Skip to lesson
Exit
AI-Editor Systems & Design1 / 3

1 min lesson

The latency levers

Connect "The budget proves the design" to the decision you would make.

Step 1 of 3

The latency levers

The budget proves the design: a big model over a far network can't fit, so the levers below are forced, not optional.

Learn more

Advanced table

Lever

Lever
Debounce + cancel in-flight
What it buys
Kills most requests before they're sent - the user is still typing
Cost / catch
Tune the window: too long feels laggy, too short wastes calls
Lever
Small, specialized model
What it buys
The only way inference fits the budget at this volume
Cost / catch
Lower ceiling on hard completions - route those elsewhere
Lever
Speculative / early decoding
What it buys
Start emitting before the full sequence is decoded
Cost / catch
Wasted compute on speculations that don't pan out
Lever
Caching
What it buys
Repeat prefixes and contexts skip inference entirely
Cost / catch
Cache keys must include enough context to stay correct
Lever
Prefetch on pause
What it buys
Compute likely completions during a typing pause, serve instantly
Cost / catch
Speculative spend; bound it so it doesn't dominate cost
Lever
Edge proximity
What it buys
Cuts the fixed network tax that no model trick can recover
Cost / catch
Operational complexity of regional inference