RESEARCHED PROVIDER · PRIMARY SOURCES

NEAR AI

NEAR AI's own confidential GPU fleet, assessed separately from its third-party and Incognito API routes.

INPUT / OUTPUT PRIVACY

Stage 1

Partial proof

ASSESSED DEPLOYMENT

NEAR-hosted Confidential TEE tier

RESEARCH REVIEW

2026-10-09 · primary-source review

Provider website

Verification labels describe capabilities established by the cited sources. This review did not perform live cryptographic attestation, reproduce production builds or complete an independent security audit.

Who actually runs the models?

The Confidential TEE tier runs models on NEAR AI's own GPU fleet. Its 3P Confidential TEE and Incognito tiers route to other providers; both are excluded from this entry.

Deployment and TEE coverage

Only NEAR's own model deployments are assessed. Use canonical model IDs and verify the NEAR model evidence as well as the gateway; gateway-only verification is insufficient to identify the inference provider.

NEAR documents Intel TDX and NVIDIA confidential GPUs for its own private inference nodes. Incognito models execute outside this boundary. Third-party TEE models have different verification chains and SDK coverage.

At least one content layer has independently checkable technical evidence, including partial proof within a layer. The full privacy boundary is incomplete. The stage tooltip distinguishes partial findings from fully met requirements for Stage 2.

STEPS FOR STAGE 2

2 of 3 requirements for Stage 2 met. Missing proof: Logs & storage.

  1. Request pathMet

    Private inputs and outputs across the request path, with E2EE or an equivalent attested channel that protects against intermediaries and binds keys to the accepted workload.

    Technically verifiable. Same-connection TLS binding

  2. Inference executionMet

    Protected inference with a verifiable workload identity and protection covering the CPU, GPU and every place content is processed.

    Technically verifiable. CPU + GPU; image provenance

  3. Logs & storagePartial

    No content logging, persistent storage, human review or training. Any temporary content-derived cache must have verifiable isolation and enforced removal.

    Partial technical proof. KV expiry not established

Identity and operational metadata are assessed separately below. They do not set the content stage; prompt or response content in telemetry remains part of the content-retention assessment.

Request path

Technically verifiable

Verified gateway. The gateway flow verifies fresh gateway evidence and binds the actual TLS connection before sending content. It also verifies every eligible NEAR model candidate. The current inference SDK supports verified encrypted Chat Completions; an ordinary OpenAI/HTTPS client does not perform those checks. Required source/build policy must be applied to both gateway and model images. Experimental direct endpoints and third-party routes remain excluded.

Inference execution

Technically verifiable

NEAR model evidence. Use canonical IDs, provider=near and x-no-aliasing, and verify every returned NEAR model candidate including applicable GPU evidence. The provenance guide hashes the exact attested Compose bytes, extracts immutable image digests and requires cryptographic build/source verification for each required image. The inference proxy publishes reproducible builds. Missing provenance or an unaccepted revision must fail the caller policy; a response signature alone is insufficient.

Logs & storage

Partial technical proof

Inspectable handling; cache gap. The public inference proxy separates diagnostic fields from request/response bodies, and its signature cache has a configurable TTL. That cache contains signature records, not the model engine's prompt KV blocks. Current documentation confirms automatic KV prefix reuse inside each model TEE and eviction as capacity is needed; no enforced removal deadline or complete per-user cache-isolation policy was established. Stateless Responses omit conversation storage but still use these caches. This is partial handling evidence, not a full no-retention proof.

Identity & metadata

Documented only

Account-linked usage. The current usage API exposes organization/API-key records, model-level input/output counts, cached-token usage, costs and request identifiers. These are billing and operational metadata rather than prompt or response content. Their collection does not set the content stage; the reviewed API and architecture do not establish a complete fixed retention or deletion schedule.

Remaining trust assumptions

  • Intel and NVIDIA hardware isolation and attestation roots.
  • The verified gateway, private model-node code and client verification policy.
  • Selection of NEAR's own Confidential TEE tier rather than a routed upstream model.

Limits of this assessment

  • The NEAR AI brand and a TEE gateway do not imply every catalog model is NEAR-hosted or confidential.
  • Gateway and model verification are distinct; experimental direct endpoints have limitations.
  • No live cryptographic verification or complete retention audit was performed.

Primary-source record

11 sources · reviewed 2026-10-09

technical docsNEAR AI model hosting and privacy tiersExplicitly separates NEAR's own GPU fleet, third-party TEE models and external Incognito routes.technical docsNEAR AI private inference architectureDescribes GPU nodes, the gateway, TLS verification, billing and direct-endpoint limitations.technical docsVerify the NEAR AI Cloud gatewayDefines separate gateway, model, TLS, response-signature and provenance checks.technical docsNEAR gateway and model image provenance policyBinds exact Compose bytes and immutable image digests to cryptographically verified source/build provenance policies.technical docsNEAR verified and encrypted inference SDKDocuments current client verification, E2EE and deployment policies separately from ordinary OpenAI clients.source codeNEAR reproducible inference proxy and logging designDocuments reproducible image/source comparison, content-free diagnostic intent and a signature-cache TTL distinct from KV state.source codeNEAR inference proxy error-handling implementationProvides inspectable diagnostic/error behavior; arbitrary upstream error messages still require review.source codeNEAR signature-cache implementationCaches signed response records with expiration; this is not a prompt KV-cache deletion mechanism.technical docsNEAR prompt-cache privacy and evictionConfirms hardware-protected per-deployment KV reuse and capacity-driven eviction, without a removal deadline.technical docsNEAR stateless Responses lifecycleDistinguishes caller-managed history and absent response storage from still-enabled model prompt caches.technical docsNEAR usage reporting and metadataIdentifies retained organization/API-key, model, token and cost information separately from inference content.
Back to providers

Sources reviewed 2026-10-09