Where this record is least certain.

A comparable industrial water price, per country, for the countries this record's fab sites are in

The survey is a paid product and this record buys no data, ever. Residential tariffs from individual utilities are not an industrial price, and averaging them into a country figure would produce a number no publisher printed. Water is a real input cost for a fab, and the honest state of the free public record is that nobody publishes it comparably.

Known

Global Water Intelligence publishes a Global Water Tariff Survey covering hundreds of cities across most countries, sold as a report. IBNET holds utility-level tariffs with uneven coverage and is largely residential.

Would unblock

the water half of any page about what a fab site costs to run

Source

https://www.globalwaterintel.com/global-water-tariff-survey

Knometa Research's Global Wafer Capacity, more than 400 fabs with capacity, wafer starts and utilisation. The 2024 edition is the LAST: the publisher states there will be no further issues.

Paid. A figure from it could inform what to look for in public sources and could never appear on this site, so buying it would be research rather than supply. Recorded because the fact that the only INDEPENDENT capacity database has stopped publishing is itself worth knowing: SEMI's numbers are becoming the only numbers, which makes a public record of what SEMI said and when more valuable, not less.

Known

Final edition on sale at US$2,750 single user, US$4,200 enterprise, read from knometa.com on 17 August 2026.

Would unblock

a check on whether this record's free-sourced picture is right

Source

https://knometa.com/services/reports/

Korean semiconductor export figures at source

Not a licensing problem but an access one. UN Comtrade covers a narrower commodity code months later, and is carried under its own key and its own row rather than being passed off as the same number.

Known

data.go.kr publishes them; all three membership types require Korean identity verification or a Korean business registration

Would unblock

a faster Korean read on the indicators page

Source

https://www.data.go.kr/

UMC's quarterly report PDFs, which carry a per-fab capacity table in thousands of 8-inch equivalent wafers - the most granular first-party capacity disclosure any foundry makes, quarterly.

⛔ THE PUBLISHER HAS TOLD AUTOMATED CLIENTS NOT TO. robots.txt disallows /upload/ for every user-agent, and that is where every quarterly report lives. This record does not fetch what a publisher has asked automated clients not to fetch, and it does not route around the instruction by another means. A HUMAN opening the PDF in a browser is not an automated client and is not covered by it, so the figures can be read and hand-entered the way every other entry in fab-capacity.json was - the file is hand-curated by design. What cannot happen is a collector. [Re-filed 18 August 2026 from method-cannot-reach, which was the wrong shelf: the method can reach these files perfectly well and the publisher said not to. The distinction got its own status when a second case appeared, see youtube-channel-feeds.]

Known

The files exist and their URLs are stable: umc.com/upload/media/08_Investors/Financials/Quarterly_Results/... UMC's robots.txt says `User-agent: *` and `Disallow: /upload/`.

Would unblock

Per-fab quarterly capacity for a whole foundry, on a stated 8-inch-equivalent basis, with a history going back to 2000.

Source

https://www.umc.com/en/Download/quarterly_results

Uploads from nine YouTube channels this record used to collect: Asianometry, TechTechPotato, Ben Eater, Applied Science, Branch Education, ASML official, imec, der8auer and High Yield.

TWO INDEPENDENT REFUSALS, and the second one is the surprise. First, the feed path is disallowed to every automated client, and this record does not fetch what a publisher has told automated clients not to fetch. Second, the sanctioned alternative closes too: the YouTube API Services developer policies allow an API client to store non-authorized data for no longer than 30 calendar days, after which it must delete or refresh it. This archive's central invariant is that it never shrinks and holds items back to 1999, so the compliant route and the archive contradict each other directly. Neither refusal is a technical fault and no user-agent change cures either.

Known

Each was collected from https://www.youtube.com/feeds/videos.xml?channel_id=... Read 18 August 2026, youtube.com/robots.txt contains, under `User-agent: *`, the line `Disallow: /feeds/videos.xml`. Six of the nine had begun returning 404 or 500 and three still answered, which is inconsistent enforcement on YouTube's side rather than a difference between the channels.

Would unblock

nothing automatically. A human subscribed to these channels in a browser is not an automated client, so an individual video worth recording can be hand-entered the way fab-capacity.json already is. Where a creator publishes their own newsletter the record takes that instead: More Than Moore was added on 18 August 2026 for exactly this reason, and Asianometry's Substack was already held.

Source

https://www.youtube.com/robots.txt

The IEEE International Roadmap for Devices and Systems (IRDS). Its More Moore chapter publishes the Logic Core Device Technology Roadmap as a SPREADSHEET, not only as a PDF, and the tables were read successfully on 17 August 2026: 2024IRDS_MM_Tables.xlsx and the 2022 equivalent both parse, and both carry a row named "Logic industry 'Node Range' Labeling" that maps each year of production to the label the industry uses for it.

This record republishes figures. A source whose own file says permission is required for republishing is not publishable on the strength of a friendlier sentence on its website, and the two cannot both be relied on. Nothing from these tables is reproduced anywhere on this site, including in this entry: what follows describes the SHAPE of the roadmap without copying it.

Known

TWO STATEMENTS FROM THE SAME PUBLISHER, AND THEY DISAGREE. The website says: "The IRDS roadmap reports can be used freely and with permission by giving proper credit to IEEE and the IRDS. IRDS images, data, and content can be shared for the greater good of the industry and research community." The spreadsheet's own index sheet says: "© 2024 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works." Note that even the website's sentence says "freely AND WITH PERMISSION", which is not a permission.

Would unblock

A SECOND PUBLISHER for node bands, and something no surface anywhere presents: the roadmap as a revision record. Comparing the 2022 and 2024 editions of one row shows the roadmap moving in three separate ways - the step between production years shortened, one step moved a year earlier while being renamed, and the naming convention changed from nanometre-equivalent to angstrom-equivalent, so the same physical target reads as a different label in the two editions. That is the stagger-chart pattern in a place nobody looks for it, and ten editions exist.

Source

https://irds.ieee.org/editions/

UN Comtrade: trade flows by HS code, reporter and partner. HS 8542 is electronic integrated circuits, 8486 is semiconductor manufacturing machinery, 3818 is doped wafers. Probed directly on 17 August 2026 rather than read about.

The licence prohibits, by name, the two things this record would do with it: AUTOMATED DOWNLOADING and PUBLICATION. An API key would not change that - a key raises a rate limit, it does not grant a licence - so registering would only add a credential to an activity the terms forbid. Nothing from UN Comtrade is collected or published here, and the endpoint was probed by hand a handful of times to establish exactly this.

Known

TWO FACTS, AND THE SECOND OVERRIDES THE FIRST. The free endpoint works with no key: comtradeapi.un.org/public/v1/preview returned Netherlands imports of HS 8542 for 2023 and 2022 without any credential. The rate limit is severe and exposed in no header - four calls in quick succession gave one 200 and three 429s, one call after a 25-second pause succeeded. Rows carry estimation flags: both read had isReported=false and isNetWgtEstimated=true, so UN Comtrade compiled or estimated them rather than the reporter filing them. The UN Comtrade License Agreement then says, of the material: "Except as may otherwise be expressly permitted by applicable law, any copying, automated browsing or downloading, redistribution, publication, or commercial exploitation of any material contained on or otherwise made available to you on United Nations COMTRADE is strictly prohibited without the prior written permission of the United Nations."

Would unblock

Trade flow as a public series, which is the only evidence that would show the industry's physical movement between countries. Worth one written request to comtrade@un.org, on the same pattern as the IRDS letter.

Source

https://comtradeapi.un.org/public/v1/preview/C/A/HS

Archive items about Arm Holdings, indexed on the company page

Company pages list an item when a headline contains one of a declared set of strings, matched literally and case-sensitively at word edges. 'Arm' is three characters, and 'arm', 'arms' and 'armed' are ordinary words that a Title Case headline capitalises, so the short name cannot be used. 'Arm Holdings' would be safe and would match almost nothing, because coverage writes 'Arm' alone. A list built from either choice would read as coverage while being either wrong or nearly empty, so no list is published. Same rule as everywhere else here: a gap is never closed by guessing.

Known

The items are collected and are already published on /archive/, reachable by source and by free-text search.

Would unblock

An indexed item list on the Arm company page. The page itself still carries the catalogue facts, its fab sites and its links.

Source

https://siliconalmanac.com/archive/

WSTS press releases older than the eleven its archive currently lists

The archive is the only index WSTS publishes and it is a moving window. The media ids do not follow from the date, so a URL for a release that has scrolled off cannot be constructed - only found somewhere that is not the publisher, which is not this record's source of truth.

Known

The archive at /76/PRESS-ARCHIVE listed eleven items on 17 August 2026, the oldest being the Spring 2023 forecast round. Each PDF sits at a path containing an opaque numeric media id.

Would unblock

a 2023 target line with three vintages and a result, and the result marker on the 2024 line

Source

https://www.wsts.org/76/PRESS-ARCHIVE

The WSTS full-year 2024 result

Deriving 2024 by dividing the 2025 result by its growth rate would produce a number WSTS never printed, presented as if it had. The 2024 target therefore carries a forecast and no result.

Known

The 2025 result release states 26% growth to USD 795.6 billion, which implies a 2024 figure

Would unblock

the result marker on the WSTS 2024 line

Source

https://www.wsts.org/76/103/Global-Semiconductor-Market-grows-26-in-2025-to-796B

WSTS quarterly releases other than Q2 2025

WSTS's public press archive holds only the most recent releases - eleven items in August 2026, the oldest a Spring 2023 forecast - and there is no index of past quarterlies. A URL cannot be guessed: the PDF paths carry an opaque numeric media id (7740 for Q2 2026, 7175 for Q2 2025) that does not follow from the date.

Known

WSTS issues quarterly releases that update the standing forecast. Two are now in the record: Q2 2025, which moved the 2026 target from 760.7 to 800, and Q2 2026, which moved 2026 to 1,655 and 2027 to approximately 2.1 trillion.

Would unblock

denser WSTS lines on every target

Source

https://www.wsts.org/61/Forecasts

300mm fab equipment spending for target 2024

A single point is a first forecast, not a stagger. It is held out rather than drawn as a line of one.

Known

The June 2023 release gives US$82 billion for 2024

Would unblock

a 2024 row in the SEMI 300mm matrix

Source

https://www.semi.org/en/news-media-press-releases/semi-press-releases/global-300mm-fab-equipment-spending-forecast-to-reach-record-$119-billion-in-2026-semi-reports

Company logos on the 161 company pages

A logo is a trade mark, and the free-and-public rule that governs the data does not extend to marks. More immediately: every number on this site links to the body that published it, so a logo taken from a search result would be the only asset here with no recorded origin. Files are therefore sourced from each company's own press kit or brand centre, one at a time, or the company renders as text with the reason published.

Known

The machinery is built and held: data/company-logos.json records provenance per logo, tools/check_company_logos.py refuses a file with no entry, and the pages render the company name as text until an entry is complete.

Would unblock

Faster recognition on a dense page. A reader recognizes a mark before they read a name.

Source

https://siliconalmanac.com/companies/

One worldwide series of industrial electricity prices, on one basis

Stitching several publishers into one map or one colour scale would assert a comparison none of them made: consumption band, tax treatment, currency and period all differ between them. This record would rather show each publisher's own panel, named and dated, than invent a world series. It would also be wrong in a second way that no caption fixes: a national average is not what a large fab pays, because industrial consumers of that size negotiate contracts.

Known

Eurostat publishes electricity prices for non-household consumers by consumption band, twice a year, free and redistributable, for the EU and several neighbours (nrg_pc_205). The US EIA publishes industrial-sector prices by state in the public domain. The IEA publishes a global set under licence.

Would unblock

a per-country electricity panel beside the fab map, one publisher per panel

Source

https://ec.europa.eu/eurostat/databrowser/view/nrg_pc_205/default/table?lang=en

TSMC's own company-wide capacity figure, from its Fab Capacity page: "Annual capacity of the manufacturing facilities managed by TSMC and its subsidiaries exceeded 17 million 12-inch equivalent wafers in 2025." The page then names the facilities counted, including JASM as a majority-owned subsidiary.

The page carries no date and no dated meta field; the only timestamp available is the server's Last-Modified, which is when the page was served and not when the sentence was written. Every figure in this record is positioned by the date its publisher issued it, and a guessed date puts a real number on the wrong day - the same refusal the WSTS Spring 2024 entry makes. The dated version of this sentence is almost certainly in TSMC's annual report, which is where to look next.

Known

The figure, the scope, the bound - "exceeded" is a lower bound - and a unit this record does not yet hold: 12-INCH EQUIVALENT WAFERS PER YEAR. That is a fourth wafer basis, distinct from the 300mm ACTUAL basis already in use, and an annual period where every other capacity figure here is monthly or weekly.

Would unblock

A COMPANY-WIDE capacity figure from the company itself, which is the one kind fab-capacity.json refuses to compute: that file holds per-fab statements and never adds them up. A publisher's own total is the only honest way to have one, and TSMC restates this sentence every year, which makes it a revision series rather than a fact.

Source

https://www.tsmc.com/english/dedicatedFoundry/manufacturing/fab_capacity

The Taiwan dollar leg of the cross-rate, as a series rather than a single date

The series could not be read in bulk from the machine this was built on. One date is enough to prove the cross-rate arithmetic and not enough to convert a history.

Known

The Federal Reserve H.10 series is free and current; its most recent observation is in the file

Would unblock

Taiwan figures in euro across time, which is the case that prompted this whole layer

Source

https://fred.stlouisfed.org/series/DEXTAUS

Data revisions: every value a statistical agency has published for a period that already elapsed

The collector was written on a machine with no outbound network. It is unit-tested against fixtures and raises rather than guessing when the CSV shape disagrees with what was requested, but it has never met alfredgraph.csv.

Known

ALFRED holds decades of vintages for the three FRED series this site already collects, free and public domain

Would unblock

the entire second table of the record, past-to-past

Source

https://alfred.stlouisfed.org/

Euro reference rates before 2026-05-15

Only the 90-day file was readable from the machine this was built on. Every conversion of a figure older than the covered span therefore refuses rather than substituting a rate, which is the correct behaviour but a thin demonstration of it.

Known

The ECB publishes its full history as one file, free

Would unblock

conversion of every figure on the site, not just recent ones

Source

https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/index.en.html

Items from four company newsrooms configured in data/sources.json

Found by the first live run of the per-company generator, which reported own_total 0 for companies that have a source configured. The cause is not diagnosed: the feeds are named correctly, since the checker verifies every own_source against sources.json and the five that work join on exactly the same field. So the feeds either answer with nothing, have moved, or are being dropped by the collector for a reason no fault has yet named. Guessing which would be closing a gap by guessing.

Known

Nine company-published sources are configured. Five have delivered: ASML newsroom (479), ASML official YouTube (15), Samsung Global Newsroom (52), Nvidia Newsroom (20), AMD Investor Relations (10), imec YouTube (15) — 591 items in total, measured 17 August 2026.

Would unblock

The strongest list on four company pages, the one that carries no inference at all because the publisher IS the company.

Source

https://siliconalmanac.com/method/

Schedule revisions: a promised milestone date moving

It is the table about named companies. The neutral visual and verbal language is being proved first on statistical-agency and trade-body figures, where no party is embarrassed by the chart.

Known

The schema is decided and written down

Would unblock

the third table