Methodology

1. How sources are selected

We prefer, in order: primary sources (company filings, investor relations, official press releases, government and energy-agency data), then established trade and technical publications. Automated collection uses only public RSS/Atom feeds and public APIs intended for syndication. We do not bypass paywalls or authentication, circumvent anti-bot protections, ignore robots.txt, or scrape LinkedIn. Every source is registered once in a source registry with its publisher, URL, type, publication date and collection date — see Sources.

2. How records are normalised

All data is converted into shared schemas — Company, InfrastructureProject, GPUDeployment, DataCentre, PowerProject, FundingEvent, NewsEvent, Observation, MarketSignal and Prediction — validated on every build. Quantitative values are never stored as bare numbers: each carries its unit, an “as of” date, a verification status, a confidence level and at least one source. A missing value means not yet sourced, never zero.

3. How duplicates are handled

News items are identified by a hash of their canonical URL (tracking parameters removed). Items whose titles are near-identical (token overlap ≥ 60%) and published within three days are treated as the same event: the earliest report is kept and later ones are recorded as corroborating sources, which raises the item’s importance score.

4. How company and project information is reconciled

Each company and project has a single stable identifier. The corporate site’s baseline records and the NeoCloud module’s detailed records are merged by identifier, so enrichment never creates a duplicate company. News is associated with companies by name and alias matching with word boundaries; ambiguous names use explicit match terms. Conflicting values are resolved in favour of the more primary, more recent source, and the choice is recorded in the record’s notes.

NeoCloud Observatory research

Company, project, financing, contract and capacity records for NeoCloud providers come from the NeoCloud Observatory research snapshot, where every value carries its source URL, source type, publication date and a value class (reported, derived, estimated, inferred). HYBSIS keeps those distinctions: capacity is typed by maturity (operational, under construction, contracted, secured, announced, potential, speculative) and by MW type (IT load, facility, grid connection, power pipeline). Headline totals use IT-load figures only and never add different types or maturities together; announced GPUs never count as operational; contract values state their basis (total contract value, ceiling, annualised or programme). Source confidence follows source type: company disclosures and filings high, industry and financial journalism medium, secondary aggregators low.

5. How estimates are identified

Every figure is labelled Reported (stated by the company or a filing), Announced (a plan or target), Operational (stated as delivered), Estimated (derived by HYBSIS or a third party) or Unverified (collected but not yet checked against its source). Announced and operational capacity are always kept separate.

6. How signals are generated

Each market signal is a published rule: a set of topic categories and keywords that make an item count as evidence, plus curated observations and structured events (financings, acquisitions, project announcements) from the shared dataset. The engine compares evidence in the most recent 90 days with the 90 days before:

• Direction — rising if recent evidence is at least 1.5× the prior window and at least 2 items higher; falling for the reverse; otherwise stable. Fewer than 3 items across both windows is reported as insufficient data.
• Strength — by recent volume: strong (≥10), moderate (≥5), weak (≥1).
• Confidence — by source diversity: high (≥5 distinct sources and ≥8 items), medium (≥3 and ≥4), otherwise low. Evidence consisting only of analyst baseline observations is always low confidence.

Signals describe what is observable. They do not claim causation, and they are not forecasts. Every signal card shows its arithmetic and its latest evidence.

7. How confidence is assessed

Record-level confidence reflects the quality of the source (primary vs secondary), its recency and whether independent sources agree. Signal confidence is computed as above. Prediction confidence is a stated HYBSIS judgement (low / medium / high) and is shown on every prediction.

8. How predictions work

Each prediction separates four things: Fact (sourced), Observed signal (computed), Interpretation (our reasoning) and Prediction (a forward-looking claim with a horizon and a review date). Predictions are stored in immutable, dated snapshot files. A later snapshot may change a prediction’s status — Open, Tracking, Supported, Partially Supported, Not Supported, Superseded — and record an assessment, but it may never delete a prediction or change what was predicted. To revise a claim we supersede it and publish a new one. This lets anyone see what we predicted, what happened, and how accurate we were.

9. How historical snapshots work

Important datasets are preserved as dated, normalised snapshots (for example, the NeoCloud dataset). A snapshot is written only when content has changed, and existing snapshots are never overwritten. Committed history files are protected by a SHA-256 manifest checked in automated tests. The daily pipeline also keeps a monthly news archive and daily signal snapshots, so trend analysis survives later changes to public sources.

10. Machine assistance

The platform runs without any AI service. If an optional language-model adapter is enabled, it may only draft short summaries for items that lack one; those summaries are labelled as machine-generated. Language-model output is never used as a source of numbers, entities, classifications or signals.

11. Corrections

If you believe a record is wrong or out of date, please contact us with a source. Corrections are applied to the current record with a note; historical snapshots and predictions are not rewritten, so the correction itself remains visible over time.

Current coverage

The dataset currently holds 38 companies, 38 projects, 50 financing/M&A events, 63 curated observations, 800 collected news items and 236 registered sources. The full dataset is published at /data/intelligence.json.

HYBSIS Intelligence is independent research based solely on publicly available information. It does not contain or rely upon confidential, proprietary or non-public information obtained through employment, clients, suppliers or other professional engagements. References to companies, products and projects do not imply affiliation, endorsement or a commercial relationship with HYBSIS.