Confidential computing network Arcium has introduced Blackthorn, a new confidential AI infrastructure designed to transform existing NVIDIA GPUs into what the company describes as “one encrypted supercomputer”. The launch marks Arcium's expansion beyond blockchain-focused confidential computation into artificial intelligence, with an emphasis on protecting sensitive data throughout the entire AI workflow.
According to Arcium, Blackthorn encrypts every prompt, uploaded file, AI model, and generated output end-to-end. The company says this prevents cloud providers, infrastructure operators, and even Arcium itself from accessing user data during AI inference.
The announcement addresses growing concerns surrounding AI privacy. Arcium argues that conventional AI services require users to expose prompts and files to the companies operating the models, creating large repositories of sensitive information that could become targets for breaches or unauthorized access.
Replacing Hardware Trust With Cryptography
A central feature of Blackthorn is its approach to confidential computing. Rather than relying on Intel or AMD Trusted Execution Environment technologies, which Arcium describes as security components with a history of vulnerabilities, Blackthorn replaces the traditional trusted host CPU with maliciously secure Multi-Party Computation (MPC).
Under the system, user prompts, documents, and other inputs are encrypted with the user's key before Arcium adds an additional layer of encryption. The encrypted information then passes through an untrusted host CPU connected to an NVIDIA GPU without exposing any plaintext data.
Inference executes inside the NVIDIA GPU's confidential computing mode, where encrypted memory protects workloads during processing. The generated output returns through the same encrypted process, allowing only the user to decrypt the results.
Arcium describes this architecture as replacing trust in individual hardware vendors with cryptographic guarantees that no single company controls.
2500× Increase in Confidential AI Without New Hardware
One of Blackthorn's primary goals is increasing the availability of confidential AI infrastructure. The company estimates that only about 2,000 confidential AI GPUs are currently accessible worldwide because existing confidential computing deployments require specialized hardware configurations built around trusted security chips. According to Arcium, that capacity supports only about 4 million users, a small figure compared with the rapidly expanding global AI market.
Blackthorn instead works on existing bare-metal NVIDIA GPUs and HGX servers via a software upgrade rather than requiring new hardware deployments. Arcium says this creates a 2,500× increase in confidential compute capacity by enabling millions of already deployed AI chips to participate in confidential workloads.
Targeting Sensitive Industries
Arcium says Blackthorn could expand AI adoption in sectors limited by privacy concerns. The platform keeps personal data, documents, and business records hidden from AI providers.
In healthcare, hospitals can process patient and genomic data while keeping it encrypted. Financial institutions can analyze trading strategies and customer data without exposing sensitive information. Governments, defense groups, and law firms can also handle highly sensitive data while maintaining full encryption throughout computation.
Built on Arcium's Confidential Computing Foundation
Blackthorn builds upon Arcium's broader confidential computing platform, which focuses on enabling secure computation without exposing sensitive inputs. The network uses Multi-Party Computation (MPC) to enable applications to compute on confidential information while preventing any individual participant from accessing the underlying data.
Arcium outlined several products that will become available in the future. The company plans to launch Blackthorn Units, which will provide dedicated confidential compute powered by bLLM, Arcium's confidential AI runtime. A separate encrypted model API will allow developers to access major open-weight AI models without deploying their own infrastructure. Organizations that already operate NVIDIA Hopper or Blackwell GPU infrastructure will also be able to install bLLM directly onto their existing hardware to enable confidential AI without purchasing additional equipment.
Arcium launched its mainnet alpha in February and has positioned itself as a privacy layer for applications across the Solana ecosystem. The Blackthorn announcement follows another milestone for the project. Arcium's governance and utility token, $ARX, launched on Solana on June 22. The network has now surpassed 1 million cumulative confidential computations and processes more than 200,000 confidential computations each day.
Read More on SolanaFloor
Crypto ETFs Lose $5 Billion in 30 Days as Bitcoin, Ethereum, and Solana Flows Turn Negative
26 Solana Frontier Winners Revealed After Crypto’s Biggest Hackathon Ever
Has This Been Solana’s Biggest Mistake?
