Data privacy and security have long been major concerns for enterprises, and they've been elevated in recent weeks as the two largest U.S. generative AI startups, OpenAI and Anthropic, faced public disputes over exactly what their systems saw and retained from customers.
Now, Canadian AI startup Cohere says it can do enterprises one better, assuring them their data remains private and off-limits to itself and cloud providers using Model Vault, the single-tenant inference product it already sells, and a technique called confidential computing.
First, readers should understand that even outside of Cohere, enterprise data has been protected at rest and in transit for years. Encryption locks down a database on disk. TLS locks down a request crossing the network.
But the moment a prompt is actually processed by an AI model, that protection disappears. It sits in plaintext in memory, readable by the host operating system, the hypervisor and anyone with privileged access to the machine it runs on. For security teams handling regulated or proprietary data, that gap has forced a choice between the convenience of hosted inference and the control of running the infrastructure themselves.
Cohere's solution runs the workload inside hardware that encrypts and isolates its memory, so neither the cloud provider nor Cohere itself can read what is inside. According to the company, that protection now covers the moment a model actually reads the prompt, not just the moments before and after, and the capability went live today.
"Confidential Computing support in Model Vault is newly available," Manoj Govindassamy, Cohere's director of serving inference told VentureBeat. "As of today, there's no additional cost for Model Vault customers, pricing remains the same as the standard Model Vault offering."
The launch lands the same day Cohere signed its definitive merger agreement with Germany's Aleph Alpha, a roughly $20 billion deal that creates dual headquarters in Toronto and Berlin.
Cohere CEO Aidan Gomez pitched the combined company around trust and governability rather than raw model capability, calling it AI "powerful enough to compete, but secure and governable enough to trust." The two announcements are not connected beyond the calendar, but they lean on the same argument.
Why enterprise data privacy concerns are rising alongside AI use
The timing sits against a backdrop where what a vendor can see, not just what it promises not to look at, has become the live question in enterprise AI procurement.
According to reports this month, Nvidia, Palantir and Booz Allen Hamilton have restricted their use of Anthropic's models after a data retention policy change, with Nvidia routing sensitive work to its own models and Booz Allen barring the models from proprietary cybersecurity work.
Separately, OpenAI could not rule out that a mathematician's private research helped its own model solve part of the Navier-Stokes problem, as VentureBeat reported.
Neither story involves Cohere, but both turn on the same unresolved question: whether a vendor's promise not to look is enough, or whether a customer needs proof the vendor structurally cannot. That is the specific gap that confidential computing and a hardware attestation report are designed to close.
How Cohere's Model Vault and Confidential Computing system works
Model Vault is Cohere's single-tenant inference platform, a dedicated deployment of Cohere's models rather than a shared multi-tenant API.
Confidential computing. The approach runs that deployment inside hardware that encrypts a workload's memory and isolates it while the workload runs. Nobody operating the surrounding infrastructure, cloud provider included, can see what is inside.
The hardware. According to Cohere, Model Vault's new Encrypted tier pairs a confidential VM on the CPU side with Nvidia GPUs running in confidential computing mode. The CPU side runs on Intel TDX or AMD SEV-SNP. This is the hardware boundary behind the claim in the lede. Protection now spans both the CPU and the GPU, rather than stopping at the CPU the way most earlier confidential computing deployments did.
The data path. The data stays encrypted as it travels from the customer to the Confidential VM (CVM)
"Within the confidential environment, it remains protected throughout the hardware path, including in memory, across CPU and GPU interconnects, and in other hardware components," Govindassamy said.
Govindassamy noted that the data is only decrypted at the point where it must actually be processed inside the protected CPU and GPU execution environment. Even processor state and registers are protected as part of the confidential computing boundary.
Outside the CVM, the data remains encrypted and inaccessible to Cohere, the cloud provider, or other administrators. The encrypted response is returned back to the user, who is the only one able to decrypt the response using their keys.
The boundary. It runs from the customer's own keys on one end to the customer's own keys on the other. Everything in between happens inside hardware Cohere says it cannot read. That includes the inference step itself.
Verify, don't just trust the vendor
A hardware boundary is only as useful as a customer's ability to confirm it holds. Govindassamy addressed how far the "not even us" claim actually extends, and where its limits sit.
"No one, including Cohere, the cloud provider or the cluster operator can see or access any customer data," Govindassamy said. "Every inference returns an attestation report that lets customers verify the exact hardware, software, and security policies protecting their workload."
Going a step further, Cohere plans on making the full serving stack used in Model Vault open source so that independent auditors will be able to validate that it doesn't log, export, or leak data. The hardware and software 'golden values' are registered with Intel's Trust Authority, so when a confidential VM boots, Intel can confirm it matches the approved, tamper-proof configuration from Cohere, Intel and Nvidia.
Customers can also use open-source tools to independently check policies and confirm their inferencing workloads are running inside the verified and secure environment.
What's not entirely clear is whether the verification happens once when the environment boots or on every individual request.
"Confidential inference becomes operationally credible when attestation is part of the request path," developer Zhao Liu Liu wrote in an X message. "Bind the signed measurement to model version, GPU policy, code digest, and retention rules so a customer can verify the enclave they actually used."
The competitive field for confidential inference
Confidential inference is not a new category. Several vendors already sell some version of the promise that the operator cannot read the prompt, and most of them run on the same hardware Cohere is using now.
Tinfoil runs open-weight models like Mistral and DeepSeek inside trusted execution environments, with public attestation backed by a transparency log. Edgeless Systems built Privatemode for European government and regulated-industry buyers who want an EU-based trust story.
Phala rents attested GPU capacity and hosts a catalog of open models rather than shipping its own. Maple AI encrypts client-side and talks only to confidential computing backends.
Confer, built by Signal creator Moxie Marlinspike, offers client-held keys and attested open-source server code, modeled explicitly on Signal.
What separates those products is not the hardware. It is who owns the model and who the buyer is. Tinfoil, Phala, Privatemode and Maple are largely running someone else's open weights inside an enclave. Confer is a private app built on attested open code. Cohere is doing neither.
Encrypted Vault applies confidential computing to Cohere's own enterprise models inside Model Vault, the single-tenant inference product it already sells alongside its North platform.
The buyer it is built for is a bank, a hospital or a public-sector team that wants its own model running in infrastructure Cohere operates, with a vendor-signed attestation token it can hand directly to an auditor. The category already existed. What is new is a frontier-adjacent enterprise lab shipping this as a standard SKU rather than a separate privacy-focused product.
What this means for enterprises
Confidential computing does not resolve every data protection question on its own. What it does is change what an enterprise can ask for. A compliance letter is a promise. An attestation report is something a security team can check against hardware measurements without taking anyone's word for it.
The practical test for any inference vendor, Cohere included, is no longer whether it says it protects customer data. It is whether it can produce a signed attestation report and show which parts of the stack that report actually covers.