Observatory status
Amsterdam time: Last data received: - arXiv records ingested: 0 GitHub records ingested: 0 Wikipedia records ingested: 0 Zenodo records ingested: 0 Hacker News ingestion: stopped on 2026-08-08 Reddit ingestion: stopped on 2026-08-08 Stack Overflow ingestion: stopped on 2026-08-08 Zenodo ingestion: started on 2026-08-11 Primary literature source: Zenodo, since 2026-08-15 Observatory active since 2026-02-15: 0 days Archive with updated backlog since: 2025-01-01
RAG service — beta
LLM-assisted search and analysis across Zenodo, arXiv, GitHub, ComputeCosts reports and selected Wikipedia statistics. Coverage and analytical capabilities are being expanded.
1,000 COSTS provides 24-hour access (currently on the order of €0.01).
Beta service — availability is currently provided on a best-effort basis.
Observatory reports
ComputeCosts publishes a series of technical observatory reports describing the measurement framework, archive construction and empirical signals extracted from the observatory datasets.
All ComputeCosts reports are published and archived on Zenodo and assigned DOI identifiers, providing an independent, persistent and openly accessible publication record.
Report 005
Zenodo acquisition and document-level infrastructure evidence.
Zenodo DOI
Report 004
Comparison of cloud computing mentions and RTX workstation disclosures in arXiv full text papers.
Zenodo DOI
Report 003
Workstation scale computing in contemporary scientific research.
Zenodo DOI
Report 002
arXiv archive, epoch 1.
Zenodo DOI
Report 001
Observatory framework and first measurement baseline.
Zenodo DOI
What this is
ComputeCosts.nl is an independent observatory that measures the real cost of compute usage, including not only financial expenditure, but also user time loss, operational friction and loss of focus.
Evidence is evaluated by source, provenance, reproducibility and direct relevance. Journal prestige or peer-review status is not treated as a substitute for evidence quality.
The initiative originated from repeated observation that in many practical and scientific workloads, local compute solutions are often faster to start, cheaper to operate, involve less dependency friction and provide substantially more privacy than cloud-based alternatives.
While cloud computing clearly offers advantages at sufficient scale, this threshold appears to be reached less frequently than commonly assumed. ComputeCosts therefore builds a forward-looking measurement framework to map these trade-offs using current, verifiable signals rather than retrospective commentary.
Agents as sensors
The Observatory currently operates four active ingestion agents, monitoring Zenodo, arXiv, GitHub and Wikipedia. Each source functions as a distinct observational sensor and is interpreted according to its own characteristics rather than treated as equivalent evidence.
Since 2026-08-15, Zenodo is the Observatory's primary literature source. Its open repository model, persistent identifiers, broad research coverage and reproducible API access make it particularly suitable for independent literature research. ComputeCosts also uses Zenodo to publish and archive its own Observatory reports.
Zenodo includes peer-reviewed publications, preprints, reports, datasets, software and other research outputs. ComputeCosts treats peer review as useful context, not as a gatekeeper. Evidence is assessed on provenance, relevance, reproducibility and what the source actually demonstrates.
arXiv remains a supplementary sensor for preprint and early-publication activity, GitHub for implementation-level and operational friction, and Wikipedia for attention toward computational infrastructure concepts. No single journal, repository or review system defines the Observatory's evidence base.
Hacker News, Stack Overflow and Reddit were used during the initial development phase as exploratory cross-reference sources. Their exploratory role was considered complete, and ingestion was discontinued on 2026-08-08. The associated locally retained source data was deleted. They are not part of the current Observatory analysis, retrieval system or language-model context.
Each active archive is collected in a controlled, repeatable manner, deduplicated into a single canonical record per item, systematically feature-scored and cryptographically anchored, creating an auditable evidence stream for analysis.
Balanced observation scope
The observatory measures concrete, documentable signals for both cloud-based and local compute environments. For cloud systems, this includes references to GPU availability, quota limits, billing discussions, egress fees, vendor lock-in and reported instability in primary sources.
For local systems, the active Observatory sources capture explicit mentions of self-hosted deployment, local GPU usage, high-end workstation hardware such as RTX-class cards and large memory configurations, and offline inference feasibility.
Current Observatory signals are derived from Zenodo, arXiv, GitHub and Wikipedia and processed into deterministic feature matrices, enabling comparisons between cloud and local infrastructure based on measurable frequencies, retrieved document evidence and time-series trends.
Archive architecture and verifiable logs
The Observatory currently operates four active ingestion archives: Zenodo as the primary source for scientific literature and broader research outputs, arXiv as a supplementary source for preprint and early-publication signals, GitHub for implementation-level and operational evidence, and Wikipedia for attention toward computational infrastructure concepts.
Each active archive is harvested incrementally and transformed into an append-only dataset with reproducible feature extraction and time-series aggregation.
Each active archive maintains its own dedicated NFT on the Algorand blockchain. Dataset state hashes are committed as zero-value transactions, creating a public, tamper-evident log of ingestion baselines, recorded dataset states and derived analytical outputs.
This structure separates signal acquisition, canonicalization, feature construction and blockchain anchoring into explicit stages, allowing recorded dataset states and analytical outputs to be independently timestamped and verified.
About the COSTS token
Each dataset hash is committed twice on Algorand: once to its archive NFT and once as a one micro COSTS transaction to the central archive address. This creates two independent indexing paths, each optimized for a different retrieval scope.
COSTS is the utility token used to access the ComputeCosts Observatory search service. A COSTS payment provides 24-hour fair-use access to LLM-assisted search and analysis across the active Zenodo, arXiv, GitHub and Wikipedia archives.
The service is designed for analysis and synthesis, not redistribution of source material. It does not provide access to or copies of the underlying datasets or documents. Answers are generated from relevant retrieved evidence, with source links or identifiers provided where available. Complete documents, sections, issues, comments or substantial source passages are not reproduced.