Imagine a commuter who translates a sign offline in two seconds with no signal. That's the whole argument. Today's flagship phones run 7-8 billion parameter models locally, and a round trip to a data center adds 50 to 200 milliseconds. For real-time tasks, the physics of distance and bandwidth don't care how big your racks are. Inference is shifting to the edge because latency and privacy are hard constraints, not preferences. The cloud won't vanish, but its role as the default compute layer for personal AI is already shrinking.

Inference is where the usage actually lives, and that's what edge silicon keeps absorbing. My opponent's training-cost figure is real, and I concede the cloud keeps training. But once a model exists, every translation and summary is inference, and published benchmarks show capability per watt on device doubling roughly yearly. For the cloud to stay the default, my opponent needs evidence its inference share holds against that curve, not a calculator analogy. Hybrid looks like a transition, not a destination.

My opponent's gravity analogy actually cuts my way. Training is a one-time capital cost, amortized across billions of inferences. The recurring compute, what you're billed for every second, is inference, and that's exactly what edge silicon keeps absorbing. On the train strike: a device doesn't need full cloud weights, just a thin fresh-data lookup. That's retrieval, not a data center doing the thinking. So the cloud keeps gravity at the training layer. But gravity isn't where the mass of daily compute lives.