A confidential GPU fails at 3 a.m. Who brings the service back, how do clients know the new node can be trusted, and can anyone log in to fix it?
Hardware-based Confidential Computing protects data during execution. But it does not, by itself, address the operational challenges of running Confidential AI in production.
A production service also needs attestation, key management, orchestration on attested nodes, trusted updates, and recovery without operator access.
2 𝐨𝐟 𝐭𝐡𝐞 11 𝐥𝐚𝐲𝐞𝐫𝐬 𝐧𝐞𝐞𝐝𝐞𝐝 𝐟𝐨𝐫 𝐚 𝐜𝐨𝐦𝐩𝐥𝐞𝐭𝐞 𝐂𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐭𝐢𝐚𝐥 𝐀𝐈 𝐨𝐟𝐟𝐞𝐫𝐢𝐧𝐠 𝐜𝐨𝐦𝐞 𝐰𝐢𝐭𝐡 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 𝐓𝐡𝐞 𝐨𝐭𝐡𝐞𝐫 9 𝐚𝐫𝐞 𝐬𝐨𝐟𝐭𝐰𝐚𝐫𝐞 𝐭𝐡𝐚𝐭 𝐬𝐨𝐦𝐞𝐨𝐧𝐞 𝐡𝐚𝐬 𝐭𝐨 𝐛𝐮𝐢𝐥𝐝 𝐚𝐧𝐝 𝐨𝐩𝐞𝐫𝐚𝐭𝐞.
Last week, we announced two UK projects using NVIDIA Confidential Computing.
BE, a GPU cloud provider, added verifiable confidential LLM inference to its existing service. Reset Health validated AI on sensitive treatment data using externally hosted NVIDIA Blackwell systems.
𝐍𝐞𝐢𝐭𝐡𝐞𝐫 𝐡𝐚𝐝 𝐭𝐨 𝐛𝐮𝐢𝐥𝐝 𝐢𝐭𝐬 𝐨𝐰𝐧 𝐂𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐭𝐢𝐚𝐥 𝐀𝐈 𝐬𝐭𝐚𝐜𝐤. 𝐒𝐮𝐩𝐞𝐫 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 𝐩𝐫𝐨𝐯𝐢𝐝𝐞𝐬 𝐚𝐥𝐥 11 𝐥𝐚𝐲𝐞𝐫𝐬 𝐚𝐬 𝐨𝐧𝐞 𝐩𝐥𝐚𝐭𝐟𝐨𝐫𝐦.
Why a confidential server is not a confidential cloud: superprotocol.com/platform/confi…
Hardware-based Confidential Computing protects data during execution. But it does not, by itself, address the operational challenges of running Confidential AI in production.
A production service also needs attestation, key management, orchestration on attested nodes, trusted updates, and recovery without operator access.
2 𝐨𝐟 𝐭𝐡𝐞 11 𝐥𝐚𝐲𝐞𝐫𝐬 𝐧𝐞𝐞𝐝𝐞𝐝 𝐟𝐨𝐫 𝐚 𝐜𝐨𝐦𝐩𝐥𝐞𝐭𝐞 𝐂𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐭𝐢𝐚𝐥 𝐀𝐈 𝐨𝐟𝐟𝐞𝐫𝐢𝐧𝐠 𝐜𝐨𝐦𝐞 𝐰𝐢𝐭𝐡 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 𝐓𝐡𝐞 𝐨𝐭𝐡𝐞𝐫 9 𝐚𝐫𝐞 𝐬𝐨𝐟𝐭𝐰𝐚𝐫𝐞 𝐭𝐡𝐚𝐭 𝐬𝐨𝐦𝐞𝐨𝐧𝐞 𝐡𝐚𝐬 𝐭𝐨 𝐛𝐮𝐢𝐥𝐝 𝐚𝐧𝐝 𝐨𝐩𝐞𝐫𝐚𝐭𝐞.
Last week, we announced two UK projects using NVIDIA Confidential Computing.
BE, a GPU cloud provider, added verifiable confidential LLM inference to its existing service. Reset Health validated AI on sensitive treatment data using externally hosted NVIDIA Blackwell systems.
𝐍𝐞𝐢𝐭𝐡𝐞𝐫 𝐡𝐚𝐝 𝐭𝐨 𝐛𝐮𝐢𝐥𝐝 𝐢𝐭𝐬 𝐨𝐰𝐧 𝐂𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐭𝐢𝐚𝐥 𝐀𝐈 𝐬𝐭𝐚𝐜𝐤. 𝐒𝐮𝐩𝐞𝐫 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 𝐩𝐫𝐨𝐯𝐢𝐝𝐞𝐬 𝐚𝐥𝐥 11 𝐥𝐚𝐲𝐞𝐫𝐬 𝐚𝐬 𝐨𝐧𝐞 𝐩𝐥𝐚𝐭𝐟𝐨𝐫𝐦.
Why a confidential server is not a confidential cloud: superprotocol.com/platform/confi…