How Cohere Wrote the Fund That's Now Funding Everyone Else

On Christmas Eve 2024, the Government of Canada issued a Letter of Intent to Cohere for a Strategic Innovation Fund contribution of $240 million.

The timing is the first thing worth noticing.

Christmas Eve is when governments send letters they want done quickly and quietly.

The Trudeau government was already wobbling.

Chrystia Freeland had resigned on December 16, eight days earlier, in a public letter that called out the Prime Minister's "costly political gimmicks."

Less than two weeks later, Trudeau would announce his own resignation.

The LOI was not a routine grant decision.

It was a $240 million commitment from a government in the last weeks of its political life, made to a single Canadian company at the centre of its AI strategy.

Eleven weeks later, on March 10, 2025, the actual Contribution Agreement was executed.

Trudeau was technically still Prime Minister but had announced his exit. Mark Carney would be sworn in four days later.

The agreement runs through April 30, 2049 — a 24-year term.

A few weeks ago I filed an Access to Information request and received the briefing documents that surround this decision.

Disclosure: I operate in Canadian AI infrastructure, so I read these from inside the ecosystem, not above it.

What's in them changes how the entire Canadian AI policy architecture reads.

The Four-Point Proposal

Inside the briefing materials prepared for the Minister of Innovation, Science and Industry — ahead of a meeting with Cohere CEO Aidan Gomez in late 2024 — there is a document labeled Annex B. It is titled "Cohere Compute Proposal."

It is Cohere's own written pitch to the Government of Canada, recommending how the federal AI compute strategy should be designed. The proposal makes four design recommendations.

First, subsidies should go to buyers, not suppliers.

Cohere argues that subsidizing compute operators — data centres, GPU rental companies — creates rent-seeking and doesn't solve the underlying problem of Canadian AI companies lacking compute access.

Subsidies should instead go to Canadian AI companies purchasing compute, so they can scale faster.

Second, focus on pre-training, not inference.

Pre-training large foundation models is where the capital intensity lives.

Inference is comparatively cheap.

The strategic value of a subsidy accumulates when applied to pre-training, where a single training run can cost tens of millions of dollars and one subsidized dollar moves the needle on capability.

Third, a higher subsidy for Canadian-located compute.

Tapering over years two and three to zero for non-Canadian compute.

Recognize that Canadian compute capacity doesn't exist today but should be built — and create a glide path that rewards both immediate Canadian buying and longer-term Canadian infrastructure build-out.

Fourth, leverage experienced global operators.

Don't try to build a full Canadian stack from scratch.

Use Oracle, Google, Nvidia and similar providers in the short term, and incentivize them to expand into Canada through the subsidy structure.

These four recommendations became the architecture of the AI Compute Access Fund — the federal program that announced 44 funded projects totalling $66 million at Web Summit Vancouver on May 12, 2026.

The document in the ATIP release is Cohere recommending, in writing, the structure of the funding architecture under which Cohere would become the single largest beneficiary.

The Anchor Catalyst Argument

The proposal is honest about the strategic logic.

Cohere positions itself as the "anchor catalyst": the argument that small Canadian AI companies, individually, cannot create the demand signal needed to convince hyperscalers like Google, AWS, Microsoft or Oracle to build GPU capacity in Canada.

The aggregate demand from a hundred Canadian startups is still smaller than the demand from one Cohere.

Cohere's claim, in the proposal:

"None of the requisite compute capacity exists in Canada, nor are there any plans by compute operators to build it in Canada at this time."

The proposed solution: subsidize Cohere's pre-training purchases, structured to favour Canadian-located compute.

Cohere's resulting purchase orders to hyperscalers become the demand signal that gets Canadian GPU capacity built.

Smaller Canadian buyers then benefit from the infrastructure that follows.

This is internally coherent. It is also, on reflection, an argument that the policy should make Cohere the buyer-of-record for Canadian AI infrastructure.

The Generational Term

The Strategic Innovation Fund contribution to Cohere is the largest single AI investment by the Government of Canada to date.

From the ATIP-released Quick Facts briefing:

A $240 million contribution against total project costs of $724.75 million, structured as a partially conditionally repayable contribution, running March 10, 2025 to April 30, 2049.

The named compute partner is CoreWeave, providing "first access to the latest Nvidia superchip," the GB200.

The 24-year term is the line that stops the eye.

Most government contributions to private companies run five years, sometimes ten.

A 24-year horizon is a generational bet — longer than the average career of a Member of Parliament, longer than two complete election cycles, longer than the median tenure of a senior public servant.

The funding ratio is worth doing the arithmetic on.

A $240 million federal contribution against $724.75 million in total project costs implies Cohere is putting roughly $484 million of its own capital alongside the federal investment.

Cohere's investor stack includes Nvidia, Salesforce Ventures, Oracle, EDC (the federal export agency), PSP (a federal pension), Cisco Investments, SAP, Inovia, Radical Ventures, Section 32 and Tiger Global, among others.

The federal share is approximately 33 percent. This is a co-investment, not a subsidy. The government is buying a position in it.

The speed matters as much as the size.

The agreement was approved by the Minister on March 7, 2025, executed on March 10, and publicly announced on March 20.

From Minister approval to public announcement was twelve days. Approval to signature was three working days.

By comparison, ISED contribution agreements at smaller scales typically take months from initial discussion to executed signature.

The Cohere SIF was moved with unusual political urgency.

The MOU

In August 2025, the Carney government did something the Trudeau government had not.

It issued a formal Memorandum of Understanding between the Government of Canada and Cohere.

The MOU was signed by Minister Evan Solomon (AI and Digital Innovation), Minister Joël Lightbound (Government Transformation, Public Works and Procurement) and Aidan Gomez.

The substantive sections — Objectives, Specific Areas of Commercial Collaboration, Forms of Cooperation — are entirely redacted in the ATIP release under Section 18(b) of the Access to Information Act, which covers federal economic interest.

But the cover memo from Deputy Minister Mark Schaan, recommending approval to Minister Solomon, is largely visible. It contains the language that defines the federal position.

Schaan writes that the objective is "to signal the GC's commitment to Cohere as the only sovereign, cloud-agnostic large language model operating in Canada and solidify their role as a key partner in the GC's AI and digital innovation agenda."

The framing he uses elsewhere in the memo, in a sentence I read twice to confirm: "champion one of our champions."

The MOU also positions Cohere for national defense contexts.

The memo specifies that Cohere is being designated for both civilian government productivity and "the nation's international commitments in mutual defense agreements."

That is a Five Eyes, NORAD and AUKUS hook surfaced explicitly in a Government of Canada document.

The MOU is not legally binding — either party can terminate on 60 days notice.

But the act of signing it, co-signed by two cabinet ministers and the CEO, is the act of public political designation.

Under Trudeau, Cohere received the SIF financial commitment.

Under Carney, Cohere received the additional formal political designation as Canada's national LLM.

That is the inheritance that matters.

The Trudeau government's financial decision could have been left to execute quietly.

The Carney government chose, in its first eight months, to publicly re-endorse it.

On May 12, 2026, Minister Solomon named Cohere "Canada's foundation model" at Web Summit Vancouver.

What It Means For Everyone Else

The Government of Canada has designated one company as its national champion at the foundation model layer:

$240 million in direct contribution, a 24-year term, an MOU defining the partnership, an explicit defense framing, and public re-endorsement across two governments.

It is the largest financial and political commitment in Canadian AI policy history.

Every other Canadian AI company operates downstream of that designation — including the 44 AI Compute Access Fund recipients announced May 12, who are disbursing public compute money under a structure Cohere designed.

For a Canadian organization actually buying AI — a hospital network, a 200-lawyer firm, a provincial agency — the practical read is narrower than the headlines suggest.

"Sovereign Canadian AI" has been defined, at the federal level, as one hosted foundation model.

Whether that definition fits your regulatory constraints, your procurement rules or your data-residency obligations is a question the designation does not answer for you.

The federal champion solves the foundation model problem. It does not solve yours.

The Honest Reading

There is a version of this that reads as an accusation. It isn't one.

Cohere did what any well-run company would do when offered the chance to shape policy in its sector.

They wrote the proposal, submitted it, and the government accepted it. None of that is improper.

Cohere is also building real technology, hiring real engineers and generating real economic activity.

Two hundred forty million dollars of public money matched against roughly $484 million of private capital, across a 24-year horizon, is a meaningful bet — and it may prove a sound one.

The federal logic is defensible too. Canada needed a foundation model champion. Cohere is the only Canadian company plausibly at that scale. The alternative — letting US foundation model providers be the only options for Canadian AI work — carries real strategic costs.

And the structural design of the fund (subsidies to buyers, a Canadian-location premium, a pre-training focus) is competent policy, even if Cohere wrote it.

Government decision-makers cannot evaluate dozens of competing foundation model proposals. They consolidate, and they pick a champion.

The mechanism for that consolidation, here, was Cohere's own proposal to design the program in which it would be the anchor.

That is not corruption. It is also not neutral.

It is how policy gets made when a sector has one obvious anchor and dozens of smaller players who lack the time and political access to write proposals.

The thing worth surfacing is the information asymmetry. Operators living inside this funding architecture have been making decisions — how to deploy compute, how to position work, how to talk to ISED program officers — without having seen Annex B. Now it's readable. That changes how the downstream decisions read.

The information was released to me, and most of it is technically public. It deserves to be readable by people who aren't paid to read ATIP packages.

The structural mechanism by which Canada's AI policy gets made should be visible to the operators living inside it.

The Canadian AI policy architecture has one anchor. The ecosystem around it is wide open.

Receipts

This piece is based on four ATIP release packages:

• A-2024-00994 (Cohere meeting briefing, including the Annex B compute proposal),
• A-2025-01015 (SIF Quick Facts briefing on the $240 million contribution),
• A-2025-00997 (the MOU approval memo from DM Schaan to Minister Solomon) and
• A-2025-01507 (the signed MOU). They are public records, released under the Access to Information Act, and referenced throughout.

Specific receipts:

• $240,000,000 SIF contribution to Cohere, agreement executed March 10, 2025
• $724,750,000 total project value
• 24-year project term, March 10, 2025 to April 30, 2049
• Letter of Intent issued December 24, 2024
• MOU signed August 13, 2025 by Solomon, Lightbound and Gomez
• "The only sovereign, cloud-agnostic large language model operating in Canada" — DM Schaan, August 2025
• "Champion one of our champions" — DM Schaan, August 2025
• "Canada's foundation model" — Minister Solomon, Web Summit Vancouver, May 12, 2026
• 44 AI Compute Access Fund recipients announced, $66 million total, May 12, 2026

The fund was designed by Cohere. The 44 projects funded under it operate downstream of the same architecture.

This is what I learned when I asked.