Google’s AICore eats gigabytes

Android’s AICore story is really about what happens when AI stops living in the cloud and starts living on your phone. The upside sounds great — faster responses, more privacy, some features that keep working offline. But the bill shows up in a place people understand immediately: storage. This week, Google finally explained why AICore can suddenly look enormous on Android devices, and the short version is that the bloat is mostly intentional. ### What is AICore, exactly? AICore is the Android system component that runs local generative AI features on supported devices, especially the Gemini Nano models Google uses for things like smart replies, summarization, proofreading, transcription, translation, and scam detection. It’s part of the push to do more AI work on the handset itself instead of sending everything back to Google’s servers. ### So why did people think something was wrong? Because AICore can look absurdly large for a background system app. Users have been spotting multi-gigabyte storage use, and reports around this week’s explanation pointed to cases as high as roughly 11GB. If you open storage settings and see a system service eating that much space with no obvious warning, “bug” is the natural conclusion. ### Why does the size jump happen? The jump happens during model updates. Google’s explanation is that when a new Gemini Nano version arrives, AICore downloads the new model in the background but also keeps the old one around for up to 3 days. Basically, your phone is temporarily carrying two copies of a very large AI model. If the fresh model has a problem, the phone can switch back immediately instead of forcing another giant download. That is a sensible reliability trick. Think of it like keeping the old parachute packed until you know the new one opens. The catch is that this safety margin is paid for with local storage, and users usually discover it only after the fact. ### Does the extra storage clear itself? Usually, yes. Google says the extra space should be removed automatically once the system decides the new model is stable. The 3-day window is the important part — not permanent duplication, but temporary duplication that can still feel permanent if you check at the wrong moment or run low on free space during that window. How is this different from a normal app update? Because normal app updates are small enough that most people never care. On-device AI models are not small. They are more like shipping a compact software package plus a chunk of trained weights that has to sit locally to deliver the privacy and speed pitch. That changes the user contract. “Runs on device” sounds elegant, but turns out it also means your device is now part data center. ### Is this only about storage? No — storage is just the easiest symptom to spot. Background model installs can also look like unexplained battery or CPU activity while the phone downloads, verifies, and swaps models. Even if the engineering logic is sound, silent multi-gigabyte behavior creates a trust problem when users aren’t told clearly what is happening and why. ### Why does this matter beyond one Google app? Because this is the real economics of on-device AI showing through. The industry has spent a year selling local AI as private, instant, and convenient. All of that can be true. But local AI also consumes scarce things — storage, bandwidth, battery, thermal headroom — and platform makers now have to make those costs legible instead of hiding them behind system labels. The bottom line is simple: AICore probably is not “broken” when it balloons. It is doing exactly what Google designed it to do. But that design choice makes a bigger point — on-device AI is not free magic. It takes up room, and now users can see the footprint.