UPISTEPS
Trending
•Small language models cut inference bills by 90%•Passkeys hit mainstream adoption across major browsers•Edge runtimes reshape how teams deploy APIs•GPU spot pricing falls for the third quarter running•Rust creeps into frontend build toolchains•Post-quantum TLS lands in production stacks•On-device AI pushes battery engineering forward

Hardware

On-Device AI and the Battery Problem

Ravi Menon
5 min readLast updated Jul 11, 2026
On-Device AI and the Battery Problem

Every flagship phone now advertises an on-device model. Far fewer of them run it continuously, because sustained inference is a thermal problem long before it is a compute problem.

Peak numbers mislead

Vendor benchmarks quote burst throughput. Sustained throughput after ten minutes is frequently half that, once the device throttles. Design your feature around the sustained figure.

Featured

Batch, cache, and defer

Do the work while charging

Indexing photos, generating embeddings and summarising documents can all wait for a charger and a cool chassis. Users never notice the delay; they notice a hot phone immediately.

#hardware#mobile#ai#npu#battery
Featured

Related reading

Small Language Models Are Quietly Winning
Artificial Intelligence

Small Language Models Are Quietly Winning

Compact models now handle most day-to-day tasks at a fraction of the cost. Here is why teams are downsizing their AI stack.

6 min read