Cohere Turns Privacy Into a Product by Hosting Rival AI Models
Cohere just opened its secure Model Vault to rival open models, proving that enterprise AI is now about data security, not just model power.
Business · Source: Cohere
What happened
Cohere is now hosting third-party open-weight models inside its Model Vault service. The company added models from major rivals including Google Gemma 4, OpenAI GPT-OSS, Microsoft Phi, and DeepSeek V4.1 Flash. Other additions include GLM, MiMo, Nemotron, and Hy3. This means enterprise customers can run competing models right alongside Cohere's own products on dedicated infrastructure.
The core pitch is strict data security and sovereign inference. Model Vault uses single-tenant infrastructure with zero data retention by default. A higher tier called Model Vault Encrypted uses confidential computing to keep data encrypted even during active processing. Not even Cohere or the underlying cloud provider can see the prompts or the outputs.
Availability depends heavily on the region, the hardware, and the specific model. The standard vault runs in the United States, the European Union, and Canada. The encrypted tier is currently restricted to the US, with EU and Canada support coming later. Notably, models like DeepSeek, Nemotron, GLM, MiMo, and Hy3 are excluded from the encrypted tier entirely.
Key facts
- US$4.25 — Promotional rate per H100 GPU-hour for standard and encrypted deployments through March 31, 2027
- DeepSeek V4.1 Flash — One of the third-party models added, though it is not supported on the Encrypted tier
- B200 or L4 — The specific processors supported for European Union deployments
Why it matters
Builders no longer have to choose between the specific open model they want and the strict security compliance their enterprise clients demand. Managing inference for open or on-premise models is historically difficult due to severe capability and capacity constraints. You can now deploy a Microsoft or Google model inside Cohere's walled garden. This solves a massive headache for founders selling AI agents to banks, hospitals, or government departments who flatly refuse to send sensitive data to shared APIs.
Trust is rapidly becoming the actual product in the enterprise AI space. Model builders are realizing that having the smartest weights does not matter if large organizations refuse to use them due to data leak fears. By becoming a secure host for its own competitors, Cohere is shifting its business model. They are moving from simply selling artificial intelligence to selling highly secure, sovereign infrastructure where companies control exactly where their data lives and who can access it.
For builders
Sell to highly regulated enterprise clients
You can now build products using top open models without failing enterprise security audits. Healthcare and finance clients pay a massive premium when you can guarantee zero data retention and confidential computing. Cohere handles the hosting and scaling so your team does not have to build a custom cluster.
Navigate strict regional hardware limits
You lose deployment flexibility depending on where your customers live. EU deployments only support models running on B200 or L4 processors. Canada supports GB200 or L4, but Canadian GB200 access requires direct contact with Cohere sales. Plan your architecture around these physical bottlenecks.
Budget for dedicated GPU hours
Cohere charges based on graphics processing unit type, deployment size, and total running time. They offer a promotional rate of $4.25 per H100 GPU-hour for standard and encrypted deployments through March 31, 2027. Published prices do not guarantee GPU availability, so you carry the cost and risk of securing compute.
My take
Cohere inviting competitors into its own house is a brilliant pivot. They figured out that enterprise fear of data leaks is much stronger than the desire for the absolute smartest model. The AI race is no longer just about who builds the best brain. The real money is going to whoever builds the safest vault.
Original reporting: Cohere. This is my rewrite and opinion.