Inference Engineering
Edge Deployment
8 min+80 XPStep 1/10
Why ship the model to the user
Running on-device inverts the economics and the physics of serving:
- Latency — no network round-trip; a phone chip answers in milliseconds
- Privacy — data never leaves the device, which unlocks features no server could offer credibly
- Cost — the user's hardware and battery pay the bill; your GPU-hour problem becomes their silicon problem
- Offline — works in airplanes, tunnels, and dead zones
The catch is stark. A phone SoC budgets ~5 W where an H100 draws 700 W, and offers maybe 8 GB of unified memory shared with the OS and every app. You don't get a 70B model. You get 3B, quantized, grateful for the NPU.