| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| tags: |
| - ai |
| - ai-economics |
| - finance |
| - reproducible-research |
| - primary-source |
| pretty_name: AI Cost Watch |
| --- |
| |
| # AI Cost Watch |
|
|
| Reproducible, primary-source analysis of the AI industry: whether the buildout's unit economics are holding. A recurring, dated note, where each issue states a falsifiable test for its forward call. Every figure is published with its data and a script that regenerates it, so any number can be checked at source. |
|
|
| - Author: NM AI Research (independent analyst) |
| - ORCID: 0009-0003-4213-7769 |
| - Concept DOI: https://doi.org/10.5281/zenodo.20541643 |
| - Interactive tool: https://nmairesearch.github.io/cost-watch/ |
| - Source and code: https://github.com/NMAIResearch/cost-watch |
|
|
| ## What this is |
|
|
| The dataset behind the AI Cost Watch series. `costwatch.json` holds the frozen data: the series metadata, each issue with its status and developments, and the tracked indicators with their per-issue readings. `build.py` regenerates the interactive front-end from that JSON using only the Python standard library, so the published output cannot carry an unchecked number. |
|
|
| ## The tracked signal |
|
|
| The series watches one indicator: a down-revision in the Big Four hyperscalers' forward capital-expenditure guidance. That is the trigger it is built to catch. It has not fired in any issue to date. |
|
|
| ## Files |
|
|
| - `costwatch.json`: the frozen dataset (series metadata, issues, indicators, readings). |
| - `build.py`: standard-library reproducer that reads the JSON and writes the front-end. |
| - `LICENSE`: Creative Commons Attribution 4.0 International. |
|
|
| ## Method |
|
|
| Separate an announced figure from the delivered one, tag each source by incentive, keep human judgement over the model's output, and set a falsifiable test for every forward call. Drafting is AI-assisted; the judgement is not. |
|
|
| ## Citation |
|
|
| NM AI Research. AI Cost Watch. Zenodo. https://doi.org/10.5281/zenodo.20541643 . Licensed CC BY 4.0. |
|
|
