Arcium has launched a new infrastructure for AI, which seeks to extend its original capabilities for confidential computing to a broader model reasoning landscape. The company disclosed that the Blackhorn system could be operated on the existing NVIDIA GPU and HGX servers and that the hints, upload files, models and output were encrypted during the AI reasoning process.
Main deduction all-way encryption
Arcium states that Blackhorn's goal is to prevent cloud service providers, infrastructure operators and Arcium itself from reading user data. According to its description, the user input is encrypted with the user key, followed by a layer of encryption, then enters the host environment that connects to the NVIDIA GPU, without revealing the specified content.
According to the company, this design responds primarily to concerns about AI privacy in enterprises and institutions. Traditional AI services usually require that the content of the alerts, documents, etc., be handed over to the model operator, which can form a centralized and sensitive data pool, as well as risk leakage and unauthorized access.
Replace hardware trust with MPC
Blackthorn ' s core programme is to replace reliance on a specific hardware security environment with multiple calculations (MPCs). In Arcium ' s view, the existing confidential computing programme relies heavily on the credible implementation environment of Intel or AMD, which in the past has seen security gaps in such components.
Arcium disclosed that Blackthorn no longer built security on a single hardware manufacturer, but instead turned to a password assurance. In other words, the system wants to complete the AI reasoning in encryption without trusting a single host or operator.
Existing GPU directly accessible
Arcium judges that the currently available AI GPU is still very small in size, about $2,000, and only supports about 4 million users. This is due to the fact that existing deployments are usually built around dedicated security chips, which are more expensive to expand.
Blackthorn is trying to circumvent this limitation through software upgrades. According to the company, the programme could operate directly on the existing nudity NVIDIA GPU and HGX servers without the need for additional specialized hardware, thus increasing the available security algorithm approximately 2,500 times and involving millions of deployed AI chips in encryption missions.
For medical and financial scenarios
Arcium considered that such infrastructure would be more appropriate for industries with high data confidentiality requirements. By way of example, hospitals can process patient and genetic data in a state of encryption, financial institutions can analyse transaction strategies and client data, and governments, defence agencies and law firms can operate tasks without exposing information at the bottom.
The company also disclosed its follow-up product plans, including Blackthorn Uniteds, an encryption model for developers, and bLLM running hours that could be installed directly on the NVIDIA Hopper and Blackwell GPU infrastructure.
In February this year, Arcium went online to the main network Alpha, which is located in the solana ecological privacy calculator. According to the company, its governance and functional token ARX had been online on 22 June in Solana; the cumulative number of network security calculations had exceeded 1 million, with more than 200,000 being processed today.
Additional information:Blackthorn is an extension of Arcium ' s product from block chain security calculations to AI infrastructure, and at this stage the relevance and actual deployment effects continue to come from unilateral corporate disclosures.