Arcium has brought privacy AI to a trial stage, but the performance metrics still need further evaluation. The key focus of Blackthorn is how sensitive data enters the model service. ① On October 21, Arcium announced that Blackthorn has entered a private alpha phase for select users, offering a privacy-preserving AI inference pathway. This remains a restricted test and should not be interpreted as universal access to a production-ready service for all developers. ② Official materials describe usage methods including APIs, dedicated GPUs, and proprietary hardware, and propose splitting GPU channel keys via secure multi-party computation. This involves trust design for data transmission and execution environments, requiring full understanding of the threat model. ③ A boundary often overlooked in marketing is that reducing reliance on CPU TEEs does not eliminate hardware trust entirely—the solution still depends on NVIDIA GPU confidential computing capabilities. ④ Descriptions such as “free trial” and “low overhead” are contingent on specific phases or project testing conditions and cannot be treated as independent performance benchmarks at public scale, nor can they be used to infer that all AI computations occur on the Solana chain. My own view: Whether this approach succeeds depends on whether customers are willing to pay for data confidentiality. I’ll be monitoring reproducible latency, throughput, fault recovery, and real-world customer cases—particularly seeking direct comparisons with conventional inference solutions. @Solana_zh @solana @Arcium