Does running AI on a phone drain the battery or overheat it?

Where the energy actually goes

Generating text moves model weights through memory as well as performing arithmetic. Moving data can use substantially more energy than the arithmetic itself in some workloads, so energy use depends on how much data moves, the memory hierarchy and the device design.

A smaller or quantized model may reduce the amount of weight data moved during inference, which can reduce memory traffic and energy use in some workloads. Actual speed, energy and heat depend on the model, format, runtime, device and task; they should be measured for the intended use.

A burst and a sustained load are not the same workload

A phone has no fan, so sustained work can warm its case and trigger thermal limits. The device, workload, ambient conditions, and design all affect how quickly that happens; there is no single surface-temperature limit or burst duration that applies to every phone.

Past that point the phone protects itself. Clocks come down, tokens per second falls, and the case stays warm until the load stops. This is the whole difference between an assistant and a benchmark. A spoken request is a few seconds of work followed by minutes of idle, and the accumulated heat drains away between turns. A script that generates continuously never gives it that chance.

The two halves of a generation behave differently too. Reading a long prompt is compute-bound and can saturate the matrix engines in a short, sharp spike. Writing a long answer is memory-bound and lower in peak power but relentless, one full pass over the weights per token. A long input is a heat problem. A long output is a battery problem.

What it costs a battery, in round numbers

As a hypothetical illustration, a 5-watt load for 10 seconds uses 50 joules, or about 0.014 watt-hours. That arithmetic does not predict a phone’s actual power draw or battery impact; both depend on the device, model, workload, screen, and radio.

Longer sustained use can consume more energy and produce more heat than brief use. A phone may reduce performance or limit charging as it warms, but the timing and effect vary by device and conditions.

• Energy per token is set mainly by the size of the weights read, so quantization moves it directly.
• Interactive use is duty-cycled, so the average draw across an hour is far below the peak during a turn.
• Sustained generation is limited by heat first and charge second: the phone slows down before the battery runs out.

The screen and the radio are on the same bill

The display can also contribute to energy use, especially at high brightness. If you are looking at the phone while it answers, some of the energy goes to the screen as well as computation.

Sending the work to a server is not free either. Transmitting costs power, the modem climbs to a high-power state to do it, and it stays there for seconds after the last packet. On a weak signal the phone pushes harder to be heard. For a short request, a network round trip can cost as much energy as computing the answer where you are standing.

Does it damage the phone?

Sustained heat can accelerate lithium-ion battery aging, especially while charging. Phones may reduce performance or limit charging when they get hot. These protections reduce risk, but they do not guarantee that heat cannot affect the device or battery.

The worst case is generating continuously while fast charging, because that stacks two heat sources in the same aluminum. Short of that, the load looks unremarkable next to a 3D game or an hour of turn-by-turn navigation with the screen lit, both of which hold a phone at high power far longer than an assistant turn ever does.