# FAIR data tax costs Europe €10.2bn a year _VAST Data demo writes FAIR metadata at ingest. Europe loses €10.2bn a year to non-FAIR data._ **Published:** 2026-11-16 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/fair-data-tax-costs-europe-10-2bn-a-year --- €10.2bn a year is the minimum cost to Europe of research data that is not FAIR, per PwC for the European Commission in 2018. [VAST Data](https://www.vastdata.com/blog/fair-data-platform-curating-research-data-where-it-lands) says its fix is a demonstration pipeline that curates data where it lands, described by engineers Chris Snow and James McClung. It does not ask researchers to file another form. A researcher or HPC job writes a folder over S3 or NFSv4 and VAST DataEngine triggers fire on each file, the company said. Events go to Kafka on the VAST Event Broker, files get profiled for format, headers, variables and checksums, and the folder gets a draft record, a FAIR score and a review page. The design matches its broader pitch for DataEngine as a runtime that runs functions, triggers, streams and filters inside the platform. The loss numbers are stark. The odds a dataset behind 516 papers still existed fell 17% for each year since publication, according to Vines et al. in 2014. Of 1,792 authors who said data was available on request, about 7% sent it, according to Gabelica et al. in 2022. In the 2025 State of Open Data survey, 69% said they get too little credit for sharing. On six public datasets, the draft got 62 of 64 fields right before human edits, across two runs on different clusters, the company said. The sets spanned NOAA sea surface temperature, a TESS light curve, CMS open data, gnomAD, BBBC039 images and UK traffic counts, plus four renamed dark copies. In all seven cases with no stated licence or creators, it left the fields blank. One renamed physics file drew a wrong guess at 0.7 confidence, which review is meant to catch. Scores rose after review. Documented sets went from 24 to 52 out of 100 to 72 to 92 on the demo checks, the company said. An export_4.csv with no documentation went from 8 to 76 once a person added licence and creator. Independent assessor F-UJI 3.5.0 scored published sets 83 to 88, capped because test identifiers are not registered with DataCite. The pipeline is about 3,500 lines of Python, almost all about metadata, with snapshots, Croissant files for ML use, vector search and OAI-PMH feeds, the company said. Search by meaning put the NOAA set first at 0.63 against 0.39 for the next on ocean temperature from satellites. What the test does not show is production behavior on billions of files with real DOIs, and Trino over published tables was not deployed in this demo. Live DOIs and a full archive backlog would change the read. Who pays for curation that mostly benefits the next lab? --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.