Arcium Launches Blackthorn: Encrypted AI Inference on Any NVIDIA GPU via MPC

Arcium has launched Blackthorn, a confidential AI inference framework that enables any NVIDIA H100, H200, or B200 GPU to process AI workloads with inputs, model weights, and outputs remaining fully encrypted throughout computation. Arcium announced early access on June 26, with first real-world pilots expected in the coming weeks.

Blackthorn addresses a structural problem in enterprise AI adoption: organizations in healthcare, finance, law, and government hold sensitive data they cannot legally or competitively share with a third-party AI operator in plaintext. The framework targets this constraint by ensuring neither the cloud provider, the operator, nor the Arcium network itself can see raw data at any point during inference.

How Blackthorn Encrypts AI Inference on NVIDIA GPUs

The core mechanism pairs two encryption layers with Arcium's multi-party computation network. A user's prompt or dataset is first encrypted with an inner key. Arcium then adds an outer encryption layer before the data reaches an untrusted relay host CPU connected to an NVIDIA GPU. The host CPU passes the doubly encrypted payload to the GPU without ever seeing plaintext.

Inference runs on the NVIDIA GPU inside what Arcium describes as its encrypted memory region, a confidential computing mode where the GPU processes data it cannot expose. The decentralized Arcium network coordinates computation across independent nodes using secure multi-party computation (MPC), splitting inputs into cryptographic secret shares so no single participant can reconstruct the original data. Results returned to the user are verifiable without revealing the underlying inputs.

This architecture replaces Intel and AMD CPU-based Trusted Execution Environments (TEEs) with MPC, which Arcium describes as "maliciously secure" and resistant to the supply-chain and firmware attacks that have historically compromised hardware-based approaches.

The scale argument for this approach is straightforward. Arcium claims this software-only method expands the addressable confidential compute pool by approximately 2,500 times, from roughly 2,000 purpose-built confidential GPU units to the roughly 5 million NVIDIA H100, H200, and B200 units already deployed in data centers globally. Blackthorn can therefore be provisioned through existing cloud infrastructure rather than requiring customers to source specialized hardware.

Healthcare, Finance, and Sovereign AI: Target Sectors for Encrypted Inference

The four markets Arcium is targeting are defined by their inability to share data with AI operators under current architectures.

Healthcare organizations need inference on patient records, genomic sequences, and clinical trial data without exposing raw inputs to the model operator or the inference provider. Financial institutions need analytics and model queries that comply with data residency requirements such as GDPR without data leaving a controlled boundary. Legal and government use cases include confidential document processing and sovereign workloads that cannot run on shared cloud infrastructure.

The fourth category is private RAG and collaborative AI: multiple institutions, including banks, pharmaceutical companies, or research consortia, can collectively run inference over pooled sensitive data and extract shared insights without any party seeing another's raw inputs. Only the verifiable result is returned.

Arcium has cited the market for regulated-sector AI as a trillion-dollar opportunity for confidential compute. The company has raised more than $15 million from investors including Jump Crypto and Coinbase Ventures, and was admitted to NVIDIA's Inception Program in 2025.

Arcium Mainnet Has Passed 1 Million Confidential Computations

Blackthorn arrives on the back of Arcium's mainnet alpha reaching operational scale. According to a June press release, the network has processed more than 1 million confidential computations and executes over 200,000 daily computations, which Arcium claims runs 20 times faster than competing confidential compute approaches.

The ZINC protocol, built on Arcium, reached Solana's top-three protocols by 24-hour revenue in June, attracting $18 million in deployed capital within two weeks and generating $1.8 million in fees during the same period.

Arcium launched its ARX token on June 22 on Binance Alpha with approximately 209 million tokens in initial circulation, representing 20.88% of the 1 billion fixed supply. ARX trades around $0.29 as of June 27, with a market cap near $293 million per Solana Compass token data.

peaq Deploys Arcium for Confidential Compute on Robot Fleets

Alongside the Blackthorn announcement, peaq, an IoT and DePIN coordination layer, announced that Arcium is now live on its robotic.sh services platform on Solana. The integration enables robot fleets running peaqOS to perform confidential computation on sensitive operational data without exposing it to any party.

The integration demo involves a Unitree G1 warehouse robot completing a shift, transmitting its operational data to Arcium as encrypted input, and receiving a verifiable result with no raw data exposed. peaqOS handles machine identity, permissions, and job coordination via Solana transactions; Arcium runs the confidential computation.

The partnership extends Arcium's use cases beyond human-directed AI inference into machine-to-machine data processing, covering fleet learning, insurance underwriting from private telemetry, and machine credit ratings built from performance data that operators cannot currently monetize because sharing it would compromise competitive information.

Arcium's Approach vs. TEE-Based Confidential AI

The more interesting technical question Blackthorn raises is whether MPC can displace TEEs as the standard approach for confidential AI inference. TEE-based confidential computing, which uses dedicated hardware enclaves in Intel or AMD CPUs, has been the dominant method for protecting data during processing, but those hardware enclaves have been compromised repeatedly through supply-chain, firmware, and side-channel attacks.

Arcium's claim is that MPC, by distributing computation across independent nodes that collectively agree on a result without any one node seeing the full data, removes the single hardware point of failure that TEEs carry. Whether that holds under production load at scale, and at what cost in latency and throughput relative to unencrypted inference, are questions the early access period is presumably designed to answer.

As Arcium described the product: Blackthorn "runs the most powerful AI models with every prompt, file, and answer encrypted from the cloud, the operator, and even Arcium."

Arcium runs on Solana, which provides the settlement and verification layer for computation jobs. For background on the ARX token launch and Arcium's mainnet alpha, see our coverage of the June 22 ARX TGE and Arcium's mainnet alpha launch.