Browser AI needs a capability ladder, not a WebGPU requirement_

In-browser inference can improve privacy and remove API cost, but device support, memory, first-load size, and thermal behavior vary widely.

A product that works only on the founder’s laptop is a demo. Visitors need a useful experience when WebGPU is unavailable or the model cannot fit.

What changes in practice_

Detect capability, offer smaller profiles, cache progressively, keep retrieval narrow, and provide a deterministic non-model fallback.

• Test low-memory devices.
• Make download size visible.
• Never block core site navigation on model startup.

Our take_

The durable advantage is not adopting the newest tool first. It is building the identity, state, evidence, and operating boundaries that let a real team own the system after launch.