The compute bill of AI papers (research idea)

The compute bill of AI papers

Inference latency and resource-constrained deployment are among the fastest-growing keywords in the AI journals, yet only 15 of 226 such papers carry an energy, carbon or green-AI keyword. How much computing do the methods in AI journals cost to train and run, and would reporting it change which methods win?

What prompted this

In the AI journals, "inference latency" had 12 papers in July–September 2026 against about 5 expected and "resource-constrained deployment" 10 against 4 (see the AI dashboard). Since October 2025, 226 papers carry an inference-latency, resource-constrained or edge-deployment keyword (199 in AI journals). Only 15 of them carry an energy, carbon, green-AI or power-consumption keyword, and "green AI" itself appears on 6.

Questions

What share of recent AI-journal papers report training compute, inference time or energy, and do they report them for the baselines too?
If results were compared per unit of computing, how often would the winner change?
Would a short reporting checklist (hardware, hours, energy) be adopted, as code release was?

Possible approach

Code a sample of recent papers for what they report about computing, and re-rank the reported results by accuracy per GPU-hour where both are given.

Next step

Code 100 recent papers from the AI journals for whether they report hardware, training time, inference time and energy use.

Related research activity

The keywords this note follows, in the tracked finance, accounting, economics, management, operations research and AI journals as of 2026-10-26
Keyword	Papers, last 30 days	Rank	Papers, last 36 months
inference latency	4	#525	43
resource-constrained deployment	2	#1303	51
green AI	1	#2881	14

@misc{scholar4ai-ideas-the-compute-bill-of-ai-papers, author = {{Scholar4AI}}, title = {The compute bill of AI papers}, year = {2026}, howpublished = {\url{https://scholar4ai.com/ideas/the-compute-bill-of-ai-papers}}, urldate = {2026-10-23}, note = {Journal Research Rankings; data from OpenAlex (CC0); licensed CC BY-NC 4.0} }