ungovernable.tech

[ COMPARISON ][ CONFIDENTIAL COMPUTING ]

Confidential Computing Across Cloud Providers

Also Cloud CC Comparison.

Which cloud does what, with which silicon, and what you actually get when the marketing page says "confidential."

Usable updated 2026-10-25 4 min read
Contents

State: early 2026. This space moves quarterly; verify before committing.

Capability matrix

Capability Azure AWS GCP Oracle OCI IBM Cloud Alibaba
Intel SGX (app enclaves) DCsv3 (GA) — (deprecated in cloud) — — — ECS ebmre6p (GA)
Intel TDX (CVM) DCesv5 (GA) Roadmap c3-standard-* (TDX) — LinuxONE-adjacent only ECS g8i (GA)
AMD SEV-SNP (CVM) DCasv5 / ECasv5 (GA) Selected M6a/C6a/R6a sizes; check tenancy and region requirements N2D Milan (GA June 2024); C3D uses SEV (not SNP) E5 (GA) — ECS g7a (GA)
NVIDIA H100 CC NCC H100 v5 (GA) — (roadmap) A3 Confidential (preview as of late 2025) BM.GPU.H100.CC (GA) — —
Nitro Enclaves (non-DRAM-encrypted VM isolation) — Supported Nitro instance types, with exclusions — — — —
Attestation service Microsoft Azure Attestation (MAA) KMS attestation conditions + Nitro attestation Confidential Space attestation OCI attestation (beta) HPVS attestation Alibaba-specific
Key release service Azure Key Vault Secure Key Release AWS KMS with attestation conditions GCP KMS + Confidential Space OCI Vault HPCS KMS
Confidential Containers / CoCo support AKS CoCo preview EKS via peer-pods (community) GKE Confidential GKE Nodes (SNP) OKE (community) IBM HPVS ACK
Confidential GKE / AKS at scale AKS Confidential EKS (partial) GKE Confidential Nodes GA OKE CC — ACK

Provider-specific notes

Microsoft Azure

  • The broadest SKU catalog. Currently the only public cloud with SGX, TDX, SEV-SNP, and H100 CC at the same time.
  • Azure Attestation (MAA) is mature, multi-platform, and returns a JWT a relying party can verify.
  • Secure Key Release (SKR) in Key Vault with Managed HSM is the strongest KMS-gating story among the majors.
  • Confidential AKS integrates CoCo and Kata.
  • Caveat: MAA is an Azure-operated verifier. For strongest assurance, verify raw quotes against Intel/AMD directly, not via MAA alone.

AWS

  • Nitro Enclaves are AWS's in-house solution: isolated mini-VMs carved out of a parent EC2 instance, with signed attestation documents. Not encrypted DRAM. Isolation is hypervisor- and silicon-level via Nitro, not memory encryption. For many threat models this is enough because the Nitro hypervisor is smaller and more trusted than a general-purpose hypervisor, but it is a different model than SGX/TDX/SEV-SNP.
  • AWS documents SEV-SNP support for selected M6a, C6a and R6a configurations, with separate requirements for Shared, Dedicated Instance and Dedicated Host tenancy.
  • KMS plus Nitro attestation conditions is the cleanest KMS-gated-key pattern in the industry. A single IAM policy clause can enforce "only release key to enclave with PCR0=X."
  • No GA H100 CC yet. Watch Trainium and Inferentia; AWS is building in-house confidential accelerators.

Google Cloud

  • Pioneered public-cloud SEV-based Confidential VMs in 2020.
  • Confidential Space is a serverless-style product that packages a hardened Confidential VM, attestation, and workload identity federation to GCP APIs. A good default for multi-party computation and attested workloads.
  • GKE Confidential Nodes provides SNP nodes at scale; TDX is in preview.
  • A3 Confidential is the main offering for confidential LLM inference with H100 CC.

Oracle OCI

  • Fewer SKUs but aggressive on H100 CC bare metal. Notable because you own the whole node, removing co-tenant side channels.
  • OCI Vault and attestation are less mature than Azure or AWS; you will do more plumbing.

IBM Cloud

  • Hyper Protect Virtual Servers use IBM Z Secure Execution (s390x), a different TEE history from x86 with strong pedigree for regulated finance and health.
  • Not relevant to NVIDIA GPU confidential inference today.

Alibaba / Tencent / Huawei

  • Alibaba ECS has SGX, TDX, and SNP offerings. Attestation collateral flows through Alibaba's own services. Relevant for APAC deployments; audit the trust chain carefully because the operator's verifier is also the cloud operator.

Decentralized and specialist operators

  • Phala, Marlin, Oasis, Secret Network, Fluence: TEE marketplaces where nodes are community-operated, attestation is verified on-chain, and payment is usually crypto.
  • CoreWeave, Lambda, Crusoe, TensorWave: GPU-specialist clouds. H100/H200 CC availability varies, and these often require bring-your-own orchestration.
  • Tinfoil, Edgeless, Anjuna, Fortanix: layer atop the majors with hardened images and attestation UX.

Choosing a provider

You care about… Go with
Fastest path to confidential LLM inference Azure NCC H100 v5 or GCP A3 Confidential
Strongest KMS-gating UX AWS KMS + Nitro Enclaves / SNP
Confidential Kubernetes at scale GCP Confidential GKE or Azure Confidential AKS
Bare-metal, no co-tenancy Oracle OCI or CoreWeave
Regulated finance/healthcare (EU/US) Azure Confidential + Fortanix, or IBM HPVS
Minimum trust in cloud operator Decentralized TEE marketplace (Phala, Marlin) + third-party verifier
Open-weights LLM for end users Tinfoil, Phala, or self-host on Azure/GCP with an open verifier

The trust-is-not-free small print

Every cloud provider controls: - The BIOS and bootloader you cannot audit remotely (though they are measured and attested). - The attestation service endpoints (unless you bring your own verifier). - Key management infrastructure (unless you use split-key or external KMS). - Network routing and egress logging. TEEs do not hide traffic metadata. - Supply-chain access to silicon. A cloud can in principle receive silicon the vendor is not shipping elsewhere.

Confidential computing reduces the cloud's plaintext-access surface dramatically; it does not remove the cloud from your threat model. Pair with transparency logs, multi-vendor quorum, or on-chain verification when you need to actually minimize cloud trust.

Primary sources

Page history

Updated
2026-10-25
Origin
Compiled from the ungovernable.tech research notes.
Improve this page
Suggest an edit — issues and pull requests are welcome.