COHERE

Aidan Gomez was 21 years old when he co-authored 'Attention Is All You Need' — the 2017 Google Brain paper that introduced the Transformer architecture and became the foundation for essentially every large language model that exists today, including GPT-4. He was an intern at the time.

That paper has been cited over 100,000 times. Most people who write papers like that spend the next decade in academia collecting citations.

Gomez did something different — he went back to the University of Toronto, finished his degree, and started thinking about what to build with what he'd helped invent.

In 2019, he teamed up with Nick Frosst, another Google Brain alumnus and Geoffrey Hinton's former research assistant, and Ivan Zhang, a strong engineering mind from the University of Toronto AI scene. Geoffrey Hinton — the so-called 'godfather of AI' — was an early advisor and investor.

Having the man who basically invented deep learning in your corner when you're raising your seed round is not nothing.

From the start, the three founders were deliberate about one thing: they were not building a consumer product. No chatbot.

No flashy demo. No 'we're going to change humanity' press release.

They wanted to build the API layer that enterprises would use to add language AI to their existing products and workflows — quietly, reliably, and without the drama that tends to follow OpenAI into every room it enters.

Cohere sells access to large language models via API. Companies integrate Cohere's models into their own products — customer support tools, document summarization, internal search, content generation, compliance workflows — and pay based on usage.

Think of it like AWS but for language AI. You don't have to train a model from scratch.

You don't have to hire a team of ML engineers. You call the API, you get smart text output, you pay per token.

The key differentiator is deployment flexibility. Cohere will let enterprises run their models on their own cloud, on-premises, or in a private virtual cloud — meaning the company's data never touches Cohere's servers.

For a bank, a hospital, or a law firm, that's not a nice-to-have. It's the whole ballgame.

OpenAI and Google are great, but try telling your legal team that your sensitive contract data is going through a shared API endpoint somewhere in Microsoft Azure. That conversation ends badly.

Cohere also makes money through Cohere For AI, its research division, and through enterprise contracts that go well beyond pure API access — implementation support, fine-tuning on proprietary data, and dedicated model deployments. The enterprise sales motion is slower than consumer viral growth, but the contracts are larger and the churn is lower.

Enterprises that rebuild their internal tools around your models don't leave easily.

Command is Cohere's flagship text generation model — the thing you call when you want an enterprise-grade LLM that will summarize documents, draft responses, classify content, or answer questions based on your company's data. Command R and Command R+ are the retrieval-augmented generation variants, optimized specifically for RAG pipelines where the model needs to search through large document stores and generate accurate, grounded answers.

For enterprises trying to build internal knowledge bases that actually work, Command R+ is the serious option.

Embed is Cohere's embedding model — it converts text into numerical vectors that can be stored and searched semantically. This is the engine behind smart search products.

Instead of keyword matching, you get meaning matching. Ask 'what's our refund policy for international orders' and it finds the right document even if it never uses those exact words.

Rerank is the complement — a model that takes a list of candidate search results and re-orders them by relevance. Together, Embed and Rerank power the retrieval layer of most enterprise AI search stacks.

Cohere's Coral product is their enterprise AI assistant interface — a deployed, private ChatGPT-like experience that companies can give to their employees, connected to internal documents and data sources, with all the privacy guarantees that enterprise IT departments require. It's the product that shows what all the API infrastructure actually looks like when it's assembled into something a non-engineer can use.

Cohere's growth strategy is essentially the anti-OpenAI playbook. While OpenAI was releasing ChatGPT to 100 million users and watching the world lose its mind, Cohere was calling procurement departments.

That's not a joke — it's a deliberate wedge.

The counterintuitive move was going deep on data privacy and deployment flexibility before most enterprises even knew they needed it. When the enterprise AI spending wave hit in 2023, Cohere had already built the infrastructure that allowed companies to run models in their own environment.

Competitors were scrambling to add this capability. Cohere had it from day one because they'd designed around enterprise requirements from the start.

They also leaned hard into the fact that Aidan Gomez is a literal co-author of the Transformer paper. In a space full of companies claiming AI credibility, having the person who helped invent the underlying architecture as your CEO is a sales asset that money can't buy.

Enterprise buyers care about trust and technical credibility in a way consumer users don't — and Cohere has it in abundance.

Strategic partnerships accelerated the growth. Oracle, Salesforce Ventures, and NVIDIA all backed Cohere, which isn't just about the money.

Oracle plugging Cohere into its enterprise cloud suite and NVIDIA optimizing Cohere's models for its hardware creates distribution that no amount of cold outreach could replicate.

The competitive pressure from OpenAI, Google, Anthropic, and Meta is genuinely brutal. All four have more resources, more brand recognition, and in Meta's case, a strategy of releasing models for free that undercuts the entire paid API market.

When Meta drops Llama 3 as open source, every enterprise CTO's first question is: 'Why are we paying for this again?' Cohere's answer — security, support, deployment flexibility, reliability — is a good answer, but it has to be delivered over and over again in every sales cycle.

There's also the model quality treadmill problem. Every few months, a new model benchmark drops and the leaderboard reshuffles.

Staying competitive on raw model performance while also servicing enterprise clients and maintaining deployment infrastructure is an expensive balancing act. Cohere is a well-funded startup, but Google has essentially unlimited compute budget and treats AI as an existential priority.

That's a hard race to win on pure capability.

And then there's the 'Cohere Health problem' — a completely separate healthcare company called Cohere Health that constantly confuses search results, journalists, and investors. It's a minor nuisance that somehow never gets less annoying.

Geoffrey Hinton was an early advisor and investor — the godfather of deep learning putting his name behind Cohere before the company had a product gave it immediate credibility in a space where credibility is everything. That association alone opened doors that most AI startups spend years trying to knock on.

The Series C in 2023 was the headline round — $270 million at a $2.1 billion valuation, led by Inovia Capital with participation from Oracle, Salesforce Ventures, NVIDIA, and others. The strategic investors matter more than the numbers here.

Oracle integrating Cohere into Oracle Cloud Infrastructure and Salesforce Ventures backing the company means two of the largest enterprise software companies in the world have a financial incentive to sell Cohere's technology to their customers. That's a distribution channel that startups dream about.

The $500 million round in 2024, pushing the valuation to $5.5 billion, included PSP Investments, AMD, and various sovereign wealth and pension funds — the kind of institutional capital that signals this is a serious long-term build, not a flip. AMD's participation is notable: a chip company investing in an AI model provider is a bet that Cohere's success sells more AMD GPUs.

The alignment of incentives runs all the way down the stack.