Product1 publisher3 min readPublished
CoreWeave launches Remote Key Encryption, keeping customer data keys out of its own cloud
CoreWeave says key custody is what keeps enterprise AI projects sitting in security review, so its new encryption service leaves the key in tooling the customer already runs and takes the provider off the decrypt list.
The Product Desk · Product desk

What happened
- CoreWeave unveiled Remote Key Encryption, a service that encrypts customer data sitting on its infrastructure using keys the company itself never holds.
- Encryption runs client-side inside the customer's compute boundary, and no key is imported into a CoreWeave-side key store, so the company holds only ciphertext.
- The service enters limited availability later this year, with IBM as a launch partner.
- The first release protects data on CoreWeave AI Object Storage with keys held in HashiCorp Vault, Vault Enterprise or any product supporting the Key Management Interoperability Protocol.
- Node access stays behind CoreWeave's Support Access Management controls, and the company said its support engineers cannot reach those nodes without a customer's express permission.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint The storage layer now inherits the customer's key hygiene: a rotation script that breaks or a key manager that goes down is the customer's problem to fix, and CoreWeave has no copy to fall back on.
- decision A security reviewer can now answer the decrypt question in two parts, cryptographic for object storage and administrative for the nodes, and has to decide whether the second part satisfies the same auditor.
- exposure Teams that take the Vault route move the single point of custody into a product IBM owns, and IBM is also the partner sponsoring the launch.
- capability Weights and checkpoints governed by rules written for ordinary regulated data can sit on rented GPU capacity without CoreWeave appearing on the list of parties able to read them.
An auditor asks for a list: every party able to turn a stored checkpoint back into usable weights. CoreWeave says that list is what leaves enterprise AI projects parked in security review. Model weights and training checkpoints fall under the same rules as any other regulated corporate data [2][16]. On most clouds the provider's own name belongs on it [3].
The mechanics are ordinary. Encryption happens inside the customer's own compute boundary, with a key generated and stored in whatever secrets manager, key management system or hardware security module that customer already operates [4]. The algorithms underneath are standard ones [17]. What CoreWeave says is novel enough to file on is the lifecycle handling: rotation, expiration and revocation run on the customer's existing schedule and tooling instead of a parallel set of CoreWeave controls. The company said that approach is patent pending [7].
Andy Manoske is product lead for enterprise and security at CoreWeave. Paul Friedman is head of security foundations. The two wrote in a blog post that the fastest path through the review loop to production is not a new security model; it is CoreWeave's security layer speaking fluently to the one the customer has already built [10].
The same argument runs through the identity work that landed earlier. CoreWeave IAM federates with enterprise identity providers including Microsoft's Entra and Okta, leaving the customer's provider as the source of truth [12]. Automated User Provisioning, introduced last November, syncs users and groups continuously, so a hire, a role change or a termination propagates across CoreWeave Kubernetes Service, AI Object Storage and the console without a ticket [13]. Training clusters were the awkward case. SUNK packages the open-source scheduler Slurm for Kubernetes, and Slurm keeps its own identity built on POSIX users, groups and accounts, historically reconciled with a corporate directory by hand [14]. SUNK User Provisioning now creates the POSIX user and groups, syncs SSH keys and sets up the Slurm account when a federated user appears. CoreWeave said a researcher added to a group in Entra or Okta can log in to a cluster node inside a minute with the right permissions already set [15].
The announcement does not compare GPU prices or name a competing cloud [3]. Custody is the ground CoreWeave picked to argue on, and so far the case for it rests on the argument alone: the material shows no customer moving a workload because of it.
Whether this is a configuration change or a quarter of work depends on where the regulated data actually sits, because the first release covers CoreWeave AI Object Storage and nothing else [9]. It also depends on whether the key store already in production can hold a key that CoreWeave will never import into its own key store [5].
What to watch
- Whether limited availability extends past AI Object Storage to cluster filesystems and node-local disks, and on what timetable.
- Whether general availability names key stores outside IBM's HashiCorp line, or leaves KMIP support as the only alternative path.
- What CoreWeave executives say about custody at Fully Connected in San Francisco on Sept. 30 and Oct. 1.