The rising costs of training frontier AI models

A preprint on the rising costs of training frontier AI models. Both papers cite it for one fact: energy is a few percent of the amortized cost of training a frontier model. Swarm ecology uses this to argue that energy does not bind a lineage; capital and legal standing do, so identity checks and providers of priced services are the chokepoints.

Cited in

both papers (Flagship: The threshold; Companion §3)

Used for

P2P2; R_m (identity checks); trophic independence

Strand

Compute, energy & the physics of information · Energy, data centres & compute supply

Cite

@misc{cottier2024rising,
  title = {The rising costs of training frontier AI models},
  author = {Ben Cottier and Robi Rahman and Loredana Fattorini and Nestor Maslej and Tamay Besiroglu and David Owen},
  year = {2024},
  url = {https://arxiv.org/abs/2405.21015},
}