Encryption at rest and in transit are solved problems this series configured months ago. The remaining exposure is data in use: while your code computes on it, data sits in memory in cleartext, visible in principle to the hypervisor, the platform operator, and anyone who compromises either. For most workloads that residual risk is accepted without ceremony. For some, regulated data, multi party computation, IP whose threat model includes the infrastructure itself, it is not, and confidential computing closes it with hardware.

The Hardware and the Spectrum

The mechanism is the trusted execution environment: CPU enforced memory encryption and isolation such that even privileged software outside the boundary reads only ciphertext. Azure’s spectrum runs by granularity. Confidential VMs (AMD SEV-SNP on the DCa and ECa families, Intel TDX on their counterparts) wrap the entire VM: memory encrypted with keys the hypervisor never holds, no application changes, lift and shift confidentiality, the option most estates actually adopt. Application enclaves (Intel SGX on the DCsv3 line) shrink the trusted boundary to a fragment of your process, strongest isolation, real development cost, chosen when the threat model justifies engineering against the enclave SDKs or the frameworks that paper over them. And up the stack, confidential AKS node pools run your containers on confidential capable hardware, while confidential containers aim the enclave boundary at the pod itself, the granularity the platform world is converging toward.

Attestation: The Part That Makes It Real

Encrypted memory without proof is a brochure claim. Attestation is the proof: the hardware signs evidence of what is running, the platform measurements, the launch state, and Microsoft Azure Attestation validates that evidence against policy, issuing a token your systems can gate on. The pattern that turns this into architecture is secure key release: keys in Key Vault Managed HSM carry a release policy requiring a valid attestation claim, so the decryption key for the sensitive dataset literally cannot leave the HSM except into a verified confidential environment. That is the design worth internalizing, because it converts confidentiality from a deployment property into an enforced precondition: the pipeline from the DevOps posts deploys the workload, the workload attests, the key releases, and any tampered or non confidential environment simply never obtains the material. SQL’s Always Encrypted with secure enclaves applies the same idea to databases, rich queries (comparisons, ranges) over encrypted columns computed inside the enclave, ending the old tradeoff where client side encryption cost you every WHERE clause.

Who Needs It, Honestly

The adoption test is a named threat or a named clause, not vibes. Clean triggers: regulated data whose compliance interpretation demands infrastructure operator exclusion; multi party analytics where competitors compute jointly on pooled data none may see (confidential clean rooms are the emerging productization); sovereign requirements layering technical guarantees over contractual ones; high value models and keys whose exfiltration threat model includes host compromise. For those, the cost is modest and concrete, confidential VM families price near their standard siblings, attestation is a managed service, and the deployment deltas are a VM size, a disk encryption option, and a key release policy in the Terraform you already have. For everyone else, the honest sequencing is that the controls from this series, private networking, managed identities, PIM, policy, Defender, retire vastly more probable risks first, and confidential computing is the capstone you place when the walls exist. It is the rare security technology whose pitch is not fear but arithmetic: the trusted computing base shrinks from the cloud provider’s software stack to a CPU and your code, and for the workloads that need that number small, nothing else gets it there.

Cheers
Osama