curio

news

Why One AI Answer Can Cost More Than Another

Every token an AI writes pushes data through physical chips, so a 500-word answer burns computation a 5-word one never touches. A bigger model moves more data for each token, too.

Providers batch many requests to keep the hardware busy, but a batch takes time to fill — cheaper compute, slower answer.

The bill never follows the model's name — it follows the work: how many tokens, how big the model, how crowded the servers.

Source: arxiv.org

Get Curio — more like this

Short reads that go somewhere. No feed to scroll forever.