{"slug": "google-open-sourced-a-compiler-that-runs-ai-on-data-the-server", "title": "Google Open-Sourced a Compiler That Runs AI on Data the Server", "summary": "Google has open-sourced HEIR, a compiler that enables AI model inference on encrypted data without the server ever decrypting it, marking a practical step toward homomorphic encryption for AI. The compiler optimizes models to run under homomorphic schemes, addressing the historical performance gap, though it remains slower than plaintext inference. This could allow regulated industries like healthcare and finance to process sensitive data in place, avoiding privacy breaches and compliance issues.", "body_md": "# Google Open-Sourced a Compiler That Runs AI on Data the Server\n\nGoogle open-sourced HEIR, a compiler that runs model inference on encrypted data the server never decrypts. It points at a future where AI can process…\n\nEvery AI privacy story you've read this year has the same shape: a company stored your data, and someone got to it. Google just shipped the technical foundation for a different shape entirely, one where the server never holds your data in readable form in the first place.\n\nIt's called HEIR, and Google open-sourced it last week. It's a compiler that runs model [inference](/glossary/inference) on encrypted data. The server does the math. It produces an encrypted answer. It never decrypts anything.\n\n## The Trick That Took Four Decades to Get Practical\n\nThe technique is homomorphic encryption, and it's been a cryptographer's white whale for forty years. The idea is simple to state and brutal to implement: perform useful computation directly on encrypted values, so that adding two encrypted numbers produces an encrypted sum, without anyone ever seeing the numbers.\n\nThe catch has always been speed. Fully homomorphic encryption made early implementations thousands to millions of times slower than plaintext, which kept it on paper. HEIR is a compiler, not a new cipher. It takes the model, rewrites it into a form that runs efficiently under homomorphic schemes, and optimizes the parts that have historically killed performance.\n\nIt won't be fast enough for a general [chatbot](/glossary/chatbot) tomorrow. But for the workloads where privacy is the whole point, it changes the calculus.\n\n## What It Actually Unlocks\n\nThink about the data that's genuinely stuck behind walls right now. Medical records that can't leave a hospital. Financial transactions that can't cross a border. Government data that can't touch a third-party cloud.\n\nRight now, running AI on any of that means either decrypting it, which is a privacy hole, or shipping it to a vendor, which is a compliance nightmare. HEIR points at a third option: process the encrypted data in place and get an encrypted result, where the model owner never sees the input and the data owner never hands over the plaintext.\n\nThat's not a niche trick. It's the answer to the question of how regulated industries adopt AI without giving up their single most sensitive asset, which is the data itself.\n\n## The Honest Part\n\nI'm not going to oversell the timeline. Homomorphic inference is still painfully slow compared to plaintext, and HEIR is early-stage. The gap is narrowing, but a model that takes a millisecond in the clear might take minutes under homomorphic encryption, and for most consumer apps that's a dealbreaker.\n\nBut here's the thing. The companies holding the most valuable data are exactly the ones that can afford to wait a minute for an answer if it means they never expose the question. A hospital that can run a diagnostic model on encrypted patient records without a business associate agreement, without a data transfer, without a single plaintext byte leaving the building, will happily trade latency for that.\n\nWatch this space. The AI privacy conversation has been stuck on policy and promises. HEIR is the first real sign that the bottleneck is becoming an engineering problem instead, and engineering problems are the ones the industry actually solves.\n\n*Sources: Google HEIR open-source release, August 2026; AI Tools Recap daily briefing, August 16, 2026.*\n\nGet AI news in your inbox\n\nDaily digest of what matters in AI.", "url": "https://wpnews.pro/news/google-open-sourced-a-compiler-that-runs-ai-on-data-the-server", "canonical_source": "https://www.machinebrief.com/news/google-heir-encrypted-inference-compiler-homomorphic", "published_at": "2026-08-21 13:04:36+00:00", "updated_at": "2026-08-21 13:15:10.164454+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-infrastructure", "ai-safety"], "entities": ["Google", "HEIR"], "alternates": {"html": "https://wpnews.pro/news/google-open-sourced-a-compiler-that-runs-ai-on-data-the-server", "markdown": "https://wpnews.pro/news/google-open-sourced-a-compiler-that-runs-ai-on-data-the-server.md", "text": "https://wpnews.pro/news/google-open-sourced-a-compiler-that-runs-ai-on-data-the-server.txt", "jsonld": "https://wpnews.pro/news/google-open-sourced-a-compiler-that-runs-ai-on-data-the-server.jsonld"}}