You Pay for Training Once and for Answering Forever

1. SAME MODEL, DIFFERENT WORK — Training: forward pass, loss, backprop, gradients, weight update; stores batch activations and repeats across epochs. Inference: forward pass only; no gradients, history, or updates; frozen weights can be quantised to 8-bit. | 2. TWO COST CURVES — Training is a large, schedulable one-time event; LoRA may add capability without full retraining. Inference spends GPU time per request forever; monthly cost = request count × per-request GPU cost. Warm capacity also costs while idle. | 3. INFERENCE BECOMES OPERATIONS — Every endpoint inherits latency budgets, availability targets, scaling, monitoring, and on-call duty. Better training improves expected quality, but offline accuracy estimates a sampled distribution; it cannot guarantee the next answer. | 4. FALSIFY THE INTUITION — Do not ask only which run is expensive. Compute 12 × actual monthly requests × measured GPU cost per request, then compare it with the training bill. Mark the crossover month: lifetime inference can become the larger line item far earlier than expected.