Cohere’s Model Vault Beta Wants Your Sensitive Data Without Ever Seeing It

There is a sentence that security teams have been repeating for years: the data you cannot protect is the data you cannot use. Hospitals sit on clinical records. Banks sit on fraud investigations. Law firms sit on privileged documents. All of it could be put to work with AI, and almost none of it is, because feeding it to a model means handing plaintext copies to infrastructure you do not control.

Cohere wants to remove that last excuse. On October 22, 2026, the Toronto AI company announced the public beta of confidential computing on Model Vault, its dedicated model inference platform. The pitch is simple enough to explain in one line: your data stays encrypted from the moment it leaves your systems until it comes back with an answer. Nobody in the middle gets to read it.

What Model Vault actually does

Model Vault is Cohere’s fully managed platform for running its models in a dedicated deployment, aimed at enterprises that cannot share infrastructure with other customers. The new confidential computing option wraps that deployment in hardware-enforced isolation. The beta runs on NVIDIA H100 GPUs inside Microsoft Azure’s confidential computing infrastructure.

The mechanics break down into three parts. Azure provides the underlying confidential computing environment. NVIDIA’s confidential computing technology carries that protection onto the GPUs themselves, so data stays shielded while the model is actively running on it. Cohere supplies the managed serving layer, an encrypted request path, and remote attestation, which is the part where your software checks signed proof of what is running on the other end before sending anything.

The workflow goes like this. A customer sets up an encrypted vault through Model Vault and connects it through the same Cohere APIs they already use. Before each request, the customer’s software checks signed proof of what is running on the other end against an approved security policy. If the environment fails the check, the client refuses to send the request. Request contents stay encrypted as they pass through intermediary systems and are decrypted only inside the verified confidential environment.

Who this is actually for

Cohere names a fairly specific list of workloads where security, contractual, or regulatory rules have historically kept data out of managed AI platforms. Fraud investigations in financial services. Privileged document search and due diligence for legal teams. Sensitive public-sector information. Patent and research material. Proprietary engineering design and architecture documents. Search and analysis across proprietary source code.

These are cases where the obstacle was never model quality. A bank’s fraud team knows what AI could do with its transaction histories; the problem was always that “could” came with a compliance nightmare attached. Confidential computing does not erase the compliance work, but it shrinks the circle of trust to a single verified machine rather than a chain of intermediaries.

Cohere is also careful about what the technology does not promise. The company says confidential computing does not eliminate every risk, and it does not replace application security, compliance checks, contractual controls, or data-residency requirements. It provides a hardware-enforced, verifiable boundary around data and computation during inference. That honesty matters, because the first wave of confidential computing marketing tended to imply a magic shield, and buyers are wiser now.

Why the timing is not an accident

This launch sits inside a much bigger industry pattern. Data has spent the last decade encrypted at rest and encrypted in transit, with a gaping hole in the middle: while a model processes your data, it sits in plaintext in memory.

NVIDIA is pushing the same direction from the hardware side. The same week as the Model Vault beta, NVIDIA announced Confidential Bare Metal capabilities, with Cohere named as a company planning to build them into Model Vault, using NVIDIA’s trusted execution environments to carry protection down to the GPU itself. According to reports, Cohere’s chief technology officer Phil Blunsom said the goal is to extend hardware-rooted security to the GPU, protecting customer data in memory and across interconnects during execution. NVIDIA framed confidential computing as the infrastructure layer under its OpenShell reference architecture for AI agents, also announced that week.

The competitive logic is worth spelling out. For enterprise AI, the differentiator is increasingly not who has the smartest model but who can be trusted with the most sensitive data. Regulated industries hold enormous amounts of unstructured data, and the vendors that crack the trust problem get first access to that market.

What to watch next

Public beta is not general availability. The technology is available now, and customers can start at vault.cohere.com, but beta features in security infrastructure get scrutinized heavily before anyone bets a compliance certification on them. The real test will be whether auditors and regulators accept the attestation evidence as a substitute for the physical and contractual controls they currently demand.

There is also the question of scope. The beta runs on NVIDIA H100 GPUs on Azure. Enterprises with commitments to other clouds, or with on-premise requirements, will be watching for expansion.

For Canada’s AI scene, this is another brick in a wall that has been going up fast. Cohere raised $500 million in August at a $6.8 billion valuation, according to reports, signed a deal to acquire the AI cybersecurity startup AISLE in October, and is now shipping enterprise-grade security infrastructure. The national story is AI sovereignty, keeping Canadian-built AI in Canada. The product story is simpler: sell to the banks, hospitals, and governments that were told AI was not safe for their data, and prove that wrong.

Frequently Asked Questions

What is confidential computing on Cohere Model Vault?

It is a public beta feature, launched October 22, 2026, that lets enterprises run AI inference on Cohere’s Model Vault platform while their data stays encrypted during processing. It uses NVIDIA H100 GPUs on Microsoft Azure, with remote attestation that lets customer software verify the processing environment cryptographically before sending any data.

Does confidential computing mean my data is completely safe?

No. Cohere itself cautions that confidential computing does not eliminate every risk and does not replace application security, compliance checks, contractual controls, or data-residency requirements. It provides a hardware-enforced, verifiable protection boundary around data and computation while inference takes place, reducing how much infrastructure you must trust with plaintext data.

Who can use the Model Vault confidential computing beta?

The beta is available now through Model Vault, with customers starting at vault.cohere.com. It is aimed at enterprises with sensitive workloads, including financial services fraud investigations, privileged legal document search, sensitive public-sector information, patents and research, and proprietary source code analysis.