Arcium has launched Blackthorn in private alpha, an AI inference stack that encrypts prompts, model weights, and responses end to end. It replaces processor-based trusted execution environments with cryptography distributed across multiple parties, combining user-controlled key encryption with an outer secure multi-party computation layer administered by Arcium. The system runs on standard NVIDIA Hopper GPUs, including the H100 and H200, and Blackwell GPUs such as the B200 and its variants. According to Arcium, this approach expands the pool of GPUs available for confidential computing from around 2,000 specialized units to approximately 5 million existing NVIDIA GPUs, a 2,500-fold increase. Blackthorn targets sectors including healthcare, finance, legal services, and government, while its production-scale performance, latency, and costs remain unproven.
Source: Read the original article
