Cohere finished 2025 with $240 million in annual recurring revenue, beating its $200 million target by 20% and growing more than 50% quarter over quarter. A week ago, it open-sourced a family of multilingual models small enough to run on a phone. Both moves point in the same direction: a public offering that could come as early as this year.
The Toronto-founded startup is doing something its larger rivals are not. While OpenAI and Anthropic chase ever-bigger models and ever-larger funding rounds, Cohere is building a company designed to survive contact with a balance sheet.
The IPO Preparation Is Obvious
The hiring tells the story. In August 2025, Cohere brought on Francois Chadwick as CFO. Chadwick took Uber public in 2019 and spent years navigating the regulatory and financial scrutiny that comes with a public listing. Companies don’t hire IPO-specialist CFOs to raise another private round.
The company also added Joelle Pineau, formerly VP of AI Research at Meta, as Chief AI Officer. Both hires signal a company preparing to present itself to public market investors who want credible leadership, not just impressive demos.
CEO Aidan Gomez, a co-author of the original transformer paper that started all of this, said in October that an IPO is coming “soon.” With $240 million in ARR, 70% gross margins, and a $7 billion valuation from its September 2025 Series D led by Nvidia, Salesforce, and AMD, the numbers are starting to look like a viable public company.
Tiny Aya: The Open-Source Play
On February 17, Cohere released Tiny Aya, a 3.35-billion-parameter open-weight model that supports more than 70 languages. The model is small enough to run locally on a laptop or phone without internet access.
This matters for two reasons that have nothing to do with philanthropy.
First, the benchmark results are strong. Tiny Aya Global outperformed Google’s Gemma 3 4B in translation quality across 46 of 61 languages tested on WMT24++. On the GlobalMGSM mathematical reasoning benchmark for African languages, Tiny Aya scored 39.2% compared to Gemma 3 4B’s 17.6% and Qwen 3 4B’s 6.25%. For a model less than half the parameter count of some competitors, those numbers represent a genuine technical achievement in multilingual capability.
Second, and more important for Cohere’s business: free models drive paid platform adoption. Developers who build on Tiny Aya for prototyping and edge deployment become natural customers for Cohere’s commercial products — the North enterprise platform, Command models for production workloads, and Rerank 4 for search and retrieval.
This is the same strategy that made Red Hat, Elastic, and MongoDB into multi-billion-dollar companies. Give away the engine, sell the service contract.
The Sovereign AI Strategy
Cohere released regional variants alongside the base model: Tiny Aya Earth for African and West Asian languages, Tiny Aya Fire for South Asian languages, and Tiny Aya Water for Asia-Pacific and European languages. Each variant was created using a technique called SimMerge that combines regional fine-tuning with the global model while preserving safety properties.
This isn’t just a technical novelty. It’s a direct play for the sovereign AI market — governments and enterprises that need AI capabilities but can’t or won’t send data to American cloud providers.
The European Union’s AI Act requires data residency compliance. India’s data protection framework mandates localization. Middle Eastern sovereign wealth funds are building domestic AI infrastructure. In every case, the buyer needs models that work in local languages, run on local hardware, and keep data within national borders.
Cohere’s Model Vault product — VPC-isolated model hosting with data residency compliance — is the commercial wrapper around this strategy. When a government in Southeast Asia or a bank in Germany wants enterprise AI that doesn’t route through Silicon Valley, Cohere is positioning itself as the answer.
The Competitive Position
Cohere occupies an unusual position. It is too small to compete with OpenAI ($300 billion valuation, assembling a $100+ billion funding round) or Anthropic ($380 billion valuation, $14 billion in revenue) on raw scale. It doesn’t have the distribution advantages of Google or Microsoft.
What it has is focus. Cohere builds exclusively for enterprise customers. It doesn’t run a consumer chatbot. It doesn’t need to spend billions on GPU clusters for a product that millions of free-tier users will hammer daily. Every dollar of that $240 million in revenue comes from businesses paying for production workloads.
That focus produces 70% gross margins — a figure that would make most SaaS companies envious and one that public market investors pay close attention to. By comparison, many AI companies are spending more on compute than they earn in revenue. Cohere appears to have found a model where the economics actually work.
The multilingual angle widens the addressable market. While American AI labs build models that work best in English, Cohere is going after the other 6.5 billion people on the planet who don’t speak it as a first language.
What an IPO Would Mean
If Cohere goes public in 2026, it would be one of the first pure-play enterprise AI companies to do so. OpenAI and Anthropic are both expected to file eventually, but neither has pulled the trigger. Databricks, valued at $134 billion after a late-2025 funding round, is another candidate. But Cohere could move faster — it is smaller, more focused, and appears to have the financial discipline that public markets demand.
The $7 billion private valuation implies a significant premium to revenue, roughly 29x ARR. Public markets have historically been less generous than private investors, particularly with companies that haven’t yet reached profitability. But in a market where every enterprise is scrambling to deploy AI and most of the world speaks something other than English, Cohere’s pitch is at least coherent.
It doesn’t need to be the biggest AI company. It needs to be the one that works everywhere.
The Bottom Line
Cohere’s combination of revenue growth, open-source strategy, multilingual specialization, and sovereign AI positioning makes it one of the more interesting companies in the AI industry — precisely because it’s not trying to build god. It’s trying to build a sustainable business that serves customers the big labs ignore.
Whether that story holds up in public markets remains to be seen. But the pieces are in place: IPO-ready leadership, a clear competitive moat, growing revenue, and a market that extends well beyond English-speaking tech companies.
The transformer paper Aidan Gomez co-authored changed everything. His company is betting that the next chapter isn’t about who builds the biggest model — it’s about who builds the most useful one.