Phương pháp luận

What this site is. AI Infrastructure Database is a research database of the AI industrial ecosystem: structured records for chips, companies, systems, data centers, power projects, technologies, software, models, products, concepts, people, benchmarks, papers and datasets, joined by typed bidirectional relationships — presented for education and research, never advice.

How records are compiled. Structured records are compiled from vendor documentation (datasheets, whitepapers, product pages), official filings and announcements, standards bodies (JEDEC, PCI-SIG, consortium specifications), peer-reviewed papers and technical reports, and reputable public reporting. Each record lists its principal sources and carries an as-of date — the date its facts were last reviewed. Records reflect public information as of that date; they are curated snapshots, not live feeds. Fast-moving areas are re-reviewed on a rolling basis: the AI-model records and current-generation chip statuses were last re-verified against public reporting on August 31, 2026.

The basis label. Every material figure is labeled with where it stands on the evidence ladder: Vendor spec (a published specification) · Vendor claim (a marketing or performance claim) · MLPerf (audited benchmark submission) · Published paper · Official (government, operator or standards figure) · Reported (credible press or filings) · Estimate (analyst or community estimate). Compute figures are dense unless noted — vendor headlines often quote 2× "with sparsity," and our records say so.

Missing values are honest. "Not publicly disclosed" means exactly that. Closed-model parameter counts, unannounced prices and unpublished capacities are never invented, inferred silently, or filled with placeholders.

Forward-looking content. Announced products carry "Announced" status and their figures are labeled as targets. Predictions and scenarios (the Future page, AGI timelines) are always attributed to the named person or organization that made them, with dates — this site makes no predictions of its own.

Market data. Quotes for public companies come from Financial Modeling Prep and are delayed (typically 15–20+ minutes) and indicative; refresh times are shown. Weekly/YTD columns are computed from daily closes. Some non-US listings return no data from our provider and show "—" rather than a substitute.

News. Headlines are aggregated from named mainstream and specialist outlets over their public RSS feeds; every item names its publisher and links to the original article. Short data briefs are written in-house with AI assistance from the figures already in this database and are labeled accordingly — see the Editorial Policy.

AI assistance. Portions of this site's content and data compilation are produced with AI-assisted research; figures are indicative and may be out of date — always verify with primary sources.

Refresh cadence. News and quotes refresh hourly; weekly/YTD market columns daily; data briefs on a daily schedule; structured records are re-reviewed and re-seeded on an editorial cadence (each record shows its as-of date). Corrections ship with the next deploy — use Contact.