Intel TDX: whole-VM trusted domains and attestation
Intel TDX creates hardware-isolated Trust Domains while leaving scheduling and resource management to an untrusted host.

Direct answer
Intel Trust Domain Extensions create whole-VM TEEs called Trust Domains. TDX protects TD private memory and saved CPU state for confidentiality and integrity against the host VMM and other platform software, while attestation lets a verifier evaluate TD measurements, attributes, and TCB state.How the Trust Domain boundary works
The Intel TDX module runs in a more privileged secure arbitration mode and mediates creation and execution of Trust Domains. Private TD memory and saved CPU state receive hardware-enforced confidentiality and integrity protection from the VMM, host OS, and other VMs. The host still allocates resources, schedules virtual CPUs, and handles emulated I/O.OVERVIEWARCH
| Inside the protected intent | Still controlled by or exposed to the host |
|---|---|
| TD private memory and protected CPU state | Scheduling, resource allocation, interrupts, and availability |
| Measured TD build and selected runtime measurements | Shared memory and deliberately shared buffers |
| Guest code/data admitted inside the TD | Emulated I/O, device input, hypercalls, and other host-supplied values |
| Attestation report data bound by the guest | Network/storage outside the TD unless separately protected |
TDX attestation in practice
MRTD represents build-time TD measurement, while runtime measurement registers can extend the boot and software chain. Guest software requests a TDREPORT and can bind application data such as a session public-key hash. A quoting enclave produces a signed TD Quote for remote verification.ATTEST
- Validate Intel certification and current TCB collateral, including expiry and revocation.
- Compare MRTD and applicable runtime measurements with approved values.
- Check TD attributes, security version and debug-relevant policy.
- Validate freshness and the report-data binding before releasing secrets.
- Operate a controlled update process for new measurements and platform security baselines.
What TDX does not protect
- Availability: a malicious host can deny CPU, memory, devices, network, or power.
- Untrusted I/O: shared pages, hypercalls, MMIO, port I/O, CPUID/MSR values, and device input must be treated as attacker-controlled unless another protected mechanism applies.
- Guest compromise: vulnerabilities in the guest kernel or application run inside the TD boundary.
- Every side channel or sophisticated physical attack: use Intel’s current security guidance for the exact platform.
- External data paths: storage, network, logging, and management services require separate controls.
Customer fit and operational constraints
TDX fits enterprises that want a familiar VM operating model while excluding a cloud or platform administrator from protected guest memory. It is particularly relevant when a supported NVIDIA confidential GPU must attach to an Intel-hosted CVM. Teams must still maintain guest images, attestation collateral, reference values, device/I/O policy, upgrade workflows, and fail-closed key release.
Frequently asked questions
Does Intel TDX protect an entire virtual machine?
It is a whole-VM TEE: the guest runs as a Trust Domain. TDX protects TD private memory and CPU state from the host, while the guest kernel and applications remain inside the trusted computing base.
Can the hypervisor still control a TDX VM?
It still schedules and allocates resources and handles untrusted interfaces, so it can deny service or supply malicious I/O. It should not be able to read or silently alter protected TD private memory within the documented threat model.
What should a TDX verifier check?
It should check Intel certification and TCB collateral, freshness, expected measurements, security versions and attributes, and any application binding carried in report data.
Sources
- OVERVIEW
- ARCH
- GUIDANCE
- ATTESTIntel TDX attestation enabling guideIntel Trusted Services
Relevant GPU availability
Verified specifications, confidential-mode support, and public listings for the accelerators this post covers.
Ready to reserve capacity?
Confidential Nodes matches bare-metal confidential GPU nodes to workloads, with verified provider data behind every listing.
