Why the Neural Engine isn't doing the work
The chat on this site runs the model on your graphics chip. The Neural Engine — the dedicated AI chip in an iPad Pro, iPhone, or Mac — is reserved for installed apps. No web page, in any browser, can reach it today.
Why the door is closed
Apple only lets software use the Neural Engine through Core ML, which is an interface for installed apps. Safari deliberately does not pass that through to web pages — partly for security, partly because there is no agreed web standard for it yet.
What browsers do offer is the graphics chip, through a standard called WebGPU. That is genuinely fast and it is what this demo uses. It just isn't the specialised chip sitting next to it.
What a native app would change
An iPad app could show this exact chat page inside itself and answer using the on-device model built into recent versions of iPadOS. That model runs on the Neural Engine: quicker replies, less battery drain, and no multi-hundred-megabyte download, because the model is already on the device.
Side by side
A proposed web standard called WebNN would let pages use neural accelerators directly. It is being trialled in some browsers, but Safari does not support it today — so on an iPad, the graphics chip remains the fastest route a web page has.