Confidential VM (CVM)

The CPU memory controller automatically encrypts all data written to RAM and decrypts it on read using a VM-specific key. This key is generated by the hardware root of trust and is inaccessible to the hypervisor, host OS, or cloud administrator. Technologies like AMD SEV-SNP and Intel TDX implement this at the silicon level, ensuring that even physical memory access or cold boot attacks yield only ciphertext. The encryption is transparent to the guest OS and applications, requiring no code modifications.

Before a Confidential VM processes sensitive data, its identity and integrity must be verified through remote attestation. The TEE generates a cryptographically signed report containing:

• Enclave measurement: A hash of the VM's initial memory state, firmware, and configuration
• Hardware provenance: Proof that the CPU is a genuine, trusted device
• Platform state: Confirmation that security features like Secure Boot are active

A relying party validates this report against known-good measurements before releasing secrets or allowing the VM to join a trusted cluster.

Traditional VMs trust the hypervisor with access to all guest memory. A Confidential VM removes the hypervisor from the Trusted Computing Base (TCB). The hardware enforces that even a privileged hypervisor cannot inspect or modify the VM's memory or CPU state. This protects against:

• Malicious insiders with hypervisor access
• Compromised cloud management planes
• Noisy neighbor attacks from co-resident VMs

The hypervisor retains only its necessary scheduling and resource allocation role, with no visibility into workload data.

Confidential VMs support live migration between physical hosts without exposing data in transit. The process involves:

• Page-by-page encryption: Memory pages are encrypted with a transport key before transmission
• Integrity protection: Cryptographic hashes prevent tampering during transfer
• Destination attestation: The target platform must prove its TEE capability before migration begins

This enables infrastructure maintenance and load balancing while maintaining the confidentiality and integrity of the running workload throughout the relocation.

Secrets such as API keys, database credentials, and decryption keys are only released to a Confidential VM after successful attestation. A Confidential Key Management Service (KMS) integrates with the TEE to:

• Validate the VM's attestation report against a defined policy
• Release keys only to verified enclave identities
• Rotate keys without exposing them to the host

This ensures that sensitive material is never available in plaintext to the underlying infrastructure, even during boot or runtime.

Data written to disk by a Confidential VM can be sealed to the VM's identity using enclave sealing. The TEE encrypts data with a key derived from the enclave's unique measurement and a hardware-bound secret. This means:

• Stored data can only be decrypted by the same application on the same platform
• A compromised host or stolen disk yields only ciphertext
• Data integrity is cryptographically guaranteed

This protects data at rest with the same hardware root of trust that protects data in use.

Enterprise deployments increasingly require heterogeneous TEE environments, where workloads run across AMD SEV-SNP, Intel TDX, and ARM CCA instances simultaneously. This demands an abstraction layer that handles attestation, key management, and scheduling across diverse hardware trust boundaries.

• Challenge: Each TEE has unique attestation protocols and measurement formats
• Solution: Enclave-aware schedulers and unified attestation services
• Key Management: Enclave-aware KMS that releases keys only after successful platform-specific attestation
• Tooling: Open Enclave SDK and SPIRE provide cross-TEE abstraction
• Goal: Single control plane for confidential workloads regardless of underlying silicon

The frontier of CVM technology extends to GPU-accelerated confidential computing. NVIDIA's Hopper architecture introduces confidential computing capabilities that encrypt data in use on the GPU, protecting model weights and inference data during AI computation.

• Technology: NVIDIA Confidential Computing with H100 GPUs
• Key Feature: Isolated execution environment for GPU workloads with encrypted memory
• Attestation: Secure GPU attestation verifies firmware integrity and device authenticity
• Use Case: Protecting proprietary foundation models during inference and fine-tuning
• Integration: Combines with CPU-level TEEs for end-to-end confidential AI pipelines