The rows that turned out to be one fab.
Seven rows on the fabs page were two descriptions of one building. This is every ruling, what it turned on, and what happened to the losing row. The suspicion is mechanical; the ruling is a person reading two sources.
Why this file exists
In the record's own words, from the file.
fabs.json was seeded from Wikipedia's list of semiconductor fabrication plants and then grown with rows sourced to operators and to the trade press, and nothing deduplicated the two passes. Seven sites appear twice. Two of the pairs contradict each other on status or wafer size, so this is a correctness question and not a tidiness one, and 194 is printed on the home page.
8 pairs ruled: 7 one fab described twice, 1 genuinely two fabs on one site.
| Verdict | What it means |
|---|---|
| same-fab | The two rows describe one physical fab. `keep` names the row that survives and `drop` the one that goes. The checker refuses once a `drop` row is still present in fabs.json, so a ruling that nobody acted on cannot read as handled. |
| distinct | They really are two fabs on one site. `why` says what established it, and it must be a publisher's statement, not a reading of the ids. |
| unresolved | Nobody has ruled. The checker reports it and does not refuse, unless run with --strict. |
One fab, recorded twice
7 pairs. The losing row was deleted from the record and the checker refuses to pass while a dropped row is still present.
| The two rows | What the sources said | Outcome | Ruled |
|---|---|---|---|
| bosch-reutlingen bosch-reutlingen-germany | One fab. The dropped row's name 'Reutlingen (main + WaferFab)' suggested two, and Bosch describes one: 'The Bosch wafer fab in Reutlingen is the company's most established and experienced semiconductor manufacturing site', with a single cleanroom figure of 44,000 m2 and frontend and backend on one location. Bosch's site list carries two process-line bullets under the single Reutlingen entry, which is two lines inside one fab. Bosch states the site 'is a pure 200-millimeter plant as of 2026', so 200mm stands and the former 150mm SiC line is not current. ⛔ Open: the type is analogue-power where Bosch's own bullet reads 'automotive ICs, MEMS, and power MOSFETs', and MEMS has been mass-produced there since 1995. | kept bosch-reutlingen-germany, dropped bosch-reutlingen | 2026-08-21 |
| intel-fab38-kiryatgat intel-kiryatgat-fab38 | Both rows were named Fab 38 at the same coordinates with the same status and wafer size, and Intel's own construction page lists ONE Kiryat Gat project: 'Kiryat Gat ... will soon welcome Fab 38. Located adjacent to Fab 28, the $25 billion expansion will produce advanced chips on EUV lithography.' Globes describes Fab 38 as one complex containing two cleanroom modules, and no publisher names or numbers those modules separately. The surviving row is renamed from 'Fab 38 (expansion)' to 'Fab 38', because '(expansion)' describes what it is relative to Fab 28 and is not a fab name. ⛔ The 300mm is this record's inference from EUV; Intel has never printed a wafer diameter for Fab 38, and Intel has published nothing on it since April 2025 while Globes and Calcalist report parts of it frozen. | kept intel-kiryatgat-fab38, dropped intel-fab38-kiryatgat | 2026-08-21 |
| microchip-colorado-springs microchip-coloradosprings-co | ⛔ One fab, and the Wikipedia row was the one that was RIGHT on the disputed field. Microchip's FY2026 10-K: 'Fab 5, located in Colorado Springs, Colorado, currently manufactures discrete and specialty products in addition to a lower volume of a diversified set of standard products on 6-inch wafers.' The 200mm on the surviving row came from the February 2023 expansion release, which says the opposite of what was recorded: 'The Colorado Springs campus currently employs more than 850 people and produces products from 6-inch wafers. The manufacturing technology that Microchip is installing will run on 8-inch wafers.' The 10-K adds that the expansion was paused through fiscal 2027. Wafer size CORRECTED to 150 and the 10-K added as the source for it. | kept microchip-coloradosprings-co, dropped microchip-colorado-springs | 2026-08-21 |
| microchip-gresham microchip-gresham-or | One fab, no contradiction. Microchip's FY2026 10-K: 'Fab 4, located in Gresham, Oregon, currently produces 8-inch wafers using predominantly 0.11 microns to 0.5 microns manufacturing processes.' The fab Microchip closed in May 2025 was Fab 2 in Tempe, Arizona; Gresham received Fab 2's process technologies. ⛔ Worth knowing: Wikipedia's Gresham row is onsemi's former LSI fab on NE Glisan Street, about 3 km from Microchip's Fab 4 on SE Stark Street, so the dropped row's Wikipedia citation looks like a mis-citation. That onsemi fab is NOT in this record and, if it is ever added, it must not be merged into this row on the shared city-centroid coordinate. | kept microchip-gresham-or, dropped microchip-gresham | 2026-08-21 |
| pragmatic-durham pragmatic-durham-uk | One fab, and the dropped row's name was wrong twice. 'FlexLogic 001/002/003' implied three: FlexLogIC-001 is a DIFFERENT site, a 200mm line at Sedgefield, and no third line exists — Pragmatic states only that the park has 'capacity to host up to nine fabrication lines'. The Durham 300mm line is FlexLogIC-002 at Pragmatic Park, Meadowfield. The type was CORRECTED from rd-pilot to foundry: rd-pilot appears in no source, and Pragmatic sells the site as 'FlexIC Foundry services' with 'low-cost, high volume production' on 'state-of-the-art 300mm wafers'. Renamed to 'Pragmatic Park (FlexLogIC-002)'. ⛔ Open: FlexLogIC-001 at Sedgefield is a genuinely separate fab and is MISSING from this record; its current status could not be established from any 2025 or 2026 source, so it was not added. | kept pragmatic-durham-uk, dropped pragmatic-durham | 2026-08-21 |
| wolfspeed-durham wolfspeed-durham-nc | ⛔ The pair that mattered. Wolfspeed's FY2026 Form 10-K states it in the past tense: 'During the period ended June 28, 2026, we closed the 150mm device fabrication facility in Durham, North Carolina and now produce all power devices in our 200mm Marcy, New York fabrication facility.' The Q2 FY2026 earnings release lists 'Completed the shutdown of 150mm device production at Durham fab, one month ahead of schedule.' The Wikipedia row said operating. The 10-K is now a source on the surviving row, which is better than the trade-press link it had. Note for anyone reading the Durham coordinates later: Wolfspeed still runs a silicon carbide MATERIALS facility in Durham and is headquartered there, and the RTP 100mm GaN fab nearby went to MACOM. Neither is this row and neither is currently in the record. | kept wolfspeed-durham-nc, dropped wolfspeed-durham | 2026-08-21 |
| xfab-lubbock xfab-lubbock-texas | X-FAB operates ONE fab in Lubbock. The 150mm/200mm conflict was not two fabs, it was an unsourced figure: X-FAB itself says 'we are processing SiC wafers in our 6" fab in Lubbock, Texas' and its careers site says 'the production of 6-inch wafers in the CMOS area as well as SiC technologies'. NIST describes the CHIPS award as 'expansion and modernization of X-Fab's Silicon Carbide (SiC) foundry facility', an expansion of the existing fab rather than a new one, and no X-FAB, NIST or trade source anywhere states 200mm at Lubbock. The 200mm was therefore DROPPED rather than merged. The surviving row keeps 150mm and gains the SiC node value and the NIST and eeNews sources. ⛔ Two things left open: its type is 'compound' where X-FAB describes a foundry running both silicon CMOS and SiC, and the 'formerly Texas Instruments' in its operator field appears on no X-FAB page. | kept xfab-lubbock, dropped xfab-lubbock-texas | 2026-08-21 |
Two fabs that share a site
1 pairs. Both rows stand, and both are on the map, because a publisher names each fab separately.
| The two rows | What the sources said | Outcome | Ruled |
|---|---|---|---|
| intel-fab38-kiryatgat intel-kiryatgat-fab28 | Intel's own newsroom names Fab 28 at Kiryat Gat as the operating fab, and the 2026 trade-press coverage names Fab 38 as a separate build on the same site. Two fabs, one campus, both stated by a publisher. | both rows kept | 2026-08-21 |
What the seven taught
The finding, which is more useful than any single ruling.
The failure mode in this dataset is TENSE, not fabrication. An expansion press release states a wafer size the company is GOING to run; the annual report states the one it runs. Two of the seven were wrong because a plan was recorded as a fact, and one was wrong because a closure announced in 2024 and completed in 2026 was never read back into the row. Where a company files an annual report, the annual report beats the press release for anything stated as current.
How a ruling is closed. Read the two rows' sources. If the operator's own page names both fabs separately, it is `distinct`. If only one fab exists and the second row is Wikipedia's older description of it, it is `same-fab`, the surviving row is the one sourced to a publisher rather than to the list, and the losing row's `sources` are merged into the survivor before it is deleted. Then delete the row and run the checker.
Every ruling names who made it and invites being overturned. A reader who reads the two sources differently is reading the same two sources.
What was acted on. 2026-08-21: all seven same-fab rows deleted from data/fabs.json in the same pass, and the corrections applied. fabs.json went from 194 rows to 187.