Provider Directory
Compare providers running AI inference inside trusted execution environments. Filter by models, pricing, and API features.
Verifiable private inference with Intel TDX and NVIDIA H100 CC
Storefront · Router · Workload operator · TEE platformPay-per-prompt confidential inference — TEE-backed models with ECDSA per-request attestation
Storefront · AggregatorConfidential inference on Intel TDX with per-workload attestation receipts
Storefront · Workload operator · TEE platformOpenAI-compatible end-to-end encrypted confidential inference on dedicated GPUs
Storefront · Workload operator · TEE platformOpenAI-compatible confidential inference across attested GPU TEE routes
Storefront · Aggregator · RouterDecentralized confidential inference on Bittensor with TEE
Storefront · Aggregator · Workload operator · TEE platformPrivate, verifiable AI — Intel TDX + NVIDIA H200 TEE with on-chain attestation
Storefront · AggregatorEU-hosted confidential inference via Cosmian VM
Storefront · Workload operator · TEE platformPrivate TEE inference via hardware-attested confidential compute
Storefront · AggregatorEnd-to-end encrypted private AI with TEE attestation
Storefront · AggregatorProvider selection
Compare the protection boundary, the evidence available to your client, and the model you actually need. The examples below are provider-described architectures, not current verified verdicts from this directory.
Tinfoil describes a confidential VM that runs the inference workload and terminates TLS inside the enclave. Its client verifier compares the TLS public key with the attested key before exchanging application data. [1]
For provider selection, ask where plaintext is available and which components sit outside that boundary. Inspect the Tinfoil provider record, then compare its models and prices.
Privatemode documents collecting hardware-signed evidence, checking it against expected reference values, and releasing trust only on a match. It distinguishes attestation from the later steps that establish the encrypted request path. [2]
Use the attestation monitor to inspect the directory's dated checks and incomplete evidence. A provider's architecture description is not proof that a particular inference request followed a verified route.
For GPU workloads, Tinfoil documents GPU verification during the confidential VM's boot process. Privatemode separately describes worker admission and GPU verification before GPU access is enabled. [1] [2]
Compare CPU and GPU coverage separately; do not infer accelerator protection from a CPU enclave label alone. The Privatemode provider record separates documented capabilities from the directory's observed evidence.
The documented request paths differ: Tinfoil describes attested TLS key binding and encrypted HTTP bodies, while Privatemode describes client-side prompt encryption and key release to attested workers. [1] [2]
Shortlist providers by available models, billing units, client integration requirements, and the evidence for the chosen request path—not by the word “private” alone. Start with the provider comparison above and compare the same model across providers.
Disclosure log
These providers make relevant confidential-inference claims, but do not currently expose enough working, public evidence to join the model and pricing comparison above.