Glossary
Confidential Computing
Encryption at rest and in transit leave one gap: data has to be decrypted to be processed. Confidential computing closes it with hardware that isolates, and can prove, the code working on your data.
General definition
Confidential Computing is the protection of data in use. The Confidential Computing Consortium, a Linux Foundation project, defines it as performing computation in a hardware-based, attested trusted execution environment (TEE). Conventional controls cover data at rest (disk encryption) and in transit (TLS), but a processor has to work on plaintext, and at that moment the data is visible to anything with enough privilege on the host: the operating system, the hypervisor, a cloud administrator, or an attacker who has compromised one of them. A TEE places the workload in a region of memory that the processor encrypts and isolates so that even privileged software outside the boundary cannot read it.
Two properties make it more than memory encryption:
- Isolation: the processor enforces that code outside the TEE cannot read or modify its memory, whether that outside code is the host OS, the hypervisor or firmware
- Attestation: the hardware signs a measurement of the code and configuration loaded into the TEE; a relying party checks that signature against the chip vendor’s root of trust before releasing secrets such as decryption keys or model weights
- Scope: process-level enclaves (Intel SGX, Arm TrustZone) protect a small piece of code, while VM-level TEEs (AMD SEV-SNP, Intel TDX, Arm CCA) protect a whole virtual machine running unmodified software
All three large clouds sell it: Azure confidential VMs, Google Cloud Confidential VMs, and AWS Nitro Enclaves plus SEV-SNP instances. NVIDIA H100 and later GPUs add a confidential mode so that model inference can run with weights and prompts protected in GPU memory. That is why the term shows up in AI procurement: it lets a hospital or a bank run inference on a provider’s hardware without the provider being able to read the data. See secure enclave for the device-side version of the same idea.
The limits are worth knowing. A TEE does not fix bugs in the code you put inside it, side-channel attacks against TEEs are an active research area, the attestation chain depends on the chip vendor’s keys, and the cloud provider still owns the hardware. It also costs more and carries a performance penalty, larger for GPUs than CPUs. For many regulated teams the simpler route is to decide where plaintext is allowed to exist and keep it there, which is a data residency and self-hosting decision rather than a cryptographic one.
In the Ethora ecosystem
Ethora does not require confidential computing and does not ship a TEE-specific build. Its answer to the same question, who can see your data while it is being processed, is the deployment model: on a dedicated or self-hosted deployment the chat server, database, file storage and AI service run inside your own cloud account or data centre, so the only parties with host access are the ones you already trust. The AI service talks to any OpenAI-compatible endpoint, so the model can be local as well, which is the private AI pattern: no prompt or transcript leaves your boundary.
The two approaches combine. Because Ethora Core installs with Docker Compose or the Helm chart on ordinary Linux hosts, it can be deployed on a confidential VM from Azure, Google Cloud or AWS, and a self-hosted model server can sit on a confidential GPU instance, with no change to the Ethora stack. Whether that is worth the cost depends on your threat model: if the concern is the cloud operator, confidential VMs help; if the concern is a SaaS vendor reading your messages, self-hosting already removes the vendor.