Google Launches HEIR Compiler to Make Homomorphic Encryption Practical for AI

What happened

On August 14, 2026, Google announced HEIR (Homomorphic Encryption Intermediate Representation), an open-source compiler that converts pretrained AI models to run directly on encrypted data. It ships with four demos: a recommendation model, credit-card fraud detection, network intrusion detection using Kitsune, and hotword spotting in audio. Performance overhead has fallen from about 100,000x slower than native in 2020 to about 1,000x slower in 2025 estimates. Google also partnered with Belfort Labs, Niobium, Cornami, and Optalysys to test hardware accelerators through the HEIR toolchain.

Why it matters

HEIR removes most of the burden of manually rewriting code for homomorphic encryption, enabling hospitals, banks, and content platforms to run AI on encrypted data without ever seeing raw user information. The cryptographic guarantee is stronger than hardware enclaves. This release signals that private AI at scale has crossed from research curiosity to engineering reality, with four peer-reviewed papers already building on HEIR.