The encrypted inference system swaps CPU enclaves for multi-party computation and runs on standard NVIDIA Hopper and Blackwell GPUs
Arcium wants you to be able to ask an AI model anything without the model’s operator ever seeing the question. Its new product, Blackthorn, is built around that idea.
The company has launched Blackthorn, an AI inference stack that keeps prompts and responses private. It uses secure multi-party computation (MPC) in place of CPU-based trusted execution environments (TEEs). The private alpha opened on October 21, 2026.
Inference is the part of AI most people interact with. You send a prompt, a model processes it, and an answer comes back. Normally, whoever runs the hardware can, in principle, see all of that in plaintext.
Blackthorn encrypts the full pipeline end to end. That covers the prompts, the model weights, and the outputs. Under its confidential computing model, no single party can access the plaintext data at any point.
A TEE is a sealed, protected zone inside a processor where sensitive code is supposed to run untouched. Arcium’s pitch is that Blackthorn removes reliance on vulnerable hardware TEEs, swapping trust in a chip for cryptography spread across parties.
The architecture uses two layers. The inner layer is key-based encryption that users control. The outer layer is an MPC mechanism administered by Arcium.
Blackthorn supports encrypted inference on standard NVIDIA GPUs, specifically the Hopper line (H100 and H200) and the Blackwell line (B200 and its variants).
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
According to Arcium’s figures, the pool of accessible confidential GPUs grows from around 2,000 specialized units to approximately 5 million existing NVIDIA GPUs. That is a 2,500x increase.
Blackthorn is not Arcium’s first product. Before this launch, the company’s network had already processed more than 1 million confidential computations. As of June 2026, it was handling over 200,000 confidential computations a day.
Arcium has built its name in crypto as a confidential computing layer, with existing DeFi features on Solana . Blackthorn is positioned as a confidential AI layer that extends that DeFi stack.
The target customers tell you a lot. Arcium is aiming Blackthorn at healthcare, finance, law, and government. A lawyer feeding privileged documents into a chatbot, or a clinic uploading patient records, needs strong guarantees.
If Blackthorn delivers on encrypting prompts, weights, and outputs together, it addresses two worries at once. Users keep their data private. Model owners, meanwhile, can serve proprietary weights without handing them over in readable form.
There are open questions worth tracking. This is a private alpha, so performance, latency, and cost at production scale remain to be proven. The outer MPC layer is administered by Arcium, and how that administration is structured will shape how much trust regulated institutions are willing to extend.