NEAR is expanding NEAR AI Cloud as confidential infrastructure for running artificial intelligence models while limiting access to sensitive prompts, outputs and model data. The platform executes supported AI inference inside hardware-isolated Trusted Execution Environments and provides cryptographic evidence of the environment used for each request, targeting developers and enterprises handling sensitive workloads.
According to NEAR AI’s official documentation, inference runs using Intel TDX and NVIDIA Confidential Computing hardware. Prompts are decrypted only inside the protected environment, while inference memory is isolated from the host operating system, infrastructure operators and NEAR AI itself, according to the platform’s security model.
Hardware Attestation Makes Inference Verifiable
The system’s key distinction is that privacy claims can be checked through hardware attestation rather than relying solely on an operator’s policy. Each supported confidential inference request can carry a cryptographic attestation binding execution to the expected hardware, code and model, allowing developers to independently verify the environment that processed the workload.
NEAR AI Cloud uses Intel TDX to isolate the confidential virtual machine and NVIDIA confidential-computing capabilities to protect GPU execution. The combination is designed to keep data encrypted not only while stored or transmitted, but also while it is being processed, addressing a major exposure point for conventional AI cloud services.
NEAR AI added another verification layer in August through Intel Trust Authority, providing independent attestation verification outside the cloud operator itself. That architecture reduces the amount of trust developers must place in NEAR AI when confirming that confidential workloads ran under the expected security configuration.
NEAR Targets Sensitive AI Workloads
The service exposes an OpenAI-compatible API, allowing developers to adapt existing applications without rebuilding their entire inference stack. NEAR AI currently supports a growing catalog of open-weight models and also offers private-model deployments for enterprise workloads, positioning the cloud as infrastructure rather than a single-model AI application.
The privacy guarantees nevertheless depend on the specific execution path. Models actually hosted inside NEAR AI’s confidential GPU environments receive hardware-backed TEE protection and attestation, while externally routed models can operate under different trust assumptions. Developers therefore need to distinguish confidentially hosted models from other model-access options rather than assuming every request receives identical hardware protections.
For NEAR, the broader objective is to make confidential computation a practical component of production AI infrastructure. The key test will be whether verifiable private inference attracts workloads that organizations have previously been unwilling to send to conventional cloud AI services, turning hardware-enforced confidentiality from a security feature into a recurring source of real compute demand.
