Movano’s Corvex Merger Yields GPU-Enforced Secure Model Weights Launch
Corvex Secure Model Weights, now in early availability, uses NVIDIA Confidential Computing and Intel TDX to keep AI model weights cryptographically isolated inside GPU silicon, preventing exfiltration on third-party GPU infrastructure. The patent-pending solution, launching March 12 2026, enhances IP safeguards for model builders, bolstering Movano’s post-merger AI infrastructure offering.
1. Product Overview
On March 12 2026 Corvex introduced Secure Model Weights in early availability, enabling AI model builders and enterprises to deploy inference workloads on third-party GPU infrastructure without exposing valuable model weights. The patent-pending solution ensures model weights remain encrypted until decrypted within the GPU’s trusted execution environment.
2. Architecture and Security Features
Secure Model Weights integrates NVIDIA Confidential Computing on Hopper and Blackwell GPUs, Intel Trust Domain Extensions for CPU isolation, remote attestation of the host environment, and post-quantum ML-KEM (Kyber-768) key exchange. This multi-layer hardware-rooted approach ensures keys and weights exist in cleartext only inside protected GPU memory, inaccessible to hypervisors or host kernels.
3. Implications for Movano
Following its all-stock merger with Corvex, Movano gains a hardware-enforced IP protection capability that addresses runtime vulnerabilities in third-party GPU deployments. This launch positions the combined company to target regulated sectors—healthcare, finance, defense—where sovereign key custody and auditable security are critical for fine-tuned AI workloads.