Phone features have historically been paid for once, in the price of the device. Generative features break that model because they keep costing money after the sale.
Inference is an ongoing expense
Every request served from a data center consumes computation, and that computation has a cost proportional to how much text or imagery is produced.
A heavy user can consume many times what an average user does, so the cost per customer varies far more than for a traditional software feature.
Bundling that into a device price would mean charging every buyer for the heaviest users' consumption, which prices the hardware badly.
Hardware margins cannot absorb it indefinitely
A phone's margin is realized once, while a device commonly stays in use for several years and may change owners during that time.
A feature included for the device's life therefore has an unknown total cost against a fixed one-time revenue.
Manufacturers manage that by offering the feature free for a promotional period, which defers the decision rather than resolving it.
On-device processing changes the arithmetic
Features that run locally cost the manufacturer nothing per use, since the buyer already paid for the silicon and supplies the electricity.
This is why local capabilities tend to be free and permanent, while server-backed ones tend to be metered or bundled with a plan.
It also explains the investment in dedicated processing hardware. Moving work on-device converts a recurring cost into a one-time component cost.
Bundling shifts who pays
Carriers and manufacturers increasingly include AI services inside broader subscriptions covering storage, backup and support.
Bundles smooth the variability, since light users subsidize heavy ones inside a single price, which is how service pricing has always handled uneven consumption.
The visible effect is that a feature demonstrated at launch becomes conditional later, tied to a plan the buyer must maintain.
Free tiers set expectations that get revised
Introductory access builds usage habits and produces the data needed to size real demand, which is worth more than the revenue forgone.
Once the cost curve is understood, limits appear: request caps, slower queues at peak, or restriction to a subset of the original capability.
Buyers reading a feature list are therefore reading two categories at once, and the distinction between them is whether the work happens on the phone.