March 22, 2026  |  Analysis by Shashi Bellamkonda

Research note: This analysis is based entirely on publicly available sources including investor memos, press coverage, and company documentation. I have not received a briefing from Cohere and this post does not reflect any proprietary or non-public information.

Sundays are when I follow up on companies that never show up in the hype cycle. This week the question was simple: whatever happened to Cohere? No consumer product. No viral benchmark moment. No drama. I had not heard much about them in months, which usually means one of two things: the company is struggling, or it is too busy closing enterprise deals to care about press.

It turned out to be the latter. Cohere's February 2026 investor memo reported $240 million in annual recurring revenue for 2025, exceeding its internal target of $200 million, with quarter-over-quarter growth above 50% throughout the year. Gross margins averaged 70%, expanding year over year. Following a $600 million funding round that included Nvidia, Salesforce, and AMD, the company is now at a $7 billion valuation and openly discussing a public market debut.

Deployment as the Core Argument

Cohere runs on AWS, Azure, Google Cloud, and Oracle Cloud. That is not the differentiator. The differentiator is that it also runs inside a customer's own environment, without any dependency on a hyperscaler. Model Vault deploys models inside a dedicated, logically isolated virtual private cloud. Beyond that, Cohere offers fully on-premises and air-gapped options for organizations where even a managed cloud environment does not meet data residency requirements. These are distinct deployment tiers, not the same product.

For a bank, a government agency, or a healthcare system operating under strict data governance rules, this matters more than benchmark rankings. The question is not which model is most capable in a controlled test. It is which model can run entirely inside a controlled environment, with sensitive data never crossing an external network. That use case has real buyers, and Cohere has been building toward it deliberately. In January 2025, the company launched North for Banking with Royal Bank of Canada, a secure generative AI platform designed specifically for financial services productivity and data security.

The hardware strategy reinforces this posture. By securing investment from both Nvidia and AMD, Cohere runs across a wider range of infrastructure than a company tied to a single chipset. That matters for procurement teams managing existing hardware investments or operating in environments with specific infrastructure requirements.

What Cohere Actually Sells

The product focus is narrow by design. Cohere is optimized for retrieval-augmented generation, semantic search, and document understanding, not general-purpose chat or consumer interaction. Its Command model family, the Embed models for semantic representation, and the Rerank model for improving search relevance are all built around the same enterprise premise: connecting the model to a company's internal data and returning verifiable, cited answers rather than generated responses from training memory.

This distinction matters when it comes to accuracy. Standalone language models, including Cohere's, hallucinate. A 2023 study by Arthur AI found Cohere's model hallucinated more than GPT-4 and Claude in isolated tests. Cohere disputed the methodology, noting that the test did not include their retrieval-augmented generation stack, which grounds answers in the enterprise's own documents and provides citations to verify the source. The argument is not that the model is inherently more accurate. It is that the system, when deployed with retrieval on top of internal data, produces answers that can be checked. That is a meaningful difference for regulated industries where auditability is a requirement, not a preference.

In 2025, Cohere extended this further with the North platform, an enterprise AI workspace that combines retrieval, search, and agent orchestration inside the same secure environment. The company also launched Tiny Aya, a multilingual model family supporting over 70 languages capable of running locally on devices without an internet connection, and Rerank 4, with a 32,000-token context window built for complex enterprise document retrieval. These releases point to a consistent strategic logic: AI that runs where the data already lives, in the language the workforce already uses.

The Question Worth Asking

One investor analysis described Cohere's position as the "Intel Inside" of enterprise AI, meaning a company content to power other organizations' workflows without requiring its brand to be visible. That framing captures something real. Cohere's commercial model is not built on consumer recognition. It is built on fitting into existing procurement structures, security frameworks, and infrastructure decisions that enterprises have already made.

The viability question for a senior technology buyer is this: if your AI deployment requires that data stay within your controlled environment, that outputs be traceable to source documents, and that the vendor not be commercially dependent on a single cloud relationship, which companies on your shortlist can actually deliver all three? Cohere's $240 million revenue base, 70% gross margins, and cloud-agnostic investor structure suggest it belongs on that list. Whether its model capabilities keep pace with the frontier as it heads toward a public offering is the open question worth watching.