and a US data giant quietly inside the tent. Both things are the real state of play.

🏦 The banks raced, the economy lagged

Four of the Big Five now carry $1B-class AI commitments: RBC (targeting up to $1B in value by 2027), TD ($1B by 2028), BMO, and Scotiabank with Scotia Intelligence across 71,000 employees. CIBC ended the half as the lone holdout. Set that institutional sprint against the economy-wide numbers and the half’s most important tension appears: KPMG found 93% of Canadian organizations using or piloting AI, up from 61% a year earlier, while only 2% could prove measurable ROI. MNP’s version: 91% satisfied, 4% transformational. StatCan’s sober floor: 12.2% business adoption in 2025. Canada also ranked 42nd of 47 countries in AI trust.

The gap between what institutions are spending and what the economy can measure is the single most important unresolved question of the half. It is also, bluntly, where the next wave of enterprise AI value gets created.

🍁 The idea that organized the half: keep it Canadian

Every thread above braids into one idea, and the half gave it a vocabulary. Solomon’s parliamentary secretary Taleeb Noormohamed said Canadians have watched IP and long-term value migrate elsewhere and that “our job is to make sure that stops.” Waabi’s Raquel Urtasun said the company must remain Canadian. Tobi Lütke’s quieter point at Toronto Tech Week was about making sure there is something durable worth keeping. Cohere retaining its Toronto HQ through a ~$20B merger is the proof case; Giesecke+Devrient planting its global AI hub inside Mila is the productive tension, foreign industrial capital choosing to sit inside the flagship institute rather than hollow it out.

Then June made the whole argument literal. Washington’s export controls disabled Anthropic’s Fable 5 and Mythos 5 models in several allied markets, and Cohere reported a surge of inbound from US allies within days. Nick Frosst’s pre-ban line to BetaKit read like prophecy a week later: if your whole stack comes from one country that can switch you off, that is not a foundation you can build on. Sovereignty spent years as a values argument. In June it became a procurement criterion, on the record, with a case study.

⚖️ The honest ledger

A half this good deserves an honest accounting of what did not go right.

• The adoption gap. 93/2. Everyone is experimenting; almost nobody can prove it is working. Until that closes, the productivity promise is a promissory note.

• The exits still point south. Ottawa’s Hyperlume sold to Credo for US$92M in cash seven months after its seed, extending the chip lineage of CentML, Tenstorrent and Untether. Hyper went stealth-to-Motorola in under a year. TrojAI of Saint John sold to A10 Networks in June. The talent stays, for now. The cap table leaves. And in the purest form of the pattern, Superhuman acquired GPTZero, built by two Toronto-raised founders to $30M ARR on $13.5M raised, a company that was never Canadian to begin with because it never felt a reason to be.

• Sovereignty is partial. The compute is increasingly Canadian-owned and Canadian-powered. The silicon inside it is essentially 100% NVIDIA, and Ottawa said plainly this half that there will be no national semiconductor strategy. Sovereign racks, imported engines.

• The bubble question. The honest answer from the half’s best debates: there is a bubble, and there is a generational platform shift, and both are true at once. Position accordingly.

🔭 Early H2 signals

The second half is already moving. Since the last regular issue: Ottawa signed its first big AI for All implementation cheques (Manitoba adoption funding, $13.85M across 63 Québec AI projects). BC retained counsel to pursue OpenAI over Tumbler Ridge. Jim Keller shot down the Qualcomm rumour. Mila and Vector posted record showings at ICML. And the ex-Maluuba team resurfaced in Toronto as Skyfall AI, with a plan to buy a company and run it with AI as CEO. The full month gets its own issue next Tuesday.

Watch these into the fall: the SCIP winner (still unnamed). Whether CIBC finally moves. The Cohere Series E close and the IPO question behind it. The first Canadian Tech Growth Fund cheque, and whether Ottawa really leads a round. C-34 and C-36 through committee, and the promised standalone AI bill. The Synapse Olds decision, round two.

The takeaway

The first half of 2026 settled the question Canada has been asking itself for a decade: the capital shows up. Funds formed, unicorns minted, procurement moved, and a government started buying and owning instead of granting and hoping. The second half asks the harder question, the one sitting inside that 93/2 gap and those southbound cap tables: whether the value built here stays here, and whether anyone can prove it is working. That is not a policy question. That is an operating question, and it belongs to the founders and operators reading this. See you Tuesday.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

The sovereign-compute conversation finally moved from press releases to procurement this week. Bell, Cohere, Hypertec, and BUZZ HPC put US$220M behind 2,304 NVIDIA Grace Blackwells in Merritt, BC, the most concrete answer yet to the question of where Canadian AI actually runs. In the same seven days, Shopify made AI shopping agents the default checkout path with its Spring ‘26 edition, Cohere tripled its UK footprint just before US export controls on Anthropic created what its team is calling “huge inbound,” and Clio paired a sixty-five-day legal-AI training blitz with a quiet Canadian-data acquisition that gives away the strategy. Around it: an $8-10B Qualcomm rumour around Toronto’s Tenstorrent, two same-day robotics partnerships out of Montréal’s Vention, and a New Brunswick crop-spectroscopy seed round that keeps making the case for AI beyond the Toronto-Montréal corridor. Let’s get into it.

💰 Deals & Milestones

• Picketa Systems (Fredericton) closed a $2.1M seed round led by Tall Grass Ventures to put ML-driven crop nutrient sensing into farmers’ hands.

  • The company’s LENS device uses spectroscopy and machine learning to translate light readings into nutrient concentrations in seconds, on the field, with no lab turnaround. Today’s standard is to mail a tissue sample and wait days for a result that may already be stale.

  • Atlantic Canada agtech with a real ML stack and a hardware moat. Geographic diversity outside Toronto and Montréal is one of the under-priced edges in Canadian venture.

• Reveal Life Science (Montréal) won the OVHcloud Startup Challenge at VivaTech in Paris.

  • The surgical-AI startup uses Raman spectroscopy combined with large molecular datasets to identify cancerous tissue intra-operatively in seconds, with data hosted on Canadian infrastructure via Qohash.

• Vention (Montréal) announced two physical-AI partnerships on the same day at Automate 2026 in Chicago.

  • The first brings FANUC America’s industrial robots onto Vention’s AI-driven hardware and software stack, including AI programming, digital twin, and collision-free path planning on NVIDIA Isaac. The second is a digital-twin collaboration with Teradyne Robotics optimized for Universal Robots cobots, with AI bin-picking via Rapid Operator AI.

  • These are separate from Vention’s Series D from earlier in the year and should not be conflated. Read together, Vention is positioning itself as the AI control plane that sits above a fragmented robotics hardware market, not as a vendor competing inside it. That is a smarter wedge.

• BrokerPlus (Toronto) launched out of stealth with more than CAD $500K in pre-seed funding to bring AI to the mortgage broker channel.

  • Backers include Michael Hyatt’s family office, Cowie Capital, Tennr’s Trey Holterman, and Bet99’s Jared Beber. The product uses AI to mine broker management systems for missed renewal and refinance opportunities within brokerages’ own client databases. Founders are CEO Swish Goswami, previously of Surf, and CTO Amir Agassi.

🏢 Large Companies

• Bell AI Fabric, Cohere, Hypertec, and BUZZ HPC announced a landmark sovereign-AI compute deal to host Cohere’s foundation models on fully Canadian infrastructure. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • The numbers: 2,304 NVIDIA Grace Blackwell GPUs (GB200 NVL72, liquid-cooled, InfiniBand), a USD $220M three-year commitment routed through HIVE’s BUZZ HPC subsidiary, hosted at Bell’s 6.5 MW facility in Merritt, BC, with delivery targeted for late 2026 into early 2027.

  • A Canadian foundation-model company running its own models on Canadian-owned compute, sitting on Canadian power, under Canadian jurisdiction, with one of the country’s largest carriers operating the fabric.

  • Worth a flag: this is a private-sector deal, not a Sovereign AI Compute Infrastructure (SCIP) grant. SCIP is federal money trying to bootstrap supply; this is private demand finally showing up to meet it. The SCIP grants will come in the months ahead.

  • Also distinct from Bell’s larger 300 MW Saskatchewan project announced earlier this year. Merritt is the one with the customer attached.

• Shopify (Ottawa) shipped its Spring ‘26 Edition with 150+ updates that make agentic commerce the default.

  • The headline change is the Universal Commerce Protocol (UCP), now on by default across every Shopify store, which lets AI shopping agents like ChatGPT, Copilot, and Perplexity read catalogs, build carts, and complete checkout on the buyer’s behalf. There is also a new “Agentic” admin section in every store, an expanded Sidekick that now reaches into the admin and POS, and a stack of merchant-side AI tooling.

• Cohere (Toronto) tripled its UK footprint with a new ~14,000 sq ft office at 100 New Oxford Street in London, capacity for up to 100 people.

  • Aidan Gomez framed the move squarely around the UK’s sovereign-AI push. Read together with the Merritt deal and the inbound interest in the In Brief section below, the picture is clear: every Western democracy that is uncomfortable with depending on a single American or Chinese lab is now a Cohere prospect, and the company is putting boots on the ground in those with the largest budgets.

• TELUS Digital (Vancouver) was named a preferred implementation partner by ElevenLabs for ElevenAgents, its enterprise voice-agent platform.

  • TELUS Digital takes the lead on implementation, integration, governance, and managed services for enterprise CX customers deploying ElevenAgents. The release cites more than 90,000 simulations run through its Fuel iX Agent Trainer and a 20% reduction in onboarding time for live agents trained alongside AI.

  • The hidden story in enterprise voice AI is that the model is rarely the bottleneck; integration into a customer’s CRM, telephony, compliance, and ops is. ElevenLabs is the best voice model team in the world. They do not want to be a service business. TELUS Digital has the implementation muscle, the CX accounts, and an existing AI platform to plug into.

• Tenstorrent (Toronto) is reportedly in advanced acquisition talks with Qualcomm, at a rumoured price of US$8-10B.

  • The Information broke it; Reuters echoed the outline. No party has confirmed and the deal is not closed. If the reporting is accurate, the price is roughly a 4x premium on Tenstorrent’s end-2025 ~$2.6B mark and would be one of the largest acquisitions ever for a Toronto-headquartered company. Jim Keller, the AI silicon legend who runs Tenstorrent, would in this scenario land inside Qualcomm.

  • Watch Qualcomm’s June 24 Investor Day. If true, the Canadian AI hardware story instantly looks different: the open-source RISC-V challenger to NVIDIA is suddenly an arm of a US public chipmaker, and the talent and IP question becomes whether Tenstorrent’s Toronto engineering team stays Toronto-based or gets gradually absorbed into a San Diego org chart.

• Clio (Burnaby) launched the Legal AI Accelerator, its largest-ever legal-AI training push, aimed at certifying 25,000 legal professionals by ClioCon in Boston this October.

  • The Accelerator opened June 15 and runs 65 days of guided practice, CLE-eligible sessions, three no-cost certifications, and a tour of Clio Connects events across multiple US cities. Free to participate.

  • The training blitz lands days after Clio’s June 10 acquisition of Canadian legal-data company Jurisage (470,000+ Canadian cases across 40+ courts, updated daily), which clears the way for Clio Work to launch in Canada later this year. Read together, the strategy is obvious, and it is the right one. In vertical AI, proprietary data is one of the few durable moats against the big labs. Clio is not trying to out-scale frontier model builders; it is locking up the trusted, jurisdiction-specific legal data and the practitioner workflows that general-purpose models cannot easily replicate, then training the entire market to run on its platform.

🔬 Research

• Giesecke+Devrient launched its AI Hub embedded inside Mila, positioning Montréal as the heart of its global AI Center of Excellence.

  • The German SecurityTech group will invest $80M over five years to develop applied AI for cybersecurity, fintech, and digital identity from inside Mila’s Montréal facility. Another global institution choosing to put applied-AI work next to Yoshua Bengio’s lab instead of inside its own corporate R&D centre. The Mila gravity well keeps working.

🏛️ Policy

• Alberta Innovates committed $14M to commercialize local technology, with AI startups among the beneficiaries.

  • The provincial program is general-tech rather than AI-exclusive, so the AI-specific share will be smaller than the headline number suggests, but the direction matters. Alberta keeps showing up on the deployment side of the AI strategy.

• MLT Aikins published Volume Two of its AI data centre legal and regulatory guide covering the physical side of building data centres in Canada.

  • The guide walks through corporate structuring, financing, Investment Canada Act and foreign-investment review, energy and grid considerations, land use, environmental permitting, and Indigenous consultation. A useful operator’s-eye companion. If you are thinking about where the next Merritt-class facility gets built, this is the playbook for what stops it.

• James Moore argued in a CTV op-ed that healthcare, and specifically rare-disease diagnosis, is Canada’s largest under-claimed AI opportunity.

  • The former federal industry minister set the argument against the national AI strategy: Canada has the data, the publicly funded health system, and the research base; what it lacks is the procurement runway to put diagnostic AI into clinics at scale. Worth reading not for the conclusion, which is familiar, but for the framing from someone who has actually sat in the cabinet seat where these decisions get made.

📊 Data

• National Bank of Canada / Leger released a survey on AI’s role in Canadian purchasing on June 19.

  • 39% of Canadians say they used generative AI to support a purchase in the past year, 61% say AI now sways them more than advertising, and Quebec leads adoption at 45% (Ontario at 41%, Manitoba and Saskatchewan tied at the bottom at 28%).

  • Stitch this to Shopify’s Spring ‘26 release above. If 61% of Canadian buyers already trust AI more than ads, and the world’s largest commerce platform just made agentic checkout the default, that is supply meeting demand on the same week.

📰 In brief

• Cohere is reporting huge inbound from US allies after Washington’s export controls on Anthropic disabled the Fable 5 and Mythos 5 models in several allied markets. Gomez and President Joelle Pineau told Bloomberg the diversification calls have picked up sharply; Prime Minister Carney referenced the same dependency dynamic in a separate appearance the same week.

• Cohere’s Nick Frosst sat down with BetaKit for a wide-ranging interview on sovereign AI recorded during Toronto Tech Week, before the Anthropic export news broke. The line that lands: if your whole stack comes from one country that can switch you off, that is not a foundation you can build on.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

Two themes this week: money and rules. On the money side, two new Canadian unicorns in the same seven days, both with AI as the centrepiece of their pitch, Toronto’s Beacon at a $1.4B valuation and Montréal’s nesto at $1.47B, with a Calgary-rooted manufacturing-vision raise and a tidy Toronto fintech pre-seed alongside. On the rules side, Ottawa tabled the first Canadian legislation to regulate AI head-on: one bill that puts safety duties directly on chatbot operators, another that gives you the right to make a company delete an AI deepfake of you. Add a sobering warning from a Toronto lab about what open-weight models can do in the wrong hands, and a rare look inside the GTA building where one of Canada’s most powerful supercomputers actually lives. Let’s get into it.

💰 Deals & Milestones

• Beacon (Toronto) closed a US$225M Series C that values the AI rollup at roughly $1.4B.

  • General Catalyst and HarbourVest led, with Lightspeed and Toronto’s Intrepid Growth Partners. The round is all-equity and all-primary, pushing total funding past US$550M and lifting the valuation from the $1B it carried on its US$250M Series B last November. The deal actually closed in January and was only made public this week.

  • Beacon is the “anti-private-equity” play: launched in 2024 by former Instacart president Nilam Ganenthiran, it buys small, established software companies and rebuilds them on a shared AI operating stack. It is reportedly closing roughly one acquisition a week.

  • AI rollups are a real opportunity, and Beacon is running the Constellation Software playbook with an AI engine bolted on. The thesis is that legacy software businesses are full of margin you can unlock with automation.

  • I met Nilam years ago when we negotiated the deal with Loblaw that brought Instacart to Canada. He is a super smart, all-around good human. He’s pulling in another ex-Instacart superstar, Marc Schaff, previously CTO at Instacart - getting the band back together!

• nesto (Montréal) raised a CAD $302M Series E at a CAD $1.47B valuation.

  • The round splits into $107M primary and $195M secondary, with La Caisse, Fidelity Canada, PICTON and Endeavor Catalyst coming in new, alongside returning backers Portage, Diagram, NAventures, National Bank, Fonds de solidarité FTQ and Fondaction. The company is profitable, with more than $37B in originations this year and over $80B under administration.

  • Will invest in AI to improve employee productivity and client experience.

  • Also launches its Maestro AI platform, which nesto says can underwrite a mortgage deal in under two minutes, work that has historically taken a day or more.

  • The secondary-heavy structure tells you early backers are taking real money off the table, which is a healthy sign for the Canadian fintech cap-table ecosystem.

• Maneva (Toronto) raised a US$27M Series A to put computer vision on the factory floor.

  • US Venture Partners led, with USVP’s Matt Garratt joining the board, and returning investors Bling Capital and Freestyle Capital plus Canadian firms Seguin Ventures and N49P. Total raised is now US$38.4M. CEO Rae Jeong is a former DeepMind engineer who started out as a welder in Alberta; CTO Kelvin Chan is a Magna alum.

  • The product turns a plant’s existing cameras into a real-time system that flags defects, catches safety hazards, and tracks output. Worth a candid flag: it also monitors workers, and Maneva’s own materials pitch up to a 10% productivity lift. That is a workforce-surveillance story as much as a quality-control one, and someone will write about it that way.

  • I’d gotten a demo of Maneva’s tech a while back, and they were already in the future. They’re continuing the trend with this impressive financing round.

• DataBraid (Toronto) secured a US$1.9M pre-seed to kill insurance brokers’ “portal hopping.”

  • The round, which is the company’s first and closed in April, came entirely from Koru Ventures, the venture studio backed by the Ontario Teachers’ Pension Plan. Founder and CEO Nick Romano previously ran Montréal’s Deeplite, the neural-network-optimization startup acquired by STMicroelectronics last year; co-founder and CTO Atif Khan joined from Messagepoint. They built and piloted the product with Welland, Ontario’s Scoop Insurance.

  • DataBraid’s software plugs into broker management systems and propagates data across the fragmented mess of carrier portals brokers juggle, eliminating the manual re-entry.

• Databiomes (Toronto) launched Ctrlvox, its first AI model, for on-device moderation of toxic game chat. (Disclosure: Databiomes is a Mistral Venture Partners portfolio company.)

  • Ctrlvox is a customizable model that runs on players’ existing CPUs with no GPU or cloud inference, now live as an Unreal Engine plug-in on Epic Games’ Fab marketplace. Databiomes says it outperforms Alibaba’s Qwen3Guard, and that the model cost $60 and took seven hours to build on its own nano-language-model platform. Co-founded by Steven Gans (ex-AMD, Intel, IBM) and CTO Tomasz Klempka; $1.2M CAD raised to date.

  • Disclosure: I led their pre-seed round. Databiomes’ tech is truly a magical unlock of LLMs. Build your own custom nano LLM for tens of dollars and deploy it on cheap CPUs and NPUs. I’m obviously a big fan and so happy to see them start opening their platform to different verticals.

🏢 Large Companies

• Cohere (Toronto) open-sourced North Mini Code, its first agentic coding model. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

  • A 30B-total, 3B-active mixture-of-experts model, open weights under Apache 2.0, small enough to run on a single H100, which Nick Frosst demoed on a Mac Studio. Independent testing by Artificial Analysis ranked it among the fastest open-weight models for output speed and ahead of comparable small models in coding.

• Scotiabank (Toronto) expanded Scotia Intelligence, its enterprise AI program, into a new phase.

  • The bank now says it has enabled more than 71,000 employees with assistive AI tools and has 5,500 engineers using AI for coding, with AI use up 30% quarter over quarter. This is an extension of the platform that was first stood up last year, not a new launch.

🔬 Research

• The Vector Institute (Toronto) signed a research MOU with Germany’s Helmholtz Munich.

  • The agreement formalizes joint AI and machine-learning research, researcher exchanges, faculty affiliations, and coordinated Canada-Germany funding, with health AI as the early focus. Vector’s Shaina Raza had already presented at Helmholtz Munich’s HAICON 2026 workshop on June 8 on AI benchmarking.

  • AI ties to Europe continue.

• University of Toronto and Vector researchers demonstrated an AI-powered “worm” that can target any online device.

  • Led by U of T’s Nicolas Papernot, a Canada CIFAR AI Chair at Vector, the team showed that free, open-weight models can power a self-adapting worm that changes its attack strategy as it moves from device to device, hijacking each machine’s compute and exploiting its specific weaknesses. The work was built and tested in a secure, internet-isolated lab, and the team flagged it deliberately, to get ahead of the threat.

🏛️ Policy

• Bill C-34, the Safe Social Media Act, would create a Digital Safety Act and is among the first Canadian statutes to put safety duties squarely on AI chatbot operators.

  • Per Osler’s read, operators of “chatbot services” would owe a duty to act responsibly: emergency measures that interrupt and redirect a user who expresses suicidal ideation, a ban on the bot posing as a human or as a licensed professional, limits on “manipulative engagement techniques” designed to foster unhealthy attachment, and synthetic-content labelling. Penalties run up to the greater of $20M or 5% of gross global revenue.

  • If you are building anything conversational, a companion app, a support agent, a coaching product, read this bill now, not after it passes. The companion-AI and “manipulative engagement” provisions are pointed directly at the engagement-maximizing design patterns that a lot of consumer AI quietly relies on.

• Bill C-36, the Protecting Privacy and Consumer Data Act, would replace the 28-year-old PIPEDA and reach AI deepfakes directly.

  • Would enshrine privacy as a fundamental right, set a higher bar for children’s data, restrict surveillance pricing, and create a right to deletion that explicitly extends to having AI deepfake images or videos of you removed from commercial platforms. It would also require companies to be transparent about their use of AI, and shift private-sector privacy authority to the new Digital Safety Commission, with fines up to the greater of C$25M or 5% of global revenue.

  • Paired with C-34, this is the post-strategy legislative engine finally turning over. Note what is not here: the old AIDA-style horizontal AI regulation from the failed Bill C-27 is not revived, and Ottawa says dedicated AI rules will travel separately. So the sequencing is privacy and online-harms first, broad AI law later. For founders, the deepfake-deletion right and the AI-transparency duty are the provisions with teeth today.

• Manitoba rejected a hyperscale AI data centre south of Winnipeg.

  • Premier Wab Kinew said a proposed 141-hectare, gas-turbine-powered facility near Île des Chênes, backed by Las Vegas-based Jet.AI and Vancouver’s Consensus Core, will not proceed. (Reported June 4, just before our window, included for the thread.)

  • This is fine, and arguably healthy. If a jurisdiction does not want a data centre, it should not have one, and there will be other places that do. There are many proposals across Canada, many of which will get built. Despite all the demand for data centres, we should resist provincial or municipal subsidies to make the math work. Let the market do its thing. The bigger question beneath it is whether the industry is building compute ahead of demand, as it overbuilt railroads in the 1800s and fibre-optic networks in the 90s. If that is what is happening here, the right posture is to let private capital carry that risk, not to backstop it with taxpayers’ money.

📰 In brief

• 1Password (Toronto) acquired Apono and launched a Credential Broker, extending its identity platform to govern what humans, machines and AI agents can access, when, and for how long.

• DRIVE Hockey (Vancouver) landed a $100K NRC IRAP grant for Coach AI, which turns sensor and “Smart Arena” data into NHL-calibre coaching insights for amateur players.

• Inside Trillium. BetaKit got a rare look inside the GTA building that houses Trillium, the University of Toronto-owned, SciNet-operated supercomputer (close to $200M in hardware, roughly 142nd in the world) that researchers use to simulate oceans and model stars. A neat, rare insider look at one of the most powerful machines in the country, sitting quietly in Toronto. Worth the read, and a useful grounding for the sovereign-compute debate that keeps running through these pages.

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

The wait is over. After slipping past three self-imposed deadlines, Canada’s national AI strategy finally landed on Wednesday, and it is big, expensive, and aimed squarely at one thing: getting Canadians to use AI. I break down below what is actually new money versus re-announced money, and why an adoption-first strategy may be pointed at the wrong target. Around it, a lively week of out-of-Toronto deals from Calgary, Kitchener-Waterloo, and Montréal, two Canadian flags at CVPR in Denver, and fresh data from the Bank of Canada and Angus Reid that should give the adoption thesis pause. Let’s get into it.

💰 Deals & Milestones

• Pricepoint (Montréal) raised a US$4.8M seed to put AI in charge of hotel room pricing.

  • An all-Québec syndicate: Brightspark Ventures led, with Boreal Ventures and AQC Capital. Brightspark’s Sophie Forest and Boreal’s JD Saint-Martin take board seats, and the money roughly doubles a 35-person team.

  • The product is autonomous revenue management, software that sets and resets room rates in real time the way airlines or larger hotel chains have for decades. Pricepoint makes it available for the individual franchisees.

  • I’ve known Mateusz Sznir, founder and CEO, for a few years. He comes from the industry and is super passionate about the problem he’s solving. A true operator and nice human.

• SensorUp (Calgary) closed a growth round led by Pender Ventures to scale agentic AI for heavy industry.

  • Climate Investment and Evok Innovations joined, alongside Occidental as both strategic investor and customer. Pender’s Cheri Corbett joins the board. No dollar figure was disclosed, which is its own tell, but the platform is already in production with five member companies of the Oil and Gas Climate Initiative.

  • Not chatbots, but software agents running emissions monitoring and field operations for the energy sector. A Calgary company selling AI into oil and gas.

• Scispot (Kitchener-Waterloo) raised a US$8M Series A to build the operating layer for AI-run life science labs.

  • Washington’s Avenue Growth Partners led, with Seattle’s Breakwater Ventures returning, taking total funding to about US$10M. Avenue’s Brian Goldsmith joins the board.

  • Scispot calls it the “self-driving lab”: software that wires together instruments, data, and AI so life-sciences experiments run with less human babysitting. Picks and shovels for the biotech-AI boom, and another Waterloo-corridor company worth watching.

• Mecka AI (Toronto talent, New York address) disclosed US$60M to build the data layer for humanoid robots.

  • Framework Ventures led, with Menlo Ventures, SV Angel, and Kindred Ventures alongside. The detail that earns it a place here: three of four founders are Canadian and 40 of 45 staff sit in Toronto, even as the company is headquartered in New York. Framework’s Vance Spencer called it the fastest-growing revenue company the firm has ever backed.

  • A caveat I’ll flag plainly: this is a disclosure of rounds that closed earlier, a US$25M Series A last November plus a follow-on, not a fresh raise this week. I include it because the talent base is unmistakably Canadian, but the cap table and the HQ are not 😔

🏢 Large Companies

• Cohere (Toronto) open-sourced co/plot, a data-visualization tool built inside its research labs and described in a post by Cohere’s Thomas Euyang. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

🔬 Research

CVPR, the year’s biggest computer-vision conference, ran June 3 to 7 in Denver, and Canada planted two flags. Both are collaborations with Apple, with the Canadian angle in the academic authors.

• Mila and Université de Montréal presented SFI-Bench, a benchmark that asks whether multimodal models understand what objects are for, not just where they are.

  • Built from more than 1,700 questions over egocentric indoor video, it tests “spatial-functional intelligence,” the difference between seeing a kettle and knowing you boil water in it. Senior author Aishwarya Agrawal is a Canada CIFAR AI Chair and Mila core member; Le Zhang is the lead author.

  • Why it matters: today’s vision-language models are good at naming and locating, and still clumsy at function and affordance. If you want robots and agents that act in the world, this is the gap that has to close, and a Canadian lab is building the ruler to measure it.

• The University of Toronto presented Velox, a method for learning compact representations of moving 3D scenes, four dimensions counting time.

  • PhD student Anagh Malik and professor David Lindell compress spatiotemporal point clouds into “dynamic tokens” usable for video-to-4D generation, 3D tracking, and even cloth simulation.

  • It is the unglamorous foundation under flashier generative-video demos. Better 4D representations feed straight into robotics perception and world models, the physical-AI thread Raquel Urtasun keeps pounding the table about.

🏛️ Policy

Canada finally has a national AI strategy. On Wednesday, Mark Carney and AI Minister Evan Solomon launched “AI for All” at Toronto General Hospital, closing out a saga that blew through every previous deadline. The full strategy is sprawling. Two things are worth your time: where the money actually is, and what the strategy is actually trying to do.

Follow the money, and separate the new from the recycled. The Globe pegged it at “more than $2.3-billion,” and Carney framed the package as building on roughly $2B already committed. Both are true, and neither is the whole picture. Here is the honest split.

• Genuinely new, strategy-attached money, roughly $2.86B:

  • A $500M Canadian Tech Growth Fund that will take equity stakes in scaling Canadian AI firms. This is the real shift, Ottawa moving from writing grants to taking ownership.

  • A $200M AI Missions program, first mission health.

  • $50M to expand the Canadian AI Safety Institute, and $130M to commercialize research out of Mila, Vector, and Amii.

• Re-announced or topped-up money doing a lot of the headline work:

  • $500M for BDC’s LIFT facility, which BDC already announced back on April 24.

  • A $500M top-up to the existing Regional AI Initiative, and a $700M top-up to the existing Compute Access Fund, the one that has already cut $66M in cheques.

  • Budget 2025’s $1.75B and April’s $25B sovereign wealth fund sit in the backdrop as the capital backstop.

  • Add up the named line items and you reach roughly $8.5B. The genuinely new, strategy-specific commitment is closer to $2.86B. The “$2.3B in new spending” headline holds up only if you count the Tech Growth Fund, Missions, CAISI, and commercialization money as new, and treat the top-ups as fresh. A fair chunk of the rest, we have reported before.

Now the bigger point: this is an adoption strategy. Strip away the funding tables and the organizing goal is to get Canadians and Canadian businesses using AI. The centerpiece target is lifting business AI adoption from just over 12% to 60% by 2034. The gap is real: barely 12% of Canadian firms used AI to make goods or services over the past year, and SMEs sit around 8%, versus 26% in Germany and into the 40s in the Nordics.

My take. Adoption is good. A country that does not use AI never gets the productivity dividend, and the Bank of Canada data below shows how far we have to go. But adoption is not where power and wealth are created. Power and wealth come from building the AI and capturing the value, owning the models, the companies, and the IP, rather than renting someone else’s. There is some value-capture logic buried in here: the Tech Growth Fund taking equity, the money to commercialize institute research. That is the right instinct, but mostly implicit. What I do not see is firm industrial policy, the kind the US and China run without apology, that bends procurement, standards, and market access to favour Canadian companies. We are funding Canadians to use AI built elsewhere and hoping a few champions emerge on their own. The countries we are competing with are writing the rules so their champions win. Not an area we need to be kind in ;)

📊 Data

• The Bank of Canada put a number on the adoption gap the strategy is chasing. A new staff paper by Chanya Chawla and Crystal Arnburg, drawing on the Bank’s December 2025 Business Leaders’ Pulse, finds AI use among business leaders is widespread personally but still limited in actual production.

  • The forward look is sobering for the “AI means jobs” framing: firms expect AI to lift capital spending, but to have a modest net negative effect on employment over the next three years.

• Angus Reid ran two surveys that, read side by side, should temper the adoption optimism.

  • On regulation, Canadians want guardrails: 68% say government should heavily regulate AI even if that slows things down, 74% doubt any government can keep pace, and 68% would oppose a large AI data centre near their home. President Shachi Kurl’s read is that there is a wide gap between where the government is going and where Canadians are.

  • On work, the verdict is a shrug. Among Canadians who use AI on the job, 49% say it has had no impact on their productivity and only 29% say it improved the quality of their work. Just 31% have an employer with an actual AI policy. And the tools people reach for are American: Gemini at 48% and ChatGPT at 47% lead, while Claude sits at 9%. A strategy built on adoption is, in practice, a strategy built on Canadians using other countries’ models.

📰 In brief

• Inovia Capital (Montréal) reported its third growth fund has reached US$365M of a US$450M target. AI-native startups took 40% of Canadian software deal value in 2025. The capital is flowing.

• S&P Global partnered with Cohere to put its financial data inside Cohere’s enterprise platform, North, for on-premise agentic workflows at financial institutions. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

Two big gatherings anchored the week, Toronto Tech Week and CANSEC, and I pulled the AI-related highlights from both. But the issue runs wider than that: a US$28.8M raise for a Montréal fintech, a $16M cyber round out of Fredericton, Rich Sutton making the case that generative AI can’t actually discover anything, fresh Osler data showing AI now drives more than half of all Canadian venture dollars, and a national AI strategy that Mark Carney says is finally landing this week. The thread tying much of it together is a familiar one: keeping Canadian AI Canadian, in who owns the companies, where the IP lives, and who controls the compute. Let’s get into it.

💰 Deals & Milestones

• Saris AI (Montréal) raised a US$28.8M Series A to automate the back offices of banks and credit unions.

  • Led by Joe Lonsdale’s 8VC, with Audacious Ventures, Homebrew, and the Btech Consortium. Founders Danial Jameel, Alice Dinu, and James Dang previously built Montréal’s Oohlala Mobile.

  • The pitch: agentic AI that automates up to 70% of consumer, mortgage, and commercial lending workflows, with integrations into Fiserv and MeridianLink. Jameel frames it as humans and AI “side by side” in financial services.

• Lastwall (Fredericton) raised $16M to defend critical infrastructure, and it’s a genuine East Coast win.

  • Led by BDC Capital’s StrongNorth Fund in what is its first cybersecurity investment, with the New Brunswick Innovation Foundation writing the largest cheque in its history.

  • The product, IDCommand, uses behavioural AI to read more than 200 login characteristics, from keystroke cadence to mouse movement, and flag anomalies before a stolen identity becomes a breach. CEO Karl Holmqvist’s framing: autonomous AI agents make stolen credentials far more dangerous, far faster.

  • Retired Major General Peter Dawe, now BDC’s VP of Defence Strategy, joins the board.

• NordSpace (Markham) led a consortium that landed $3.2M from NGen to build what it calls Canada’s first AI-powered space-propulsion manufacturing line.

  • The full project is valued at over $8M, with the $3.2M coming from Next Generation Manufacturing Canada’s Advanced Manufacturing Technology Program. The output: turbopumps for NordSpace’s Tundra rocket.

  • The consortium is entirely Canadian-owned: NordSpace, Miltera, Pegmatis, Prime Powders, and Indigenous-owned Bear Paw Manufacturing. The AI lives in the process, in-situ quality control and Pegmatis’s manufacturing-intelligence layer, not in a payload.

  • It’s the sovereign-compute logic applied to the factory floor, and another NGen cheque worth tracking alongside the broader manufacturing-AI thread.

🌆 Toronto Tech Week Recap

Many heavy hitters used Toronto Tech Week to circle the same theme: keeping Canadian AI Canadian. The marquee stages kept returning to retention: keep the company here, keep the IP here, keep the power to run it here.

• Raquel Urtasun (Waabi) told U of T president Melanie Woodin at BetaKit’s Most Ambitious Town Hall that physical AI is at its pre-ChatGPT moment: “Imagine you have the crystal ball before the ChatGPT moment. That’s where we are now.” Ignoring it would be “such a big miss for the country,” she said, before saying the quiet part plainly: “I want this company to remain, as much as possible, Canadian.” A clear marker from the founder of the largest raise in Canadian history.

• Nick Frosst (Cohere) went structural at Homecoming’s sovereignty panel. The binding constraint on sovereign AI isn’t talent or models, it’s electricity. Canada, he argued, “just need[s] more nuclear power plants” to run it. The timing helps his case: Ontario is already building the first small modular reactor in the G7 at Darlington, backed by $3B in federal and provincial funding, the kind of clean baseload that a sovereign-compute buildout will need. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

• Taleeb Noormohamed, parliamentary secretary to AI Minister Evan Solomon, gave the clearest preview of the coming strategy at ALL IN Talks Toronto. Canadians have watched industrial-scale IP and long-term value “migrate elsewhere,” he said, and “our job is to make sure that stops.” Read alongside Carney’s “coming this week,” IP retention is shaping up to be the organizing idea.

• Tobi Lütke (Shopify) pushed back on the one-person-unicorn fantasy, calling it “bullshit,” and reframed today’s tools as a “seven out of ten engineer on your phone,” not a replacement for teams. While the room talked about keeping Canadian AI here, his quieter point was about making sure there’s something durable worth keeping.

🛡️ CANSEC Recap

CANSEC was the AI-and-defence show this year, and the symbolism wasn’t subtle. Canada’s largest defence trade show ran May 27 to 28 at Ottawa’s EY Centre, now renamed the Cohere Centre after the Toronto AI firm bought the naming rights. Record crowds, a first-ever sitting-PM keynote, and a wave of sovereign-autonomy launches. The throughline, again, was keeping the capability Canadian. (Disclosure: Cohere is a Mistral Venture Partners portfolio company.)

• Carney goes to CANSEC. The first sitting PM to keynote the show, Carney pledged $180B in defence procurement over ten years and named “unimpeded access to artificial intelligence” a strategic capability alongside space and quantum. On the venue’s new namesake: “If you don’t know who Cohere is, you will.”

• Volatus unveils a sovereign autonomy brain. Toronto’s Volatus Aerospace introduced V-Cortex, an AI flight controller and autonomy OS on Canadian-controlled IP, pitched as the base of a “sovereign Canadian autonomy ecosystem” for uncrewed platforms.

• Calian launches ATHORA. Ottawa’s Calian debuted ATHORA, an open-architecture layer to stitch sensors, networks, and legacy systems across every domain. It’s the connective tissue for the agentic-AI work Calian is doing with Cohere’s North.

• INKAS goes counter-drone. Canadian maker INKAS launched ANURI, a modular AI-guided counter-UAS interceptor with visual terminal homing, claiming time-to-target under 80 seconds at five kilometres.

• A 13-company Canadian autonomous vehicle. Ottawa’s Convergence Design Services launched the MIL-V, a fully electric autonomous military vehicle on 100% Canadian IP, integrating 13 Canadian firms plus Arctic OWL, a distributed edge-AI sensing network.

• MDA’s 49North teams with Voyageur on ISR. MDA Space’s defence arm 49North signed an MOU with North Bay’s Voyageur Aviation to pair crewed and uncrewed aircraft with AI-enabled ISR data fusion for Canadian customers.

• The Palantir question. Defence Minister David McGuinty defended a $3.7M DND data-analytics contract with U.S. firm Palantir as “a legitimate procurement” that’s “moving forward,” while saying the government “will look at this question of data sovereignty.” As the Toronto Star’s Alex Ballingall reported, the deal surfaced only through documents tabled in Parliament. An awkward counterpoint to a show themed on keeping AI Canadian.

🏢 Large Companies

• Cohere (Toronto) and Calian Group (Ottawa) partnered to bring sovereign AI to the defence industry.

  • The collaboration puts Cohere’s secure agentic platform, North, into controlled defence environments for the Canadian military, allied forces, and SMEs, routed through Calian VENTURES.

  • Use cases named: mission planning, decision support, and training. It’s an evaluation-and-integration agreement rather than a signed contract, so no dollar value yet, but it’s a meaningful sovereign-AI-meets-defence signal and dovetails with the CANSEC wave above.

🔬 Research

• Rich Sutton (Edmonton - University of Alberta) says generative AI literally cannot discover anything, and he wants science to stop pretending otherwise. The Alberta RL pioneer and 2024 Turing laureate posted a recorded talk, delivered to the SAIR Foundation, arguing that supervised-learning systems are stuck choosing between novel and good, and can never be both at once.

  • His frame is an old academic joke: the work is both novel and good, except the good parts aren’t novel and the novel parts aren’t good. That, he says, is generative AI exactly. Output is either drawn from training data (good, because the source is good) or from the model’s own randomness (novel, but unmoored). Never both in the same stroke.

  • The missing ingredient he names is evaluation. Real discovery needs three steps: variation, evaluation, and selective retention. Generative models only do the first. That’s why he files AlphaGo, AlphaFold, AlphaProof, and even Claude Code on the other side of the line: they evaluate against a goal, so they actually find things. Garden-variety LLMs don’t.

  • This is the sharpest articulation yet of the Alberta school’s bet against the LLM consensus, and it’s not idle contrarianism. Sutton grounds it in his own group’s 2024 Nature paper on continual backpropagation.

• Torc Robotics partnered with Mila in the institute’s first-ever autonomous-trucking collaboration.

  • Torc, a Daimler Truck subsidiary, is embedding on-site at Mila with dedicated research space, working on generative world models, multi-agent behaviour modelling, reinforcement learning, and foundation models for physical AI. Mila’s Liam Paull (CIFAR AI Chair, UdeM) and Torc AI head Felix Heide are the named principals.

  • There’s a quiet tension worth naming here. The same week founders were on Toronto stages insisting Canadian AI stay Canadian, foreign industrial capital planted a research flag inside Canada’s flagship AI institute. Both things can be good.

🏛️ Policy

• The national AI strategy is, almost, here. Carney told reporters on May 27 that the long-delayed Pan-Canadian AI strategy is “coming out next week.”

  • The fall 2025 consultation drew more than 11,000 comments. Six pillars have been previewed: protecting Canadians and democracy, new privacy and online-safety laws, sovereign compute, scaling Canadian AI companies, international alliances, and AI training.

  • Don’t hold your breath. This strategy has already slipped past end-2025 and Q1-2026 targets, and Solomon’s own office softened Carney’s “next week” to “imminent.” And then a draft was leaked to the CBC. I’m going to wait for the real thing before I comment.

• SCIP’s $890M window closes today. Applications for the Sovereign Canadian AI Compute program were due June 1.

  • Up to ~$890M over seven years, with leads limited to Canadian non-profits, post-secondary institutions, or the consortia they head. A Queen’s University-led consortium is a named likely bidder, having recruited an ex-Nvidia supercomputing engineer and signed Simon Fraser and Bell as partners.

  • No winner has been named yet, so this is one to watch into next week. (And no, despite the rumours, Telus is not a SCIP winner. That’s a separate sovereign-data-centre program.)

📊 Data

• Osler’s 2025 Deal Points Report confirms what we all see on the field: AI is now the centre of gravity in Canadian venture.

  • AI companies made up 23.6% of all financings and a striking 54% of all capital invested in 2025. As partner Ryan Unruch put it, AI is “the single largest industry captured by the report by both deal count and dollars invested.”

  • The premium is real: median pre-money valuations for AI companies ran roughly 60% above the rest of the market, with median round sizes about 50% larger. Osler sees a lot of deals. I find their data to be one of the most reliable in Canada.

📰 In brief

• Ontario’s universities released a major AI report, “Talent, Technology and Trust,” at an Empire Club event during Toronto Tech Week. Roughly a dozen recommendations, chaired by Waterloo president Vivek Goel, with a pointed call for Ottawa to fund sovereign AI research infrastructure.

• Bill C-22 keeps biting. Public Safety Minister Gary Anandasangaree rejected calls to shorten the metadata-retention period in the federal lawful-access bill, holding firm at up to one year despite pushback from Apple, Google, Meta, and the Privacy Commissioner. Signal and NordVPN have threatened to leave Canada rather than comply. The data-governance backdrop to every “build AI here” pitch.

Have a great week! See ya 👋

— Raif

Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.

Was this forwarded to you? Subscribe below.

Subscribe now

I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.

Upper Bound 2026 wrapped in Edmonton this week, and the money showed up. Two new Canadian venture funds with AI mandates (Sagard $150M USD, N49P $25M USD first close) on top of last week’s three (Top Down, SFU Innovates, DevCap) puts us at five new funds in fourteen days. Layer in $57M of fresh federal and provincial money flowing into AI this week alone ($24M CIFAR Chairs refresh + $16.5M FedDev Ontario + $6.8M PrairiesCan + $10M Alberta/Amii Health Innovation Lab) and you’re at roughly $235M of new capital deployed or committed since Mother’s Day. Edmonton ran the table on the policy side: PrairiesCan dropped $6.8M on the Upper Bound stage, Alberta and Amii announced a $10M Health Innovation Lab, U of A and Amii wrapped their $30M recruitment campaign 25% above target, and the conference itself pulled 11,000 attendees with 53% YoY growth. Cohere had another big week too. Let’s get into it.

🚀 Deals & Milestones

Cohere had another big week with an acquisition, two sovereign-AI MoUs in Spain, and an open-weight flagship

Disclosure: Cohere is a Mistral Venture Partners portfolio company.

• Acquired Reliant AI on May 19, the second German-tied deal in seven weeks after Aleph Alpha. Reliant was founded in 2023 by ex-DeepMind researchers Karl Moritz Hermann, Marc Bellemare (a Mila CIFAR Chair) and Richard Schlegel. Its 35 employees join Cohere; Hermann and Bellemare become VPs. The Reliant team powers North for Pharma, Cohere’s biopharma agentic platform with named customers including GSK and Ipsen. Reliant had raised US$11.3M seed from Inovia Capital and Mike Volpi. Terms undisclosed.

• Signed MoUs with Indra Group and Multiverse Computing on May 20 during King Felipe VI’s state visit to Canada. IndraMind covers sovereign AI deployment across Spain, Canada, and Europe, with defence interoperability and adaptation to all five of Spain's official languages. The Multiverse partnership explores quantum-inspired AI optimization via the Canada-Spain bilateral cooperation framework.

• Released Command A+ as open-source on May 20, a mixture-of-experts model under Apache 2.0. Nick Frosst on X: “This tech can go one of two ways.” Positioning: sovereign open-source as a counterweight to China-anchored DeepSeek and Qwen.

• In seven weeks, Cohere has executed two acquisitions, signed three open-government partnerships, and shipped an open-weight flagship. That’s a coherent sovereign-AI strategy moving at IPO-prep pace.

Sagard launches a $150M USD AI Fund

• Anchored by the Power Corp family (IGM Financial and Great-West Lifeco), Sagard’s new AI fund is its first dedicated AI vehicle and explicitly designed to back early- to growth-stage AI software.

• Paul Desmarais, Sagard’s Executive Chairman, told reporters the fund will invest across North America with a Canadian lean.

• Power’s already-substantial backing of Canadian tech (including Wealthsimple) now has a dedicated AI vehicle behind it. With Sagard’s other franchises (venture, growth, credit), this rounds out the Power AI footprint.

N49P hits $25M USD first close on Fund IV

• Our good friend Alex Norman’s Toronto-based seed fund hit $25M USD first close toward a $70M USD target on May 21. Northleaf Capital Partners is a backer. The fund has already deployed into cybersecurity startup NexRisx.

FedDev Ontario deploys $16.5M across 13 GTA AI companies

• Announced by Minister Evan Solomon at MaRS on May 25 to kick off Toronto Tech Week. The most granular federal AI capital allocation we’ve seen in 2026. Vector Institute’s $4M is separated out below. The twelve company-level grants:

• Private AI (operating as Limina), Toronto, $2M. Sensitive-data infrastructure for regulated enterprise AI.

• Cosm Medical, Toronto, $1.99M. AI-driven Gynethotics platform for custom pelvic-health devices.

• ProteinQure, Toronto, $1.8M. AI protein and peptide drug discovery. First AI-designed peptide is in Phase I at Princess Margaret.

• DMD Building Systems, Scarborough, $1.7M. AI plus robotics for structural-steel manufacturing.

• Fiscal.ai (Stratosphere Technology), Toronto, $1.52M. AI for unstructured financial filings to structured-data API.

• MinuteBox, Thornhill, $700K. AI for cloud legaltech entity management.

• Edgecom Energy, Toronto, $575K. AI Energy Co-Pilot.

• Future Fertility, Toronto, $555K. AI endometrial receptivity assessment, 300+ clinics in 35 countries.

• Naryant, Oakville, $536K. AI fleet management analytics.

• MarkiTech, Oakville, $500K. CliniScripts AI clinical-workflow automation focused on autism care.

• Trax, Toronto, $394K. AI-assisted building-permit compliance.

• VisFuture, North York, $200K. Natural-language AI for SME business data.

Vector Institute lands $4M to run an AI startup-deployment program

• The largest single line in the FedDev tranche didn’t go to a startup. It went to Vector Institute, which will use the funds to run a Data Readiness, Model Development and Deployment program for Ontario AI startups.

PrairiesCan deploys $6.8M across Alberta AI

• Announced by Solomon on the Upper Bound stage on May 19. The PrairiesCan Regional AI Initiative dropped $6.8M across five recipients. U of A’s $3M sovereign-compute vault is in Research below. The four company-level grants:

• Vertical City, Edmonton, $1.3M. AI/ML ad-tech expansion.

• Darkhorse Emergency Corp, Edmonton, $1M. AI analytics for fire and emergency services.

• NTWIST, Edmonton, $1M. Industrial AI for mining and high-mix/low-volume manufacturing.

• Localintel, Calgary, $500K. Location intelligence content platform for municipalities and economic development offices.

• All four are applied and industrial AI, not pure software. Consistent with the province’s positioning, and a useful pair with the Alberta-Amii Health Innovation Lab below.

Sustainable AI Group launches in Montréal

• Sasha Luccioni, formerly Hugging Face’s AI and climate lead and a Yoshua Bengio postdoc at Université de Montréal, and Boris Gamazaychikov, formerly Salesforce’s head of AI sustainability, launched Sustainable AI Group (SAIG) on May 22. The Montréal-based advisory firm targets enterprise AI-sustainability reporting standards.

• Gamazaychikov to BetaKit: “the structure [AI] is on is incompatible with sustainability, and that presents a lot of business risks.”

• The launch lands as Canadian data-centre build-out accelerates (Telus B.C., Cohere/CoreWeave, Bell B.C.). A useful Canadian counter-narrative.

Kritik ships VisibleAI for university faculty

• Toronto-based Kritik, founded by Top Hat alum Mohsen Shahini with Carine Marette and Mark Deepwell, launched VisibleAI on May 19. The tool lets faculty see what students typed versus what they prompt-generated.

• Marette: “When the AI is constrained, the students are using more of their creativity and technical skills to provide higher quality work, as opposed to cognitive outsourcing.”

• Kritik is explicitly not selling AI detection. They’re selling AI transparency, which is a much better positioning bet in the long run. Faculty can decide how much AI is ok.

Trane and BrainBox AI open a Montréal innovation lab

• BrainBox AI co-founder and CEO Jean-Simon Venne opened a Montréal AI innovation lab and showroom focused on agentic AI for HVAC on May 20. First major public presence since Trane acquired BrainBox AI.

• BrainBox’s models claim up to 25% energy savings and 40% emissions reduction. Showroom screens at the launch displayed client savings “typically over 20 percent.”

VidCruiter ships AI Interview Scoring

• Moncton-based VidCruiter, led by CEO Sean Fahey, launched AI Interview Scoring on May 19. The product applies client-defined rubrics to pre-recorded video interviews with written explanations for each score and explicitly avoids biometric or face analysis. VidCruiter reports $15-20M ARR.

Peripheral Labs opens a basketball biomechanics lab in Toronto

• The U of T robotics spinout, led by CEO Kelvin Cui, opened what it claims is North America’s first non-NBA biomechanics basketball shooting lab on May 25. Partner is the Quantum Sports and Learning Association (QSLA) at a Dovercourt Village facility.

• Spatial-intelligence AI built for autonomous vehicles is finding a second life in sports tech. Worth watching whether the wider Canadian self-driving diaspora seeds more sports-AI startups in the next year.

🔬 Research

CIFAR AI Chairs get a $24M refresh

• Solomon announced $24M on May 21 to support 20 new and 22 renewed Canada CIFAR AI Chairs, including a renewal for Turing Award co-recipient Richard Sutton. The program now has 143 active Chairs across Amii, Mila, and the Vector Institute.

• Elissa Strome, executive director of the Pan-Canadian AI Strategy at CIFAR, described the network as “the third-highest-impact AI research cluster in the world.”

• Per The Logic, the announcement exhausts the program’s remaining $162.2M envelope ahead of an expected national AI strategy update.

U of A and Amii close $30M recruitment campaign 25% above target

• The joint $30M recruitment campaign for AI talent wrapped on May 21 with 25 researchers hired against a 20-target.

• While the national discussion is about Canadian AI talent flowing south, U of A and Amii quietly hit 125% of a $30M hiring campaign.

U of A gets $3M from PrairiesCan for the Canadian AI Compute Vault

• Part of the PrairiesCan $6.8M tranche, the U of A allocation funds CAICV, a sovereign HPC infrastructure project for Canadian AI research. The data-sovereignty framing was explicit: workloads will not move through U.S. or foreign cloud systems.

• Sovereign compute is the recurring theme of 2026. CAICV joins Telus’ B.C. cluster, Nutanix’s Vancouver center, and the federal Compute Access Fund as Canadian alternatives to U.S. hyperscalers.

Upper Bound 2026 by the numbers

• 11,000 attendees, 22 countries, 53% YoY attendance growth, per Amii’s GlobeNewswire release on May 19.

• Third year of compounding growth and likely the largest sustained AI conference in Canada now.

🏛 Policy

Alberta and Amii launch a $10M Health Innovation Lab

• Alberta Minister of Technology and Innovation Nate Glubish announced $10M over three years on May 20 to establish a joint Health Innovation Lab with Amii. The lab will deploy AI into Alberta Health Services workflows.

• The sovereign-infrastructure framing was the headline: Alberta health data will not move through foreign cloud systems.

• Pair this with the Alberta-Amii partnership on the data-sovereignty side, the U of A CAICV grant on the compute side, and the PrairiesCan industrial-AI tranche on the commercial side. Alberta is now executing a coherent sovereign-AI strategy at the provincial level.

📅 Upcoming Events

• Canadian AI Conference 2026 (Vancouver) — May 25–29, 2026

• ALL IN Toronto — May 28, 2026

• ALL IN Main Event (Montréal) — September 16–17, 2026

Thank you for reading! Subscribe, and please share any feedback 🙏🏽

— Raif Barbaros