SAN JOSE, Calif., Sept. 16, 2025 /PRNewswire/ — At the AI Infra Summit 2025, Cornami, a leader in scalable computing architectures, and DESILO, a privacy-enhancing technology (PET) startup, announced the deployment of a fully homomorphic encryption (FHE) based large language model (LLM). The solution processes sensitive data while it remains encrypted, delivering real-world speed and accuracy.
Breaking the Privacy vs. Performance Dilemma
From healthcare to financial services, enterprises have faced the same tradeoff: stronger data protection typically slows AI performance, while faster insights often compromise security. This collaboration addresses the long-standing dilemma by enabling encrypted AI inference at near plaintext speeds.
“For decades, FHE was considered too slow for real deployment. By accelerating encrypted computation, we are proving that enterprises can have both privacy and performance.” Dr. Craig Gentry, Chief Scientist of Algorithms for Cornami, also known as the father of FHE, has been leading key innovations to provide unique value to this industry for Cornami and its software partners. One such innovation is plaintext ciphertext matrix multiplication (PCMM) which affects 90%+ of LLM compute. PCMM allows these operations to be executed securely with extremely low overhead. This innovation together with scalable acceleration using Cornami’s Compute Fabric results in orders-of-magnitude faster performance over traditional encrypted compute approaches. “The value of PCMM is that it makes privacy-preserving AI practical by enabling matrix multiplication — the core of LLM compute — to be executed efficiently and securely, closing the performance gap between fully encrypted and plaintext LLM inference computation while strengthening compliance, sovereignty, and post-quantum security,” stated Dr. Craig Gentry.
Read more at prnewswire.com