Cohere AI is a Canadian enterprise AI company that builds large language models, embedding models, and agentic AI platforms for businesses and governments.
Unlike OpenAI, which chases consumer adoption with ChatGPT, or Anthropic, which positions itself on safety, Cohere made a contrarian bet from day one: skip the consumer hype, sell privately-deployed AI to banks, telcos, and governments. In 2026, that bet is paying off.
The company was founded in Toronto in 2019 by Aidan Gomez, Nick Frosst, and Ivan Zhang. Gomez was a co-author of the 2017 paper “Attention Is All You Need”—the paper that introduced the Transformer architecture that underpins virtually every modern large language model. He was 20 when he helped write it.
But here’s what the marketing materials won’t tell you: Cohere is not trying to beat OpenAI on benchmarks. Its pitch to investors isn’t “we will beat GPT-5.” It’s that sovereign deployment, multilingual depth, and improving gross margins represent a durable business category.
“Cohere is a very low drama company,” Chief AI Officer Joelle Pineau told AFP, dismissing theorizing about AGI as a distraction.
In an industry where CEOs compete to predict the arrival of superintelligence, this is almost radical.
The Transformer Origins: Why the Founders Matter
The “Attention Is All You Need” paper is the single most influential AI paper of the past decade. Every major LLM—GPT, Claude, Gemini, Llama—traces its lineage back to that work. Gomez co-authored it.
But the often-repeated description of all three Cohere founders as “former Google Brain researchers” is slightly imprecise. Gomez and Frosst worked at Google Brain; Zhang was the third technical co-founder. The distinction matters because it reflects Cohere’s positioning: deeply technical, but not a Google spin-off.
Cohere is backed by investors including Nvidia, AMD, and Salesforce, and its valuation has swelled to roughly $7 billion.
Cohere’s Product Portfolio: What You Can Actually Use
Cohere’s product lineup breaks into four categories:
1. Command Family (Generative Models)
- Command R and R+: Previous-generation models optimized for RAG (retrieval-augmented generation)
- Command A+: 218B total / 25B active parameter MoE model, released May 2026
- Command A: Earlier generation
2. Embed and Rerank (Search & Retrieval)
3. North Platform (Agentic AI)
- Enterprise workspace for building and deploying AI agents
- North Mini Code: Coding-focused agentic model
4. Specialized Models
- Aya: Multilingual models
- Transcribe: Speech recognition
- Tiny Aya: Runs locally on laptops and edge devices
Command A+: The 218B-Parameter Open-Weight Milestone
Command A+ is Cohere’s first Mixture-of-Experts model and one of the most significant open-weight releases from a Western AI company—released under Apache 2.0 license in May 2026.
Here’s the technical breakdown:
| Specification | Detail |
|---|---|
| Total parameters | 218 billion |
| Active parameters per token | ~25 billion |
| Experts | 128 experts, 8 active per token, plus 1 shared expert |
| Context window | 128K input, 64K output |
| Input modalities | Text, image, tool use |
| Languages supported | 48 languages |
| License | Apache 2.0 |
| Minimum hardware | 1× B200 or 2× H100s |
The MoE architecture is the key innovation. Instead of activating all 218B parameters for every token, the model routes each query to only the most relevant “expert” networks—about 25B parameters. This means it has the knowledge capacity of a massive model but the runtime compute cost of a much smaller one.
The deployment story is even more interesting. Cohere offers the model in several near-lossless quantizations. The result? A 218B-parameter model that can run on a single NVIDIA B200 or two H100s.
“过去一个千亿级模型要一整个GPU集群伺候，现在一台机器搞定。” — 36Kr on Command A+
But there’s a catch. The “single GPU” claim refers to a data-center-grade Blackwell B200, not a consumer card. Still, the efficiency gains are real.
The licensing is what separates Command A+ from almost everything else. Cohere previously released models under CC-BY-NC (non-commercial) licenses. Command A+ is Apache 2.0—full commercial use, no restrictions. This is a significant shift and a direct challenge to OpenAI and Anthropic’s closed models.
Command A+ is built for actual enterprise workloads: complex RAG pipelines, multi-step SQL generation, and high-stakes financial document analysis. It’s optimized for the realities of enterprise deployment, where reliability, cost efficiency, and operational performance determine whether a system can be deployed at all.
North: Cohere’s Bet on Agentic AI
North is Cohere’s enterprise platform for building and deploying AI agents—and it’s where the company sees its future growth.
The platform enables organizations to connect AI models to enterprise applications, workflows, and data sources. Think of it as a secure workspace where employees can deploy AI agents that automate real work.
Key Capabilities
- Multi-agent automations: Triage issues, generate documentation, drive releases across teams
- Secure deployment: Organizations can build and deploy AI agents behind their own firewalls, ensuring data privacy and total control
- Industry-specific solutions: North powers use cases across financial services, energy, public sector, healthcare, and technology
Enterprise Adoption
- TKMS (German naval shipbuilder) signed an eight-figure contract to deploy North across its organization
- Aston Martin F1 Team partnered with Cohere, giving all employees access to North
- S&P Global integrated its financial data directly into North
In March 2026, Cohere announced it would make an optimized version of North available on NVIDIA DGX Spark, bringing a high-performance, locally run enterprise workstation to Spark owners.
Financial Reality: $240M ARR and an IPO on the Horizon
Industry reports indicate that Cohere reached approximately $240 million in annual recurring revenue in 2025, surpassing its $200 million target with 50%+ quarter-over-quarter growth.
The Growth Story
| Metric | 2025 |
|---|---|
| ARR | ~$240 million |
| Gross margin | ~70% |
| Valuation | ~$7 billion |
| QoQ growth | 50%+ throughout the year |
85% of revenue comes from private and on-premise deployments. This is the opposite of OpenAI’s model, which relies heavily on API usage and consumer subscriptions.
Reports suggest Cohere’s gross margins have approached around 70% as enterprise deployments scale, expanding by 25 basis points year-over-year.
The comparison that matters: OpenAI at ~$25B ARR trades at roughly 34x. Anthropic at $30B ARR trades at roughly 27x. Cohere at $240M ARR trades at roughly 29x on its $7B post-money. All three are trading at roughly the same revenue multiple—the market is pricing them on category, not scale.
The IPO Question
CEO Aidan Gomez said in October 2025 that the startup hopes to make its public market debut “soon.” He told Bloomberg that he thinks investors would welcome a “pure play AI investment opportunity.”
If “soon” means in 2026, Cohere may be contending against OpenAI, Anthropic, and SpaceX/xAI, which are all reportedly weighing their own public debuts.
The Aleph Alpha Merger: Sovereignty as Strategy
In April 2026, Cohere merged with Germany’s Aleph Alpha, creating a transatlantic AI company with dual headquarters in Toronto and Berlin.
The deal is directly supported by both the Canadian and German governments—the two countries’ respective Digital Ministers attended the announcement in Berlin. Both governments view the merger as a decisive step for the technological sovereignty of both nations.
“Cohere and Aleph Alpha are joining forces—a strong signal for Germany as an AI hub. What is emerging here is a German-Canadian AI model: secure, sovereign, competitive.” — German Digital Minister Karsten Wildberger
The combined entity aims to deliver a secure alternative for customized AI in highly-regulated sectors including the public sector, finance, defense, energy, manufacturing, telecommunications, and healthcare.
Cohere shareholders hold approximately 90% of the combined company, with Aleph Alpha shareholders receiving around 10%.
Schwarz Group (owner of Lidl and Kaufland) committed $600 million to lead Cohere’s Series E.
Cohere vs. OpenAI vs. Anthropic: The Real Comparison
Most comparisons of Cohere to OpenAI or Anthropic miss the point. They’re not playing the same game.
| Factor | Cohere | OpenAI | Anthropic |
|---|---|---|---|
| Primary market | Enterprise (private/on-prem) | Consumer + Enterprise API | Enterprise + Safety |
| Business model | Private deployments, custom | API, subscriptions | API, subscriptions |
| Consumer product | None (North is enterprise) | ChatGPT (large global consumer user base) | Claude (consumer + enterprise) |
| Data privacy | Can run air-gapped | Shared infrastructure | Shared infrastructure |
| Open source | Apache 2.0 (Command A+) | Closed | Closed |
| Sovereign AI | Core strategy | Not a focus | Not a focus |
| Gross margin | ~70% (reported) | ~40-50% (est.) | ~40-50% (est.) |
Cohere’s weakness: Smaller ecosystem and less mainstream brand recognition than OpenAI, Anthropic, or Google. For risk-averse procurement decisions, brand awareness matters.
Cohere’s strength: No Microsoft-entangled compute commitments, no headline-driving safety incidents, and enterprise ARR concentrated in regulated industries that don’t churn. Cohere is well-regarded among enterprise technical buyers but lags in broader AI brand awareness.
Cohere’s thesis suggests that durable value often lies not in owning the foundation model, but in owning the workflow built on top of it.
Security and Sovereignty: The Actual Moat
Cohere’s primary competitive advantage isn’t model performance—it’s the ability to deploy AI anywhere, on any infrastructure, with maximum data control.
The company offers deployment across:
- Hosted APIs
- Third-party cloud platforms (AWS, Azure, Oracle)
- Customer-controlled virtual private clouds
- On-premise installations
- Air-gapped environments
This flexibility matters for regulated industries. Financial services, healthcare, and government customers can’t send sensitive data to public APIs. Cohere lets them run models on their own infrastructure.
Sovereign AI Features of Command A+
- Greater transparency through open-weight availability: Full visibility into model architecture and behavior
- Total data sovereignty: On-premises and private cloud deployment with no external data transmission
- Regulatory alignment: Control and deployment model designed to enable alignment with evolving global compliance and AI governance requirements
- No vendor lock-in: Predictable costs with no licensing restrictions or forced ecosystem dependencies
The company now has dual headquarters in Toronto and Berlin, with offices in San Francisco, New York, London, Montreal, Paris, and Seoul.
Why Cohere Matters in the AI Race
Cohere represents something increasingly rare in AI: a company that has deliberately chosen scale over hype.
While OpenAI and Anthropic compete on model size and benchmark scores, Cohere has focused on what actually matters for enterprise customers—reliability, security, and cost predictability. The company’s “ROI over AGI” philosophy isn’t just marketing; it’s reflected in product decisions like Command A+’s quantization support and North’s private deployment model.
The company’s success suggests a broader trend: the AI market is segmenting. There’s room for consumer AI giants, but there’s also a growing market for sovereign, private, enterprise-grade AI that doesn’t depend on US hyperscalers. Cohere is betting that this market will be large enough to support a standalone public company.
Who Should Use Cohere (And Who Shouldn’t)
Use Cohere if:
- You’re a large enterprise in a regulated industry (finance, healthcare, government, defense)
- You need to deploy AI on-premise or in a private cloud
- You have engineering resources to build custom solutions
- Data sovereignty is a non-negotiable requirement
- You want to avoid vendor lock-in with US hyperscalers
Avoid Cohere if:
- You’re a solo developer or small team looking for a plug-and-play solution
- You want a consumer-facing chatbot (Cohere doesn’t have one)
- You need the absolute best model performance on every benchmark
- You’re looking for simple, transparent pricing (custom enterprise pricing is opaque)
- You want to experiment quickly without engineering overhead
Pricing Transparency
Cohere’s pricing is complicated, and that’s being generous.
API Pricing (Command R+)
- Input: $2.50 per 1M tokens
- Output: $10.00 per 1M tokens
Enterprise Products (North)
- No public pricing—you must contact sales
The Real Cost Reality
Command R+ costs 44% less than GPT-4o. But “less than GPT-4o” still isn’t cheap. Pricing varies significantly depending on the deployment method (e.g., API, private cloud, on-premise) and enterprise agreements.
The broader point: if you’re an enterprise, you’re negotiating. If you’re a developer, you’re paying API rates that are competitive but not the cheapest.
Final Verdict
Cohere AI is the most underrated enterprise AI company in the world—and one of the strongest candidates for an AI IPO in 2026.
It’s not the flashiest. It doesn’t have a consumer chatbot. It doesn’t make bold claims about AGI. But it has something most AI labs don’t: a sustainable business model, real enterprise revenue, and improving gross margins.
The company’s contrarian bet—skip the consumer hype, sell privately-deployed AI to banks, telcos, and governments—looks increasingly prescient. While OpenAI and Anthropic compete on benchmarks and consumer adoption, Cohere is quietly building a durable business in regulated industries that value security over speed.
Key Takeaways
- Cohere is an enterprise-first company—85% of revenue comes from private and on-premise deployments
- Command A+ is a genuine milestone—218B parameters under Apache 2.0, runs on a single B200 or two H100s
- North is the future—agentic AI platform for enterprises, already deployed across financial services, energy, public sector, and healthcare
- IPO is likely in 2026—~$240M ARR, 50%+ QoQ growth, ~70% gross margins
- Security is the moat—air-gapped deployment, sovereign AI positioning, no vendor lock-in
- Not for everyone—requires engineering resources, not a consumer product
Who Should Use It
Large enterprises in regulated industries that need AI they can control. Banks, healthcare providers, government agencies, and defense contractors.
Who Shouldn’t
Developers and small teams looking for a quick, cheap, plug-and-play AI solution. Cohere requires engineering investment.
Next Steps
- Try Command A+ on Hugging Face
- Explore the North platform if you’re an enterprise
- Watch for the IPO—it’s coming
FAQs
1. Is Cohere AI a Canadian company?
A: Yes. Cohere was founded in Toronto in 2019 by Aidan Gomez, Ivan Zhang, and Nick Frosst. It is co-headquartered in Toronto and San Francisco, with offices in Palo Alto, London, and New York.
2. Who is the CEO of Cohere AI?
A: Aidan Gomez, who co-founded the company in 2019. He was a co-author of the “Attention Is All You Need” paper, which introduced the Transformer architecture.
3. Is Cohere AI profitable?
4. What does Cohere make?
A: Cohere builds large language models (Command family, including Command A+), embedding and rerank models for search, the North enterprise AI platform for agentic AI, and specialized models for multilingual and speech applications.
5. Does Cohere AI have stock?
6. Cohere AI vs OpenAI: which is better?
A: It depends on your use case. Cohere is better for enterprises that need private, on-premise deployment and data sovereignty. OpenAI is better for consumer applications, broad API access, and cutting-edge model performance.
For regular updates about the AI, AI tools, AI guides, technology & more. Visit, The TechUnfoldedAi.com.
Leave a Reply