What is confidential computing?

Confidential computing uses hardware-backed isolation to protect data while software processes it. Data is decrypted only inside a trusted execution environment, or TEE, separated from the host operating system and other workloads. Encryption at rest protects storage and encryption in transit protects networks; confidential computing addresses the remaining data-in-use stage.

A TEE creates an isolated memory area and measures the code and configuration loaded into it. Remote attestation lets another party verify those measurements before releasing data or keys. The protection depends on processor, firmware, virtualization and attestation services. It can isolate a virtual machine, an application or a smaller function, each with different operational tradeoffs.

Banks can use TEEs for fraud models, credit scoring, payment authorization, market-risk calculations and key operations. Confidential computing can protect shared analysis without handing clear data to another party or the infrastructure administrator. Azure and other cloud platforms also support protected AI models on a GPU. The method does not decide whether the processing is lawful or whether the model is accurate.

Attestation can prove which code started, but it cannot prove that the business logic is correct. Side channels, firmware flaws, rollback and weak key release policies remain risks. Operations teams also need observability without exposing sensitive memory. A bank should define the trust boundary, patch process, evidence retention and recovery path before it treats a confidential workload as a security control.

NIST defines confidential computing as hardware-enabled isolation and processing of encrypted data in memory. GDPR and banking secrecy still govern the data, while DORA governs ICT risk and providers. Confidential computing can support those controls but creates no automatic compliance. Procurement should require documented attestation, incident handling and evidence that the selected workload and processor combination supports the promised isolation.

BaFin, the FMA and FINMA assess the whole control environment, not the technology label. An EU bank that uses a TEE in a public cloud still needs provider governance and an exit path. Frankfurt banks and market operators process payment, trading and risk data that fit the use case. The ECB's supervision makes explainable evidence about keys, administrators and recovery more useful than a general encryption claim.