Skip to main content
Until now, the industry standard for confidential inference has utilized a Confidential Virtual Machine (CVM) launched on top of CPU TEEs (Intel TDX or AMD-SEV) sitting on the GPU host. However, TEEs have an extensive history of breaches and compromises. This is where trust and security collapse. Blackthorn is Arcium’s replacement for vulnerable CPU TEEs and encrypts the outer channel with AES keys secret-shared across MPC nodes. It lets you run a model on infrastructure you don’t control (a cloud GPU, a rented box, someone else’s cluster) without that operator or Arcium ever seeing your prompts or the model’s responses.

​
What you get with Blackthorn

  • Your prompts and the model’s responses are readable only on your own device and inside the GPU’s protected memory, never to the machine that hosts the GPU.
  • Trust is split, not concentrated: no single party, including the host operator, holds the keys that could open your traffic.
  • It runs on standard, shipping hardware — no custom silicon, no CPU TEE on the host.
Blackthorn runs on shipping NVIDIA Confidential-Computing GPUs. It supports NVIDIA Hopper (H100) and Blackwell (B300) GPUs.
Next: how it works.