Jev moves 80% of agent calls off frontier models to a decision layer
The author says 1.7 million of his 2.1 million daily calls are decisions rather than generation, and routing them through Jev cut that part of the bill from $11,400 to $63.
JENSEN HUANG IS WORRIED ABOUT THE WRONG THING everyone thinks the threat to nvidia is a cheaper chip out of china. it isn't. it's a model that refuses to write a single word sam altman is building stargate. elon is stacking gpus inside colossus. dario keeps signing compute deals bigger than most countries' budgets. all of it rests on one assumption: every ai call needs a giant model generating tokens jev just broke that assumption it's the model the guy who invented rlhf built after he walked out of openai. it doesn't generate, it decides. classify, route, score, approve, block. 70 milliseconds, output tokens free, nothing to hallucinate and here's the part wall street hasn't priced in yet: → pull the logs from any company running agents and 70-80% of the calls are decisions, not writing → every one of those calls currently runs on a frontier model, on nvidia silicon, billed by the token → jev makes the same call on roughly 1/200th of the compute → move that 80% over and most of the gpu demand behind it quietly evaporates the data centers don't disappear. they just turn out to be massively oversized for what the work actually needed i moved my own stack last month. 2.1M calls a day, 1.7M of them were decisions. my bill on those dropped from $11,400 to $63 multiply that by every company running agents and you get the most expensive misunderstanding in tech: we built trillion-dollar infrastructure to answer yes or no the labs will keep the writing. jev is quietly taking everything else bookmark this. in 12 months it's either the dumbest take you read this year or the one you wish you'd acted on
This post quotes another post: x.com/leopardracer/status/2101982527871131955