Unlike ordinary end-to-end encryption, which keeps data unreadable to a server but also unusable by it, fully homomorphic encryption (FHE) allows real computation — including AI model inference — on ciphertexts. Only the client holding the decryption key can read the final result. Google's open-source HEIR compiler (announced August 2026) automates converting a trained model into an FHE-compatible version, though the technique still runs roughly 100x-1,000x slower than plaintext computation as of 2026.