Arcium’s Blackthorn Alpha Combines MPC With NVIDIA GPU Confidential Computing

Arcium opened Blackthorn’s private alpha on Oct. 21, giving organizations ways to test its AI inference stack, which combines multi-party computation (MPC) with NVIDIA GPU Confidential Computing to protect data during inference.

Applications go through a portal that requires an organization email. Access options include an API, installation on compatible GPUs operated by the customer, or dedicated Blackthorn units by request.

Arcium’s design encrypts data on the user’s device and decrypts it inside the GPU’s protected memory. It removes the CPU trusted execution environment from the trust chain while retaining the GPU’s confidential-computing protections. Blackthorn is designed to keep prompts, outputs and models encrypted during use and inaccessible to the host, operator and Arcium.

The listed hardware includes an H100 HGX unit with eight H100 GPUs and 640 GB of GPU memory, and a B200 HGX unit with eight B200 GPUs and 1,440 GB. Arcium identifies AI inference providers, healthcare AI, financial AI, and sovereign or regulated AI as intended users.

Blackthorn’s public materials include no numerical benchmarks for latency, throughput or cost, and the access portal does not show API rates. TokenPost previously covered Blackthorn’s private-alpha launch, which marked Arcium’s move from blockchain privacy work into AI.