Term sheets usually describe what an investor receives. Cohere’s newest one also describes what the company must build, in which country, and by when: at least C$1.7 billion of computing infrastructure inside Canada before 2030.
The Toronto developer of enterprise AI models said on Oct. 29 that it had completed a first close on a Series E round targeting as much as C$7 billion, at a post-money valuation of C$35.7 billion. The Canadian government and private investors are both participating, and proceeds go toward expanding deployments for corporate customers and paying for computing capacity.
The step up is steep. Cohere was valued at $6.8 billion in a round closed last year, or roughly C$9.5 billion at prevailing exchange rates, which puts the new price at close to four times that mark in twelve months. In dollar terms the company is now worth something near US$25 billion, a figure that would have described a consumer platform rather than a vendor of models sold to IT departments.
Cohere’s business is not the one most people associate with the current boom. It does not run a chatbot for the public and does not compete for consumer downloads. It sells access to its models two ways: through an application programming interface billed by usage, and through deployments that run inside a customer’s own cloud account, its own data center, or on hardware shipped to its premises. Its platform for building agents and retrieval systems is sold under the name North. Much of its distribution runs through partners, including Oracle, SAP, Fujitsu and Dell, which resell its models inside enterprise software and hardware they already sell.
The company’s founding story explains the pitch. Aidan Gomez, who started Cohere in 2019 with Nick Frosst and Ivan Zhang, was a co-author of the 2017 paper that introduced the transformer, the architecture under essentially every large model now in commercial use. Cohere’s argument has been that the same research lineage, aimed at corporate data rather than at a general audience, produces something a bank or a health system can actually deploy.
The word the company uses for that market is sovereignty, and it has become a budget line. Governments and regulated industries increasingly want models that run under their own law, on their own premises, subject to their own audit, and they are willing to pay a premium over the price of an API call to a data center in Virginia. Canada has its own version of the policy, a sovereign compute program funded at C$2 billion and announced in late 2024, and the condition on Cohere’s round attaches private capital to it: public money and public procurement in exchange for private construction.
The C$1.7 billion is capital Cohere must spend, not money it receives. It is a commitment to buy accelerators and data-center capacity in a country with cheap power and a slow approvals process for industrial sites. The investment will show up in the company’s cost base long before it shows up in revenue, and it converts a software business into something closer to an infrastructure operator, with the depreciation schedule to match.
The size of the raise suggests how much capacity the company thinks it needs. Enterprise inference, the work of answering queries for thousands of employees a day, is a throughput problem before it is a research problem. Customers who put a model behind a claims system expect consistent latency and contractual guarantees, which means dedicated capacity rather than shared capacity, and that has to be bought years ahead of demand. Every large buyer of computing has learned the same lesson in the past two years: ordering early is expensive, and ordering late is worse.
Then there is the competitive problem the money does not solve. Open-weight models from Meta, Mistral, Alibaba and DeepSeek can be downloaded without a license fee and run on hardware the customer already owns. For a company whose proposition is that a model should live inside the customer’s perimeter, that is not a distant threat; it is the substitute sitting next to the product. Cohere’s answer is tuning, support, compliance documentation and the engineering work of making a model behave in a specific domain, which is valuable and also the kind of value that gets priced like services rather than like software.
Most enterprise AI spending is now routed through the three large cloud providers, each of which sells its own models first. Cohere’s partner-led distribution is a way around that gate, and it depends on partners continuing to prefer a neutral supplier to a competing one. Oracle’s stake in the company helps; it does not settle the question.
The Canadian government’s role cuts both ways as well. It is an investor, a prospective customer, a regulator and a source of industrial policy, and those roles do not always point the same direction. A change of government would not unwind a closed round, but it would change the demand the round is meant to serve. Sovereign AI is a policy fashion as much as a procurement category, and fashions change with elections.
For now the company has what it asked for: a valuation that puts it among the most valuable private AI developers outside the United States, a domestic political constituency, and enough capital to build before it sells. The risk it is taking is the mirror image of the one its customers are avoiding. They want a model that never leaves the building; Cohere is betting that enough of them will pay for one that the buildings are worth owning.
The round’s first close is not the whole C$7 billion, and the rest will arrive only if the commitments hold. What the valuation really prices is a forecast about 2030: that the demand for models which stay inside a country’s borders will be large, permanent and expensive enough to justify the concrete.
