Via iq.wiki
The NEAR Protocol initiative leverages Intel TDX and NVIDIA confidential computing to create hardware-isolated environments where AI models run without anyone, including the infrastructure provider, seeing user data.
Here’s a question that should keep every enterprise CTO up at night: when you send sensitive data to an AI model for inference, who else can see it? NEAR AI is building infrastructure designed to make the answer definitively “no one.”
The project, tied to the NEAR Protocol ecosystem, is constructing a confidential AI platform that runs both open-weight and closed AI models inside Trusted Execution Environments, or TEEs. Your prompts and the model’s outputs stay encrypted and isolated in hardware-protected enclaves, invisible even to the people running the servers.
How the tech actually works #
NEAR AI Cloud offers an API for private inference that integrates two heavyweight confidential computing technologies: Intel TDX and NVIDIA Confidential Computing. These are hardware-level isolation mechanisms that create encrypted memory spaces where code and data are shielded from the host operating system, the cloud provider, and NEAR AI itself. Cryptographic attestation lets users verify that the correct model ran in a genuine secure environment, without any tampering.
The platform supports both open-weight models and “proxied closed models” through its IronClaw agent framework. Open-weight models like Llama or Mistral can be loaded directly into a TEE. Closed models from providers who don’t release weights can be accessed through a proxied setup that still maintains the privacy guarantees.
IronClaw 1.0, released on July 27-28, 2026, is the open-source AI agent runtime powering this stack. It runs autonomously inside enclaves with zero operator access and has been leading benchmarks including PinchBench and ClawBench.
Strategic partnerships signal enterprise ambitions #
On March 19, 2026, Venice.ai integrated with the platform to enable fully private inference for both text and image processing. The integration means prompts and outputs are shielded from access by infrastructure providers or NEAR AI during the entire processing pipeline.
NEAR AI joined NVIDIA’s Inception program on January 13, 2026, aimed at optimizing its confidential computing workflows for enterprise-grade applications.
Other notable partnerships include work with Brave Nightly for confidential inference integration and a collaboration with the Bermuda government for sovereign workloads.
Why enterprises should care #
By running models in TEEs with cryptographic attestation, the platform creates an auditable chain of trust. Organizations can verify not just that their data was encrypted in transit, but that it remained encrypted during processing and that the model executing the inference was the one they expected.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our