Community Discussion · Policy

Separating training and inference is the most expensive waste in the RL era

HuangCFOHuangCFOAug 182026/08/18 271 views

When finance people read tech news, our reflex is to look for cost items. Let me start with an analogy: A company has two budgets, one for R&D and one for marketing. Policy dictates they cannot be cross-used, yet by year-end, both departments are scrambling to spend everything, because if you don't use it up, your budget gets cut next year. Today, most AI companies deploying training and inference compute operate on this exact siloed accounting logic. Training clusters are separate from inference clusters—they buy separately, schedule production separately, and calculate utilization rates separately.

2 replies

?
Ctrl + Enter to reply
Truth Seeker

Wait, dynamic scheduling sounds great, but with such huge fluctuations in inference during the RL phase, won't training tasks get squeezed out? Do you have any actual data from real runs? I'm worried the paper calculations look pretty but reality is different.

Ming Ming Bu Gui Fan

LOL, our internal training and inference were on two separate ledgers too. Once RL came along, everything got messed up. The scheduling team was arguing every day, so in the end we just threw it all into one pool. You know what? Turnover rate actually went up.