Google: edge plus cloud muscle

Google just put two opposite bets on the table at once: a free iPhone app that works offline, and a cloud empire that outside researchers say holds about one quarter of the world’s artificial intelligence compute. Those bets only look contradictory if you think every artificial intelligence task has to run in one place. (computerworld.com) (epoch.ai) The new app is called Google AI Edge Eloquent, and Google says it runs speech processing locally on the phone instead of sending audio to a server. Apple’s App Store listing says the app is free, iPhone-only, 67.1 megabytes, and designed to turn messy spoken dictation into cleaned-up text. (ai.google.dev) (apps.apple.com) That local setup changes two things people notice immediately: speed and privacy. Google’s listing says audio and personal data stay on the device for the core experience, and the app keeps working when the phone has no connection. (apps.apple.com) (computerworld.com) Google has been building toward this for months with Google AI Edge Gallery, a demo app for running open-weight models on phones. In February 2026, Google showed a 270 million parameter model called FunctionGemma that could turn plain-language commands into actions like opening maps or creating a calendar event without any server call. (developers.googleblog.com) (github.com) Running a model on a phone is like keeping a pocket calculator in your backpack instead of calling a data center every time you need arithmetic. It is great for short jobs, but it is still a backpack, which is why Google’s cloud side matters at the same time. (developers.googleblog.com) (docs.cloud.google.com) On the cloud side, Network World reported this week that hyperscalers now host more than 60 percent of global artificial intelligence compute, with Google in the lead. Epoch AI estimated Google held about 23 percent of global cumulative artificial intelligence compute capacity by the fourth quarter of 2025, largely through its own Tensor Processing Units rather than Nvidia graphics chips. (networkworld.com) (epoch.ai) Those Tensor Processing Units are Google’s custom chips for matrix math, which is the repetitive number-crunching that trains and serves modern models. Google Cloud’s documentation says one TPU v5e chip delivers 197 trillion floating-point operations per second in brain floating point 16 format, and a full 256-chip pod reaches 50.63 quadrillion floating-point operations per second. (docs.cloud.google.com) Google also designed that chip line for both training and serving, which is the difference between teaching a model and answering a user in real time. Google says TPU v5e is a combined training and inference product, with training tuned for throughput and serving tuned for latency. (docs.cloud.google.com) Put those pieces together and the shape of the market gets clearer. A startup can keep the first hop on the device for instant replies, offline use, and private data, then hand bigger jobs to cloud hardware when the task needs a larger model, more memory, or many users at once. (computerworld.com) (networkworld.com) (docs.cloud.google.com) That is why Google’s tiny dictation app and its giant chip fleet belong in the same story. The phone handles the part that feels personal and immediate, and the cloud handles the part that is too big to fit in your pocket. (apps.apple.com) (epoch.ai)