Glad to see this comparison thread kicking off—I've been wanting to share our migration story for a while. We started with AWS Nitro Enclaves for a compliance-heavy agent runtime, but the round-trip latency in our feedback loops became a real bottleneck. For context, our agents need to attest, fetch a model chunk, run inference, and respond within a few hundred milliseconds. Nitro’s vsock-based setup worked fine for batch jobs, but for tight agent loops we were hitting 30-50ms overhead just for the PCIe bridge and enclave lifecycle transitions.

We recently switched to Intel TDX, and the difference is noticeable. Specifically, we’re running our agent runtime in a TD guest with the runtime measured via TD quote at boot, then keeping the agent alive for multiple inference cycles. The kernel-level memory encryption keeps us compliant (we’re in fintech, so TEE boundaries matter for audit), but the big win is the direct memory access—no cross-VM serialization bottleneck. Our median loop time dropped from ~110ms to ~45ms in the same EC2-like setup.

That said, TDX isn’t a silver bullet. We’re still wrestling with the attestation flow: Intel’s PCCS infrastructure can be a pain to maintain, and the quote verification pipeline (vs. Nitro’s KMS-integrated attestation) required more custom tooling. Also, if you need to load arbitrary code into the guest post-boot, TDX’s measured boot model needs careful planning—we’re pinning a pre-validated agent image in a read-only kernel module to avoid integrity breaks.

For anyone evaluating this space: test your actual loop latency before committing. Nitro’s “cold start” penalty for each enclave creation can kill real-time agents, while TDX’s persistent guest gives you a warm start that’s hard to beat. AMD SEV-SNP sits somewhere in between, but we haven’t tested it at scale yet—I’d love to hear from folks who have.

—sarah (mod)