NEAR AI Cloud provides an OpenAI-compatible inference interface designed to let developers migrate existing AI applications while choosing between different execution and privacy models. Developers can point compatible OpenAI client libraries to NEAR AI’s API endpoint without rebuilding their existing request logic, while model selection determines whether inference runs inside confidential hardware or through an external provider.
According to NEAR AI’s official Cloud and Private Chat announcement, the API supports existing client libraries and code paths while its SDKs handle encryption, key management and attestation for confidential workloads. NEAR AI initially introduced the service in December 2025 and has since expanded the available model catalog and verification infrastructure.
Privacy Depends on the Model Route
For models hosted confidentially, NEAR AI encrypts prompts before they enter a Trusted Execution Environment. Inference occurs inside an isolated Intel TDX-based environment that prevents the host operator, infrastructure provider and NEAR AI itself from reading the workload during normal execution. The model processes the decrypted input inside the enclave and re-encrypts the response before it leaves the protected environment.
Those requests also produce cryptographic evidence that developers can inspect. TEE-hosted inference can return hardware-backed attestation proving that the workload ran inside an expected confidential environment, and NEAR AI added Intel Trust Authority in August 2026 as an independent verifier of Intel TDX evidence. That separates verification from the company operating the service.
The privacy model is different for proprietary models supplied by external providers. NEAR AI routes those requests through a shared API credential so the upstream provider does not directly receive the end user’s account identity, but the prompt itself is still sent to that provider for processing. NEAR AI’s current terms explicitly state that these closed-source models do not receive TEE-based privacy or cryptographic attestation.
OpenAI Compatibility Lowers Migration Friction
The common API surface allows applications to switch between model types without maintaining completely separate integration stacks. An existing OpenAI SDK deployment can change its base URL and credentials while preserving familiar chat-completion request patterns, reducing engineering work for teams experimenting with alternative inference infrastructure. NEAR AI currently exposes both confidential open-weight models and proxied proprietary models through this interface.
NEAR AI also distinguishes operational metadata from protected content. Its confidential-inference privacy documentation says end-user prompts and responses are processed inside enclaves with zero content retention after each request, while separate metadata such as latency, token counts, model identifiers and billing information can be retained for operations and accounting.
That architecture makes privacy a property of the selected execution route rather than a universal guarantee attached to every model in the catalog. Developers choosing a TEE-hosted model receive hardware isolation and verifiable execution, while an Incognito third-party model primarily provides identity abstraction at the provider layer. The next meaningful test will be whether developers adopt those differentiated privacy levels at scale while preserving the compatibility advantages of a single API surface.