Lucid Computing

Lucid Computing describes its offering as "Verifiable AI. Proven in hardware.": privacy-preserving, verifiable compute clusters built on trusted execution environments S-1406. Its verification work includes:

• Location certificates. Lucid hosts, in its GitHub organization, the draft Sovereignty Certificates specification, which packages ping-based chip location verification as short-lived certificates S-1404; its homepage links to the effort S-1406. See Lucid sovereignty (location) certificates.
• Developer platform. Its documentation lists a "Data Sovereignty & Localization" auditor, described as "Ensuring data remains within approved geographic jurisdictions" S-1407.
• Traffic shaping. A Lucid design brief proposes capping the external traffic of each pod of accelerators, on the grounds that inference sends little external traffic while training moves large volumes between machines S-1301. The brief says the design is not yet implemented or red-teamed S-1301. See Bandwidth limits and compartmentalization.
• Research cluster. Lucid reports that it builds and operates a bare-metal GPU research cluster on behalf of the Verifiable Compute Foundation, which decides who gets access S-1701 S-1702. VCF reports two H100 nodes operational in a beta launch S-0091. Lucid's cluster page announces a community opening in November S-1702.
• Red-teaming and standards. Lucid reports running red-teaming programmes with national security agencies and government research institutes, in which the teams attacking its architecture are not the teams that built it S-1701. It reports building a testbed inside an unnamed government-funded research institute, whose first configuration retrofits two eight-GPU H100 servers to test claims about where compute runs, how it is used and which model is loaded S-3363. Lucid says the institute sets the test plan S-3363. It lists standards in development for compute identity, attested audits and hardware-rooted compute accounting S-1701.