When incumbents chase the consumer mass market tethered to one platform, serve the data-sensitive customers they structurally cannot, bringing your product to the customer's environment instead of demanding their data come to you.
Cohere
Cohere is an enterprise LLM company founded in 2019 by three former Google Brain researchers, including Aidan Gomez, a co-author of the 2017 "Attention Is All You Need" paper that introduced the Transformer. It is not the fastest-growing AI lab: OpenAI and Anthropic, which each raised tens of billions and built consumer flagships, dwarf Cohere's roughly $240 million in reported 2025 revenue. Cohere's move is the refusal itself. It declined the consumer and frontier race those labs run, and sold enterprise-only, data-private, deployable models into a segment the incumbents structurally cannot follow it into.
Counter-Positioning: refusing the consumer race the incumbents cannot exit
The problem. By 2023 the AI narrative and the capital were flowing to consumer flagships, ChatGPT and later Claude, and to ever-larger frontier models, each tethered to a hyperscaler (OpenAI to Microsoft Azure, Anthropic to Amazon and Google). A startup could not win that fight; it had neither the tens of billions nor the distribution. But the same structure walled the giants off from a segment: banks, telcos, insurers, and governments that legally cannot send sensitive data to a consumer-tuned, single-cloud API.
The approach. Cohere refused to build a consumer app or monetize through engagement, and refused to tie itself to one cloud. It sold what it called "a security-first category of enterprise AI that is simply not being met by repurposed consumer models," running inside the customer's own environment via cloud API, VPC, or fully on-premises, bringing the model to the data rather than the data to the model. Canadian domicile (subject to PIPEDA, not the US CLOUD Act) and cloud-neutrality were the trust signals. This is counter-positioning in the strict sense: OpenAI cannot copy "no consumer product, any cloud, deploy in your VPC" without undercutting the ChatGPT and Azure businesses that fund it.
How it solved it. The refusal kept Cohere capital-light and let it own the segment the giants declined. It reached a reported $240 million ARR by the end of 2025, up about 287% from roughly $62 million a year earlier by Sacra's estimate, on gross margins near 70% and with an estimated 85% of revenue from private deployments. In August 2025 it raised about $500 million at a $6.8 billion valuation, still with no consumer product. The counterfactual is the point: the better-funded labs grew faster in raw dollars, so raw growth is not the win. Cohere's win is structural, a business the incumbents could not chase without cannibalizing their own.
Differentiation: RAG on private data, not one bigger frontier model
The problem. Cohere could not out-spend Google, OpenAI, or Anthropic on raw model scale, and benchmark cleverness was not a durable edge. Regulated buyers cared less about a chatbot's IQ than about grounding answers in their own documents without those documents leaking or training someone else's model.
The approach. Cohere specialized in retrieval-augmented generation and shipped a coordinated suite instead of one monolith: Command R in March 2024 and Command R+ in April 2024, optimized for RAG and tool use with a 128,000-token context and strength across ten major business languages, alongside its Embed and Rerank models. It backed this with a contractual commitment of no access to data processed in private or third-party-managed deployments. This is where Cohere separates from Mistral: Mistral sells open-weight models on European sovereignty, while Cohere sells a closed but privately deployable RAG stack and positions as the enterprise's retrieval partner, not an open-weight download.
How it solved it. RAG-plus-privacy won exactly where the consumer-tuned labs were weakest. Cohere's retrieval models landed inside products from Oracle and Notion, and it shipped North, its secure enterprise platform, on-premises through the Dell AI Factory in May 2025. That an estimated 85% of revenue comes from private deployment, at roughly 70% gross margin, is the evidence: enterprises are paying for the deployable-and-grounded stack, not for model size.
GTM: channel partners into regulated buyers
The problem. Selling into banks, governments, and large enterprises requires trust, industry compliance, and procurement relationships a young startup does not have. Building that buyer by buyer against incumbents already embedded in enterprise IT would be slow and expensive.
The approach. Cohere went to market largely through strategic channel partners and co-built vertical products rather than a direct-sales army. Oracle (June 2023) embedded its models in Fusion Cloud and NetSuite; McKinsey (July 2023) handled enterprise integration; Fujitsu and LG co-developed the Takane Japanese and a Korean LLM. It tailored North for Banking with RBC in January 2025 and shipped through the Dell AI Factory.
How it solved it. The partner motion converted into marquee multi-year contracts and strategic investment from the same partners. The August 2025 round of about $500 million at roughly $6.8 billion drew Oracle, Salesforce, NVIDIA, AMD, and Cisco, several of them also distribution channels. Riding partners' installed bases into RBC, Bell, LG CNS, and the Canadian government reached regulated buyers far faster than a cold direct-sales effort could.