When the AI ​​feature runs right on the device, you still have to pay a monthly fee

This week there are two stories that seem to have nothing to do with each other, but reading them side by side I see they are telling the same story. Apple silently recorded in the release notes of macOS 27 beta 3 that if it wants to use Apple Intelligence features in the Home app: such as summarizing security camera activities in natural language, searching for videos by description, users must have an iCloud + 2TB package, or 279 thousand VND / month. At the same time, Meta limits the Conversation Focus feature on Ray-Ban smart glasses, which helps filter and clarify voices in noisy environments, to 3 hours of use per month, unless you pay an additional $20 (I can't find an equivalent price for Vietnam) for the Meta One Premium package to get 15 hours. The two are superficially similar: both are AI features on purchased hardware, and both are placed behind a paywall. But digging deeper into the technical reasons behind, I see that they are two very different stories, and it is that difference that is worth mentioning.

Apple: the cost is real, but still chooses the most annoying way to collect fees

Apple's case also has a tenable technical explanation. The AI ​​features for the Home app, according to Apple itself, run on "powerful server models", meaning that when the camera records a video, that data must be processed through AI models located on Apple's servers to generate natural language descriptions, rather than being processed entirely on the device. Twitter processing on the server means there is a real compute cost, per request, and that cost increases linearly with the number of cameras and the number of hours recorded by the user.

Meta: when the limit does not come from server costs

The Meta case is much harder to justify. Conversation Focus is a feature that runs entirely right on the chip of Ray-Ban glasses, does not send data to the cloud, and does not consume Meta's server resources with each use. In other words, the marginal cost for a user to run the feature for an extra hour is virtually zero, because all the processing happens on the hardware the user has already paid for.

So where does the 3 hour per month limit come from? The Meta representative explained that the paid plan is "primarily for heavy users who need additional limits and support," but this interpretation doesn't really answer the question of why an offline feature needs a "limiter" in the first place. Without real operating costs behind it, usage limits only serve as a price discrimination tool, a way for Meta to divide users into groups willing to pay different prices for the same thing, based on how much or how little they need it, not based on how much it actually costs Meta to serve them. Community reaction, especially from deaf users who rely on this feature more than normal people, shows that this discomfort is not just emotional.

Hardware is like a gate, no longer the final product

Looking beyond these two specific cases, I think what is happening is a familiar shift in the history of the consumer hardware industry, only this time wearing AI. The home security camera industry has been down this path before: nearly every smart camera brand sells hardware at low or break-even prices, then collects real money via cloud storage plans or monthly smart recognition features. Apple's HomeKit Secure Video itself was once considered a rare exception, as video storage was included in the available iCloud package instead of a separate service, and that was exactly the reason why users chose the Apple ecosystem over other camera brands in the first place. The fact that Apple is now putting its latest AI feature behind an additional paywall shows that even that exception is gradually eroding. For Meta, the story is even more symbolic because smart glasses are still a young product category, where the company is trying to build trust with users to accept wearing a device with a camera and microphone constantly on their face. Putting a core feature, which costs nothing to operate, behind a paywall at this early stage is a clear signal that the hardware, even when sold outright to users, is still seen as just a gateway into a recurring revenue stream, and no longer a complete product in and of itself. If this trend spreads, the real question is not "which AI feature will be free next," but whether consumers will still be able to distinguish between what's a real operating cost, and what's just a price set because the company knows they can get away with it. Note: avatars are created by AI based on the article content