ComputeCosts Observatory Report 002

This report documents the first structural analysis of the arXiv archive within the ComputeCosts
observatory. The archive records metadata and abstracts of recently published scientific papers
and treats this collection as a measurable signal of how computational infrastructure is described in
scientific writing.

The dataset analysed in this report contains 6690 unique papers obtained through a deterministic
ingestion pipeline. The epoch 1 observation period spans from 1 January 2025 00:00 UTC until 19
February 2026, corresponding to the ingestion state frozen at the time of this report. The analysis
deliberately collapses the entire dataset into a single baseline snapshot rather than attempting to
interpret short term variation. The objective is not to measure trends but to determine whether the
archive contains measurable signals relating to cloud infrastructure, local computation and
operational constraints.

Initial measurements show that abstracts already contain operationally relevant signals, even
though detailed deployment narratives are not typically expected in scientific papers. Mentions of
self hosted and local execution vocabulary indicate that local compute set ups are explicitly used in
a measurable subset of the dataset. Hardware references to consumer GPUs, specifically NVIDIA
RTX class devices including RTX 4090, demonstrate that workstation grade hardware is used for
scientific computing in published research.

These observations are particularly valuable because the archive is accompanied by a complete
local PDF collection for the full dataset. The present report does not analyse those PDFs, but their
availability means that the abstract level signals identified here can be deepened later through full
text extraction, context reconstruction and evidence snippets. The abstract results therefore
function as a calibration step that justifies a second phase of deeper paper level analysis.