Rafay and Fortanix bring confidential AI to Kubernetes pods
Rafay and Fortanix have extended their confidential VM work to Kubernetes, running each pod in an encrypted environment where keys are released only after CPU and GPU attestation succeeds.

Why it matters
- Rafay and Fortanix launched a joint confidential AI service for Kubernetes workloads on 19 November 2026.
- Fortanix releases encryption keys only after cryptographic attestation of CPU and GPU trusted execution environments succeeds.
- 93% of AI leaders know traditional encryption fails during processing, but only half can verify protection during inference (Fortanix/Atomik Research).
- Cloud operators can package confidential AI as a self-service SKU via Rafay's Developer Hub.
- The service is available now alongside the companies' existing confidential virtual machine integration.
Rafay and Fortanix have launched a joint confidential AI service that runs Kubernetes workloads in encrypted environments, releasing data only after cryptographic proof that the underlying hardware and software have not been tampered with. The offering, announced 19 November 2026, extends the two companies' existing work on confidential virtual machines to containerized AI workloads orchestrated by Rafay's platform on CNCF Confidential Containers with Fortanix Confidential AI.
The stakes are straightforward. AI workloads in regulated industries — financial services, healthcare, the public sector — process customer records, patient data and financial transactions on infrastructure the data owner does not control. Model providers face the mirror problem: proprietary weights deployed on third-party GPU fleets can, in principle, be read by the operator. The service attempts to close both gaps at once.
What actually ships?
The division of labor between the two companies is explicit. Rafay provides multi-tenancy, self-service access and governance controls. Fortanix verifies the runtime environment through cryptographic attestation (a process where hardware cryptographically proves its own integrity before being trusted) and releases encryption keys only if that verification succeeds. If the checks fail, the workload never receives the keys.
The mechanism sits at the Kubernetes pod level. Confidential Containers, a Cloud Native Computing Foundation incubating project built on Kata Containers (a lightweight virtualization runtime for containers), runs each pod inside its own confidential virtual machine. The CPU encrypts the pod's memory, and on supported hardware, GPUs fall inside the trust boundary as well.
Two Fortanix components do the security work. Confidential Computing Manager attests the environment by checking the integrity of CPU and GPU trusted execution environments (isolated, hardware-enforced regions of a processor), firmware and software components. Data Security Manager then releases encryption keys only to workloads that passed those checks.
Rafay's contribution is turning that infrastructure layer into something a cloud operator can sell. Its platform automates cluster deployment, onboards GPU infrastructure once for multiple tenants, packages confidential workloads as self-service stock keeping units, and applies access controls and quotas across tenants and clusters. End users launch workloads through Rafay's Developer Hub rather than waiting for an administrator to provision them manually.
The takeaway: the security guarantees and the commercial packaging come from different vendors, and the service works only if both halves hold.
Who is the research behind the pitch?
The companies lean on survey data to frame demand. Research conducted with Atomik Research and cited by Fortanix found that 93% of AI leaders understand traditional encryption does not protect data while AI systems are processing it, but only half can fully verify that such protection is in place during inference (the stage where a trained model generates outputs).
That gap between awareness and verifiable protection is the market the service targets. Traditional encryption covers data at rest and in transit, but data must be decrypted in memory to be processed — the window confidential computing aims to close, particularly when a third party such as a neocloud provider, sovereign cloud operator or internal shared AI platform manages the hardware.
Half of AI leaders cannot verify protection during inference; whether this product closes that gap depends on attestation actually catching the attacks it is designed for.
What do the companies say?
Mohan Atreya, Chief Product Officer at Rafay, framed the commercial logic for both sides of the market: "Confidential computing enables model owners to deploy proprietary AI models on third-party infrastructure with reduced risk of exposing valuable intellectual property, while creating new revenue opportunities for cloud operators to offer premium, security-focused AI inference services."
Atreya also described the operational split: "Fortanix verifies the environment through cryptographic attestation and controls the secure key release. Rafay turns that into a governed, multi-tenant service, so a neocloud or sovereign cloud can offer confidential AI as a SKU, and its customers can run sensitive workloads on infrastructure they don't operate. With Confidential Containers, that model now reaches the Kubernetes pod, where AI teams already ship their workloads."
Anuj Jaiswal, Chief Product Officer at Fortanix, cast the problem as a three-way standoff: "Every AI Factory faces a trust dilemma. Model owners want their weights hidden from operators and customers; enterprises want their data hidden from model owners and operators; and operators must prove they can see neither, even via another tenant or compromised software."
His claim of resolution is expansive: "Fortanix and Rafay resolve that dilemma in one click. Rafay orchestrates the deployment of proprietary models and AI applications across on-premises, sovereign and cloud infrastructures. Fortanix Confidential AI releases keys only to attested workloads, so model owners, enterprises and operators each keep control of what they own, every time it runs."
"One click" is a marketing formulation, not a technical one. The underlying claim — that attestation-gated key release removes the need for any party to trust the others blindly — is the testable part.
Who is it for?
The companies name four target categories:
- Cloud providers, which can package the offer as a premium confidential AI service, backed by attestation evidence showing customer workloads ran inside trusted execution environments and that keys were issued only to verified workloads.
- Sovereign AI cloud operators, which can keep keys in a hardware security module (a tamper-resistant device for key storage) under their own control while attesting workloads before data is decrypted.
- Enterprises, which gain the ability to run AI workloads on sensitive datasets while limiting exposure to the infrastructure operator during processing.
- Model providers, which can deploy proprietary models into customer environments without exposing model weights before the environment has been verified.
The common thread is that none of these parties has to take the infrastructure operator's word for anything — in principle. The announcement does not specify which GPU platforms fall inside the trust boundary beyond "supported hardware," nor does it quantify performance overhead, which confidential computing deployments typically carry.
Support for Confidential Containers-based deployments with Fortanix Confidential AI is available now, alongside the companies' existing confidential virtual machine integration. Whether neoclouds and sovereign operators adopt it as a billable SKU, rather than as another checkbox feature, will determine whether the three-way trust dilemma the companies describe was a real buying constraint or a slide-deck abstraction.
By James Calloway.
Source: itbrief.com.au
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- NVIDIA Donates OpenShell to CNCF, Ships Open-Source AI Factory Tools
- SUSE pushes AI controls into Rancher Prime and SLES 16.1
- AMD 5th Gen EPYC processors host Google Cloud's confidential TPU 8tx
- Commvault and NVIDIA to Make Protected Enterprise Data Usable for AI
- OpenAI Clears FedRAMP Moderate for ChatGPT Enterprise and Its API