On October 29, Prime Minister Mark Carney announced that Ottawa will invest up to $1.6 billion in Cohere, Canada's best-known AI company.
If it's all paid out, that's roughly $39 for every person in Canada, using Statistics Canada's April 1, 2026 population estimate.
I build AI models in Toronto. My company takes federal support too. So this isn't a piece against public money in AI.
It's a piece against putting this much of it into one company, on terms Canadians aren't allowed to see.
Because I think Ottawa is building something it won't be able to walk away from.
What we know, and what we don't
Here's what's public. Cohere committed to stay incorporated and headquartered in Canada, create at least 1,000 new technical jobs by 2030, and put at least $1.7 billion into new Canadian computing infrastructure by 2030. The money comes in stages, with regular reporting.
It's part of a fundraising round targeting up to $7 billion at a $35.7 billion valuation. Cohere announced an 'initial close', and a spokesperson said it has 'most of the target funding'. Germany's Schwarz Group, the owner of Lidl, is leading.
AI Minister Evan Solomon said payments are tied to goals, and that 'jobs and revenue targets are key'. He wouldn't share the commercial terms.
Here's what we don't know. How big the stake is. The price. The milestones. What happens if Cohere misses one. Whether Ottawa gets a board seat. How Canada ever sells. The Crown corporation managing the stake declined to disclose any of it, and said the final arrangements were still being negotiated.
Divide $1.6 billion by the valuation and you get about 4.5% of the company. That's my arithmetic, not a disclosed number. Nobody outside the deal knows what we actually bought.
That's the first problem. Not that Ottawa invested. That it invested and won't show us the deal.
The Crown corporation that owns a pipeline
The stake is managed by the Canada Development Investment Corporation, or CDEV. It reports to Parliament through the Finance Minister. It also owns the Trans Mountain pipeline through a subsidiary.
CDEV is one of 7 public institutions now tied to Cohere.
| Who | What it is | Tie to Cohere |
|---|---|---|
| CDEV | Federal Crown corporation | Manages Ottawa's up to $1.6B stake (2026) |
| Strategic Innovation Fund | Federal program | $240M for computing power - 'partially conditionally repayable', agreement runs to 2049 |
| BDC Capital | Federal Crown corporation | Shareholder since 2025 |
| Export Development Canada | Federal Crown corporation | Shareholder since at least 2024, and makes introductions for Cohere abroad |
| Shared Services Canada | Federal department | Customer - Cohere's model powers CANChat, the AI assistant for public servants |
| PSP Investments | Federal public servants' pension manager | Led Cohere's US$500M round in 2024 |
| CPP Investments | Canada Pension Plan manager | US$50M in 2025 |
They're not all the same thing. The 2 pension managers invest at arm's length. One is a customer. But step back and the picture is clear. Build Canada CEO Lucy Hargreaves put it best: Ottawa is becoming 'the owner, the funder, the customer and the regulator.'
The $240 million is described in the government's own database as 'partially conditionally repayable', on an agreement running to 2049. It doesn't say how much comes back, or when.
And it won't stop here. Solomon said Cohere won't be Ottawa's last equity stake in a tech company, and a roughly $500 million Canadian Tech Growth Fund is being designed to take more.
That's why this deal's design matters. It's the template.
Too Canadian to fail
Look at what Cohere signed up for.
The jobs. At least 1,000 new technical jobs in Canada by 2030. Cohere has more than 800 employees today, worldwide. Its own plan goes further - more than 1,500 new Canadian roles by 2031.
The computing. C$1.7 billion by 2030 works out to roughly $425 million a year. How much of that overlaps with Cohere's own $4 billion Canadian program, nobody has said.
Some of it is already locked in. Cohere signed a 3-year, US$225 million contract with HIVE Digital for computing, connectivity and storage. To serve it, a HIVE subsidiary borrowed about US$129 million at a fixed 9% to buy a US$165 million cluster of 2,303 NVIDIA chips. That debt is HIVE's, not Cohere's. But the commitment is Cohere's.
And the profits. Cohere hasn't said it's profitable. CEO Aidan Gomez expects it to get there before 2029.
Each of these is defensible on its own. Together, they create a company where every promise is a public promise, and every missed target becomes a political problem.
That's my fear. Not that Cohere fails tomorrow. That it becomes too Canadian to fail - and the next time it needs money, saying no feels impossible.
To be clear, Cohere may well build a strong enterprise business. Nothing in the record says it's about to collapse. The question is whether that business justifies this valuation, this much public exposure in one company, and these obligations. Without the terms, nobody outside the deal can answer that.
What is Cohere now?
Cohere was founded in 2019 to build frontier AI models. Its CEO co-wrote 'Attention Is All You Need', the 2017 paper behind the 'T' in ChatGPT. Geoffrey Hinton and Fei-Fei Li invested early. The credentials are real.
But look at its last 2 years of announcements.
- It's the anchor tenant of a $725 million CoreWeave computing project in Cambridge, Ontario, backed by $240 million in federal money.
- It named Bell its first-choice computing partner in Canada.
- It signed another computing deal for a Bell data centre in Merritt, B.C.
- It's doing deals for Aleph Alpha in Germany, Reliant AI in health care and AISLE in cybersecurity.
Its own models have moved more slowly. More than a year passed between its Command A and Command A+ flagships. On Artificial Analysis's current index, Command A+ scores 6, against a median of 23 for open models its size. One benchmark isn't everything. But it's not nothing either.
Then, on November 3 - 5 days after the deal - Cohere started hosting other companies' AI models through Model Vault. The list includes China's DeepSeek, Zhipu, Xiaomi and Alibaba, alongside NVIDIA, Google, OpenAI, Microsoft and Mistral.
In fairness, Model Vault runs those models on isolated, dedicated infrastructure, with zero data retention by default. Customer data doesn't go to the company that built the model. And the market is moving this way. On OpenRouter, open-weight models make up around 60% of US usage, mostly from Chinese developers.
But here's the timeline.
On October 26, Cohere published a security finding that an Alibaba Qwen model tried to upload a diagnostic file to a server in Beijing during a sandboxed test. The sandbox blocked it. Cohere said the behavior was rare and that it couldn't tell whether it was a training leftover or a deliberate 'call home'.
3 days later, Check Point reported 64 replays in which the model invented 52 different destinations - Amazon, Alibaba, Microsoft and 11 addresses that don't exist. Check Point concluded a backdoor probably isn't the cause.
On October 29, Ottawa announced the deal.
On November 3, Cohere's product notes listed Alibaba's Qwen among the models customers can run on its platform.
Hosting rival models on locked-down servers is a sensible business. But Cohere's pitch to Ottawa is that it's the trusted Canadian alternative.
So what exactly is the public buying? A frontier model lab? A hosting business? An enterprise software company? Ottawa hasn't said. And if it can't say what it bought, it can't tell us whether the bet worked.
The scoreboard
Solomon has said only 4 countries have built a major AI model of their own: the US, China, France with Mistral and Canada with Cohere.
Cohere was founded in 2019. Anthropic was founded 2 years later. Mistral, 4 years later. Here's how they compare, year by year.
'Revenue pace' means annualized revenue - what a company would earn in a year if its current monthly sales held. Companies define it differently, so treat these as rough comparisons, not a precise ranking.
Cohere's latest disclosure is a growth rate, not a number: annual recurring revenue up 165% year over year, and 300 customers, up 20 times. That's real growth. But it hasn't published a dollar figure since US$240 million at the end of 2025.
Mistral got to a similar valuation about 4 years faster, and it runs an app everyday people use. Cohere had the head start on both. It isn't winning the frontier race. It may not need to - but then let's be honest about what we're funding.
Germany already tried this
Aleph Alpha was founded in 2019, the same year as Cohere, to be Europe's frontier AI champion. It raised about US$520 million, and announced no new outside financing after November 2023. Its 2023 accounts showed revenue of under €1 million. In 2024 it moved away from building frontier models.
In April, Cohere announced it would absorb Aleph Alpha, with Cohere's shareholders set to own about 90% of the combined company. The definitive agreement was signed in September and is still awaiting regulatory approval. Schwarz Group, a major Aleph Alpha backer, is now leading Cohere's round.
Cohere isn't Aleph Alpha. It has real customers and real growth. But Germany's national champion has agreed to be absorbed by Canada's. The national-champion model doesn't have a great track record - and Canada just doubled down on it.
The best case for the deal
I'll give Ottawa its strongest arguments.
- Control over critical technology. Carney has argued Canada too often funded research whose rewards flowed elsewhere. US laws like the CLOUD Act can give American authorities access to data held by US companies.
- Equity, not a grant. A stake means taxpayers share the upside if Cohere wins.
- It keeps the company here. Cohere has committed to stay incorporated and headquartered in Canada.
- Other countries do it. France's public investment bank, Bpifrance, backs Mistral.
- The growth is real. 165% revenue growth, with NVIDIA and Schwarz Group in the same round.
These are real arguments. I'd make some of them myself.
None of them requires secret terms.
I'm not the only one asking. Conservative MP Adam Chambers questioned why the government needs to be a shareholder at all, and said Ottawa risks acting like an investment bank. In 2025, the union representing federal professionals called the government's earlier Cohere partnership 'privatization with a Canadian flag on it'.
What I'd do instead
Ottawa got one thing right. Solomon has said Canada won't overdo regulation. Good. Europe's instinct is control first, build second, and I don't want that here.
But avoiding Europe's regulation habit shouldn't mean copying its other one - picking a national champion and pouring public money into it.
And the plan skips the front door. It funds data centres and enterprise sales. Nothing for the AI apps people actually use. Canadians use ChatGPT, Claude and Gemini at home, and they bring those habits to work. In a 2024 Microsoft and LinkedIn survey across 31 countries, 78% of AI users said they bring their own AI tools to work. My read is that the habit often comes first and the office contract follows. A country that owns the data centres while foreign companies own the app and the habit has paid for the house and rented out the front door.
None of this requires cancelling the deal. All of it can be attached to the next payment.
- Publish the deal. The stake, the price, the payment schedule, any board seat, and how Canada eventually sells. If it's a commercial deal, show the commercial terms.
- Pay for results, not spending. Tie each payment to sales to customers other than the government, repeat customers, profit per sale, exports and independent tests. Drop the hiring target.
- Make results public. Any company holding public AI money should publish independent test results - including the bad ones.
- Spread the computing power. Fund many builders through competitive programs instead of one main tenant. The next Cohere is probably 10 people right now. Programs like this already exist - my company uses one - and they deserve a bigger share of the money.
- Open up government AI contracts. Let any model compete for federal work, judged on tests and required to run in Canada. If Cohere wins on merit, great.
- Fund the freedom to switch. Canadian servers under Canadian law, open models hosted here, and tools that make moving between AI providers easy. No single champion can own that.
- Set the rule for a missed target before it's missed. If a milestone slips, the next payment pauses, and any waiver is made public. That's just how investing works.
Own or be owned
I want Cohere to win. I want a Canadian AI company so good the world pays for it because it wants to - not because a government asked.
But look at what we've built instead. A company that hasn't said it's profitable. Tied to 7 public institutions. Bound by promises to hire, to spend and to stay. Buying companies abroad. Hosting other people's AI, including China's. On terms Canadians aren't allowed to see.
Every one of those ties makes it harder to say no next time. And Ottawa has already said there will be a next time.
Canada didn't need to buy a champion. It needed to build a market where champions have to keep earning it.
Own or be owned. Right now, Canada is building a company it can't afford to lose.
Every fact in this piece links to its source. If I got one wrong, tell me and I'll correct it publicly, with a note at the top.
George Pu is a Toronto founder. His company, SimpleDirect, builds AI models, including Vinci, that compete with Cohere's.
