Arcium has launched Blackthorn, an AI inference stack designed to ensure privacy for prompts, model weights, and outputs. The system replaces hardware-based trusted execution environments with secure multi-party computation. The private alpha phase for the product began on October 21, 2026.
The architecture utilizes two layers: user-controlled key-based encryption and an MPC mechanism managed by Arcium. The platform supports standard NVIDIA Hopper and Blackwell GPUs, including H100, H200, and B200 models. According to the company, this expands the pool of accessible confidential GPUs from 2,000 to approximately 5 million units.
Arcium previously processed over 1 million confidential computations on its network. By June 2026, the daily volume exceeded 200,000 computations. The company is targeting sectors such as healthcare, finance, law, and government for its new AI layer, which complements its existing DeFi infrastructure on Solana.
