Here’s a question that should keep every enterprise AI team up at night: who else can see your data while it’s being processed in the cloud? The answer, for most setups, is uncomfortably long. Cloud providers, infrastructure operators, and anyone with privileged access to the hardware can, in theory, peek at data while models crunch through it.
Arcium thinks it has a fix. The Solana-based confidential computing network has announced Blackthorn, an infrastructure layer that encrypts data on Nvidia GPUs so that no single node or operator ever touches plaintext information. The pitch is straightforward: run AI inference and training at frontier scale, with privacy baked in, at speeds and costs comparable to traditional, non-confidential systems.
How Blackthorn actually works
The core technology relies on multi-party computation, or MPC. Instead of trusting one machine with your raw data, the workload gets split across multiple GPU nodes, each holding only an encrypted fragment. No single participant can reconstruct the original data.
What makes Blackthorn’s approach notable is that it runs on standard, off-the-shelf Nvidia hardware. Specifically, Arcium says it supports the Hopper H100 and H200 chips, as well as the newer Blackwell B200 series. Rather than depending on specialized trusted execution environments or custom silicon, Blackthorn achieves confidentiality through software.
