AI Could Move From The Cloud To The Phone – What Does That Mean For App Developers?

Most of the AI features in the apps on your phone don’t run on your phone. Each request travels to a data centre, gets processed there and comes back as an answer, and the developer behind the app usually pays for every trip.

That could start to change. AI companies are “starting to do phones”, according to Qualcomm CEO Cristiano Amon, who says some of them have asked for handsets that can run an AI model of at least 100 billion parameters, more or less all the time, by 2028.

The companies weren’t named and no chip has been announced to meet the request, so it’s best read as a sign of where the industry wants to go rather than a product promise.

This directional change is noteworthy for early-stage companies whose operating costs rely on cloud service fees, leaving founders with a pressing dilemma: how do commercial models hold up if devices process the workload independently?

What Changes When The Phone Does The Thinking?

Cost is where the change hits hardest.

Any product capability requiring a cloud round trip per interaction racks up expenses that increase with every new user. This proves particularly punishing for high-volume tools like smart typing assistants, meeting transcribers, automated summaries and recommendation engines. When local hardware takes over processing duties, the device absorbs the computational load and cuts the company’s per-request expenses.

Performance and speed follow close behind. Operating locally means ditching the delays of cloud round trips too, so responses can arrive faster and some features could keep working offline. For software used underground on the tube, in lifts or in remote black spots, that reliability is a game changer.

Phones are also the natural home for AI agents because they hold personal context, app data and sensor information. Amon predicts a “battle of the agents”, with assistants acting on a user’s behalf across apps and services instead of sitting in a single chatbot window.

For developers, that suggests a future where users perhaps meet products through someone else’s agent first.

Is On-Device Really The Private Option?

The privacy case is straightforward. If personal context stays on the phone instead of being sent to a server, there’s less data in transit and less sitting in someone else’s database. It’s one of the main reasons the industry is pushing processing onto the device.

The fine print is that an always-on agent that watches context and acts across apps needs broad permissions, and it has the credentials to use them. Security researchers are already flagging the rise of non-human identity theft, and agents running on millions of handsets would add a lot of new identities to protect.

Keeping data local helps, but it doesn’t necessarily remove the need to decide what an agent is allowed to see and do.

Can A Phone Really Hold A 100-Billion-Parameter Model?

Not right now. Reports suggest the most advanced mobile silicon available today runs mixture-of-experts models topping 30-billion parameters locally, which puts the 2028 ambition at more than three times that. While these models only fire up a fraction of their total parameters for any given task to ease the computational load compared to dense models of equivalent size, memory availability remains a stubborn friction point.

As a rough guide, a 100-billion-parameter model needs about 50GB of memory at 4-bit precision, or about 25GB at 2-bit. That’s far more than most phones carry today. Continuous background inference would also drain the battery and generate heat, and memory supply is already tight, with data-centre AI demand competing for the same chips.

Hardware timelines have a habit of slipping, too, as Apple’s delayed foldable iPhone shows. Getting there may depend on sparse model design, heavier compression, streaming from flash storage or a hybrid of local and cloud processing rather than simply squeezing a giant model into RAM.

The Cloud Isn’t Going Anywhere Yet

On-device and cloud AI are shaping up to complement one another as opposed to competing directly. Heavy-duty reasoning and large-scale data processing will still be anchored in data centres for some time. Mobile devices feature vastly different processing power and cooling limits. This means that developers can’t rely on uniform hardware and will need seamless cloud fallbacks for lower-spec handsets.

The features likeliest to move onto the device first are the ones called most often that need the least heavy reasoning, such as autocomplete, transcription and summarisation. They’re also where the cost savings for developers would show up first.

The Battle Of The Agents Starts In Your Pocket

If AI companies get the phones they’re asking for, the next platform fight may have less to do with apps than with which agent exists between the user and everything else.

Apps have spent years competing for a place on the home screen, and an agent that opens them on a user’s behalf changes who controls that attention. Whoever builds the agent users trust to open their apps could end up holding the position the home screen has held until now.