Corvex Announced the Launch of Secure Model Weights

Why It Matters

It eliminates the cleartext gap that exposes model weights, allowing regulated and security‑focused organizations to protect IP while leveraging scalable cloud compute.

Key Takeaways

• End‑to‑end encryption keeps weights inside GPU memory only
• Uses NVIDIA Confidential Computing and Intel TDX for hardware isolation
• Remote attestation blocks key release to compromised hosts
• Post‑quantum ML‑KEM secures key exchange against future attacks
• Open‑source stack ensures auditability and vendor‑neutral security

Pulse Analysis

The rise of foundation models has turned model weights into high‑value assets, often representing years of research and millions of dollars in compute. Traditional cloud security focuses on data at rest and in transit, leaving a vulnerable window during inference when weights are decrypted in cleartext on the host. This exposure creates a strategic risk for sectors handling sensitive data—healthcare, finance, defense—where a single breach can compromise trade secrets and regulatory compliance.

Corvex Secure Model Weights tackles the problem at the silicon level. By running workloads on NVIDIA Hopper and Blackwell GPUs in Confidential Computing mode, the solution encrypts GPU memory so that only the GPU’s trusted execution environment can access plaintext weights. Intel Trust Domain Extensions add CPU‑side isolation, while remote attestation guarantees that decryption keys are released only to verified hardware configurations. The inclusion of ML‑KEM, a post‑quantum key encapsulation mechanism, future‑proofs the key exchange against quantum‑grade attacks, ensuring that the cryptographic handoff remains secure even as threats evolve.

Beyond technical safeguards, the open‑source foundation built on the Cloud Native Computing Foundation’s Confidential Containers project offers transparency and auditability that many enterprises demand. This vendor‑neutral approach lets organizations choose cloud providers based on performance and cost rather than trust assumptions. For regulated industries, the ability to run cutting‑edge AI on third‑party infrastructure without sacrificing IP sovereignty could accelerate adoption, reduce on‑premise spend, and reshape competitive dynamics in the AI services market.