Sources & method

What this site is

A parts-first catalog of the AI data center supply chain. The unit of the site is the physical component — a cold plate, an ABF substrate, a 765 kV transformer — and companies hang off parts through a single many-to-many table. Every page you see is a query over that table.

Where the data comes from

The seed is a hand-drafted taxonomy (v0.1, drafted 2026-08-27) of 11 layers, 130 part categories and roughly 1,000 companies, compiled from public knowledge: OCP specifications and contributor lists, vendor product catalogs, exchange filings, earnings transcripts, and trade press. Placements are then verified over time against primary sources. Nightly jobs add price snapshots, news, Taiwan monthly-revenue filings and SEC customer-concentration disclosures.

The confidence system

• disclosed — the company or a regulatory filing says so. A source URL is attached to the row.
• reported — from trade press or a reference-design partner list, not a company disclosure.
• inferred — a placement from the taxonomy seed without a citation yet. Treat it as a lead.

Tags are never upgraded without a source, and the site renders them with those words. Share estimates, quantities and lead times appear only when a source URL sits in the row; unknown stays blank.

Prices

Prices, day changes and market caps come from a nightly snapshot (currently via Yahoo Finance data), shown in the listing currency with the snapshot timestamp. Failed fetches keep the previous value and are visibly marked stale. Nothing here is real-time, investment advice, or an endorsement; this site has no affiliation with any listed company.

How AI and automation are used here

The site is published by its operator under the name GridToChip; there are no individual authors, bylines or personas. You can reach the operator at chiptogrid@simbasystems.com.

Generated by code from the database, with no language model at publish time.

• The "where it sits in the stack" block, the lead paragraphs and FAQs on company pages, and every comparison page: sentences are templates filled from part and company rows, part summaries and computed overlaps.
• The capex tracker: figures are read from SEC filings by a script, each linked to the filing; quarters marked derived are computed by subtraction.
• The Taiwan revenue table (MOPS filings), the filing-language counts (fixed phrase lists counted in SEC filings), the stress test and the weekly data changelog (a diff of the data files). Each is arithmetic over filed or catalogued data.
• The daily part-of-the-day post quotes the part's own catalog fields and points to the part page as its canonical address.

Where AI is used.

• The site's code and much of its hand-written prose, including the explainers, the glossary, the supercomputer and rack breakdowns and the analysis posts, were drafted with AI assistance (Anthropic's Claude) and published by the operator. Sources for the facts in the breakdowns are shown on each line.
• Nightly news headlines are matched to parts and companies by a language model running on the operator's own hardware, with a keyword fallback when it is unavailable. Each news row records which classifier tagged it.
• Customer-concentration rows combine pattern matching on SEC filings with a model check; every row keeps the filing URL and a short excerpt.

What AI does not do here. It does not invent sources, prices, quantities, authors or credentials: a number appears only with a source URL, and unknown stays blank. There are no AI-generated images; charts, the logo and the share image are drawn by code.

Checking, and its limits. Placements carry a confidence tag, the data files are validated for broken references every night, and scripts audit page metadata and internal links before a release. No outside expert (financial, legal or engineering) reviews this site; it is informational and not advice. If something is wrong, tell us and we will fix it.

Measured accuracy

On 2026-10-22 we drew a seeded random sample of 60 of the 486 placements that carry a source link and checked each source: does the page load, does it name the company, and does it mention the part.54 passed that automated check, 5 that did not were read by hand and held up, 1 could not be read (the source host returned an error), and we found 0 that the source contradicts or fails to support. Every sampled row, its automated result and our review note is in accuracy_sample.csv.

Read that carefully. It shows the sources are live and about the right company and topic, not that every role and share is exactly right. 17 of the 60 sampled sources (28%) are a company homepage rather than a product page, which supports the company but only loosely the part; across all sourced placements the figure is 122 of 486 (25%). The ~1,360 inferred placements carry no source and are not covered by this audit, so their error rate is unmeasured; treat them as leads.

Taxonomy version

Current: v0.1. The taxonomy file is read-only; all corrections flow through an overrides layer so re-imports never silently undo a fix.

Report an error

Corrections are welcome — email chiptogrid@simbasystems.com with a public source (see contact). A public repository for direct edits is also being set up.

Machine-readable access

Every part and company page has a .md twin (append .md to the URL) and a JSON mirror under /api/. The whole catalog is downloadable as SQLite, and llms.txt describes the stable IDs.