# Mean Free Path - all essays (full text)
A working notebook on vacuum: the physics, the machines, and the people who run them, from industrial electron tubes through accelerators and light sources into fusion, quantum hardware, semiconductor fabrication, and space. Written from inside the work by Pierre F. Ribault.
Site: https://pierreribault.com
Author: Pierre F. Ribault (vacuum scientist)
Contact: pierre@pierreribault.com
License (data): CC BY 4.0, Mean Free Path, Pierre F. Ribault
---
# A number is a claim
> This site used to show outgassing values for 64 materials. It now shows 20. Here is why that is progress.
Date: 2026-09-01
Canonical: https://pierreribault.com/essays/a-number-is-a-claim/
Until this week, this site displayed an outgassing signature for 64 materials. Today it displays 20.
Nothing was measured differently. No material got worse. What changed is that I built [a database that refuses to serve a number without its source](/tools/outgassing/), then pointed the site at it. Forty-four signatures did not survive the question "who says so?"
I want to explain why I consider this an upgrade, because the instinct says otherwise. A table with 64 rows looks more complete than a table with 20. Completeness is what most references optimize for. It is also how a rumor spreads.
## Where the 64 came from
The old values came from where everyone's values come from. Vendor tables that list a single number per material, no species, no temperature. Course notes. The memory of the trade. A 1e-11 for baked stainless that I have heard in three languages and never once seen attached to a measurement.
None of these numbers were wild. Most sat within an order of magnitude of something defensible. That is precisely what makes them dangerous. A wrong number gets caught. A plausible orphan gets repeated.
## What sourcing them actually taught me
Building the base meant reading the primary documents, and the primary documents are stranger than the tables that cite them.
The most used materials list in gravitational wave detection publishes 95 material rows. Fourteen carry an outgassing figure. None of the fourteen states a pumping time. For water on unbaked stainless, that list gives a value nearly two orders of magnitude above what CERN publishes at ten hours of pumping. The two are not in disagreement about the steel. They are answering different questions, and only one of them says which. The CERN reference itself is blunt about this: a water outgassing rate quoted without its pumping time carries no meaning.
The same CERN lecture note publishes two values for silver bearing copper after the same 200 C bake, in two tables, credited to the same measurements. A factor of three apart. Neither retracted.
The LIGO acceptance threshold on the five hydrocarbon masses appears in two specifications of the same corpus. Same number, 2e-12. One document writes it per square centimetre. The other writes it with no area at all. Two different physical quantities wearing the same digits.
Two respected compilations disagree on unbaked molybdenum by a factor of 131. And the constant that converts a rate into monolayers per year, the one this site itself engraved in an earlier essay, exists in the literature as both 5e14 and 1e15 molecules per square centimetre, a factor of two, with neither source naming the molecule it stacks.
I did not find these problems by looking for problems. I found them by writing down, for every value, where it comes from and under which conditions it holds. The contradictions surfaced on their own. The base now publishes ten of them, side by side, each value under its key, with no arbitration. Picking a winner would just manufacture one more orphan.
## The rule the site now follows
Every value in the base carries a source key, the conditions as the document states them, the species, the measurement temperature, and the pumping time or an explicit declaration that the source omitted it. A value that cannot carry all of that does not enter. A material that has no defensible value gets a declared absence with its reason, which is information, where a silently reused number is noise.
That rule is what shrank the column. The 44 signatures that vanished were never deleted. They were never really there.
The same rule runs forward into the calculators. Every result now cites the key of the value it consumed. If the base holds two conflicting values for your material, the calculator returns a range and both keys instead of a quiet choice. If your operating point falls outside the validity range a source declared, the calculator refuses and says why. Sixty-four of the 379 readings on its menu exist to be refused. A tool that always returns a number is lying about at least some of them.
## The part that concerns you
Take the outgassing figure you use for stainless steel. The one in your gas budget spreadsheet, your design review slide, your mental arithmetic at the RGA screen.
Where does it come from? At what pumping time? Which species? After which bake, measured how?
If you can answer, you are holding a measurement. If you cannot, you are holding a claim, and you have been passing it along. I was. Twelve years in vacuum, and until I forced every number on this site to show its papers, I could not have told you which of my own figures were measurements and which were folklore with good posture.
The base is at [/tools/outgassing](/tools/outgassing/), the calculators at [/tools/calculators](/tools/calculators/), the raw data at [/data/outgassing.json](/data/outgassing.json) under CC BY. Twenty signatures today, 258 entries, 18 sources, thirteen and a half decades from a getter pumping krypton to neoprene shedding water. It will grow. It will grow slowly, because every new number has to answer the question first.
A number is a claim. A source turns it into evidence. Everything else is rumor with units.
## Sources
- [A] LIGO-E960050-v13, LIGO Vacuum Compatible Materials List, D. Coyne (ed.), 4 June 2014, Table 1 Approved Construction Materials. The table publishes 95 material rows; fourteen carry an outgassing figure, and no pumping time or measurement temperature appears anywhere in the table. Rows A14 and I9 carry a figure whose species column could not be resolved, and the base records them as declared absences rather than guessing. — https://dcc.ligo.org/public/0003/E960050/013/E960050-v13%20Vacuum%20Compatible%20Materials%20List.pdf
- [A] P. Chiggiato, Outgassing properties of vacuum materials for particle accelerators, CAS Glumslov 2017, arXiv:2006.07124. Source of the water rate on unbaked stainless read at ten hours of pumping, of the statement that a water outgassing rate without its pumping time is meaningless, and of the two silver-bearing copper values after the same 200 C bake, Table 3 against Table 4. — https://arxiv.org/abs/2006.07124
- [A] LIGO-E960022-v24, LIGO Vacuum Compatibility, Cleaning Methods and Qualification Procedures, B. Bland, D. Coyne and J. Fauver (eds.), 27 June 2012, section 13.17.2.1: the hydrocarbon acceptance criterion on AMU 41, 43, 53, 55 and 57, written with an area in the denominator. — https://dcc.ligo.org/public/0003/E960022/024/E960022-v24.pdf
- [A] LIGO-E080177-v2, RGA Test Qualification of components for the LIGO UHV, section 6d: the same five masses and the same figure, written with no area. The base publishes both under the key sc-hc-threshold-unit, with no arbitration. — https://dcc.ligo.org/public/0008/E080177/002/E080177-v2%20RGA%20test%20qualification.pdf
- [A] R. J. Elsey, Outgassing of vacuum materials II, Vacuum 25(8) (1975) 347-361, Pergamon Press, The Rutherford Laboratory, Chilton, England. No DOI is printed on the document. Source of the molybdenum value read at one hour of pumping that the base sets against the LIGO list. — https://doi.org/10.1016/0042-207X%2875%2991653-X
- [A] R. J. Elsey, Outgassing of vacuum materials I, Vacuum 25(7) (1975) 299-306. Source of the monolayer surface density of 1e15 molecules per square centimetre this site uses. — https://doi.org/10.1016/0042-207X%2875%2990730-7
- [A] D. K. Benson and G. A. Beitel, Development and evaluation of vacuum pressure gauge components from carbon and graphite, NASA CR-2101, Midwest Research Institute report C-2922, contract NAS 1-10738, Langley Research Center, September 1972. The footnote on page 11 defines one monolayer as 5e14 molecules per square centimetre and uses that definition throughout its table I. — https://ntrs.nasa.gov/citations/19720023813
- [A] The outgassing base itself, /data/outgassing.json, published under CC BY 4.0. Every count in the closing paragraph is read from it: 258 entries, 18 source keys, 20 material signatures, and a span of 13.6 decades from a TiZrV getter pumping krypton at 1e-18 to neoprene at 3.99e-5. The ten contradictions are its site_conflicts block.
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary. The counts in this essay were reconciled by hand against the base on 2026-09-01, the discipline set out in ESSAY_PUBLISH_CHECKLIST.md: an essay is a dated artifact and its numbers are not wired to the build. The gate check:outgassing verifies that the base does not contradict them.
---
# Clean Has No Unit
> Nine trades prepare surfaces for vacuum. No two of them measure the same thing. The unit has existed on paper since 2013. No vacuum handbook writes it on a purchase order.
Date: 2026-08-24
Canonical: https://pierreribault.com/essays/clean-has-no-unit/
import PlateMonolayerLadder from '../../components/PlateMonolayerLadder.astro';
import PlateFiveMasses from '../../components/PlateFiveMasses.astro';
import PlateWhoReadsWhat from '../../components/PlateWhoReadsWhat.astro';
import ProseImage from '../../components/ProseImage.astro';
import coupons from '../../assets/essays/clean-coupons.jpg';
import rinse from '../../assets/essays/clean-rinse.jpg';
import fingerprint from '../../assets/essays/clean-fingerprint.jpg';
A purchase order carries a line that reads UHV clean. No number follows the two words. The part arrives ten days later, wrapped in aluminium foil inside a heat-sealed bag, with a certificate that repeats the two words in a different font. Somewhere between the machine shop and the flange, a person decided what clean meant that week. The order never said.
The same two words mean nine things, depending on who reads them. A contractor who cleans oxygen pipelines, a contamination control engineer on a satellite, a technician rinsing a niobium cavity at a hundred bar, a fab engineer counting iron atoms on a wafer, a physicist etching the oxide off a resonator: each of them says clean, each of them means a number, and the numbers sit seven decades apart. This essay lines them up on one scale, names the instrument behind each one, and ends with the number a vacuum buyer could write on the order.
Clean is a relation between a surface and the process that surface will undergo. A surface is clean when what it carries stays below the threshold at which the process fails. Three terms: a quantity, a species, a failure mode. Change the process and the same surface changes status. Two standards already say this in their own words. The aerospace cleanliness standard IEST-STD-CC1246E puts its emphasis on contaminants that impact product performance. The European space standard ECSS-Q-ST-70-01C, revised in October 2025, defines a contaminant as unwanted matter that can affect or degrade performance or lifetime. The performance decides. The surface reports.
What a surface carries sits on two axes. The molecular axis is a surface density, molecules per square centimetre, readable in monolayers, about 1e15 small molecules per square centimetre for a full one. The particulate axis is a count per area by size, or a fraction of the area covered. Every trade reads one axis with one instrument and gives the reading its own name. The parts per billion, the milligram per square metre, the ISO class and the mbar litre per second per square centimetre are four projections of the same two quantities.
Line the molecular thresholds up on one scale and the trades sort themselves.
Oxygen service sits at the top. A component ordered to CGA G-4.1 arrives with at most 220 milligrams of non-volatile residue per square metre. Older editions accepted about 500. For a hydrocarbon oil of ordinary molar mass, that is a few hundred monolayers. The enemy is ignition by adiabatic compression, and a few hundred monolayers of oil under a few hundred bar of oxygen is a fuel load. The test is a solvent rinse weighed on a balance, a black light, a wipe and a dew point. NASA's oxygen ground test systems below 5,000 psi trip an approval at 5 milligrams per tenth of a square metre and a forced re-clean at 20.
Space hardware sits a decade below that. IEST-STD-CC1246E writes residue in milligrams per tenth of a square metre. The old Level A, now called R1, is one milligram: about ten monolayers of oil, give or take a factor of two on the footprint of the molecule. The same standard keeps counting down to ten nanograms, five decades below R1. The enemy is a film that arrives on an optic or a radiator after launch, when nobody can wipe it. Space is the trade that pushed the residue scale furthest in both directions, and its parts are bought, shipped and verified by people other than those who will suffer the film.
LIGO sits at A/50. The interferometer buys its piece parts to the aerospace standard, a fiftieth of the old Level A, a tenth of a monolayer, and its 2010 specification still asks for assembly in a Class 100 cleanroom per FED-STD-209, a federal standard withdrawn in 2001 when ISO 14644-1 replaced it. The enemy is a heavy hydrocarbon on a mirror, which the laser reads as phase noise.
Then the wafer. The semiconductor roadmap wrote its surface metal limit at 1e10 atoms per square centimetre after 2009, one hundred thousandth of a monolayer, and its organic limit at two nanograms per square centimetre. Read from two directions by two industries that never cite each other, the finest level of the space standard and the metal target of the fab meet at 1e10 per square centimetre.
Seven decades. Each trade sits at the top of one column and is absent from the others. The oxygen contractor's best work fails any space part. The fab's number would bankrupt the oxygen contractor. Both numbers are right, because the failure modes differ.
The vacuum trade reads a flux. Accelerator vacuum measures outgassing, gas leaving a square centimetre per second, in mbar litre per second per square centimetre. The number describes an inventory only after a conversion that nobody performs on the shop floor. One mbar litre holds 2.5e19 molecules at room temperature. A rate of 1e-12 mbar·L/s/cm² is 2.5e7 molecules per second per square centimetre, or 0.8 monolayer per year. The ITER Vacuum Handbook sets its impurity limit for torus components at 1e-9 Pa·m³/s/m² at 100 °C. Convert it. One monolayer per year, to the rounding. Chiggiato's table gives austenitic stainless steel after a 150 °C bake at 3e-12 for hydrogen, three monolayers per year, drawn from the bulk of the metal rather than from the surface. His warning covers every unbaked figure ever quoted: without the pumping time, the number means nothing.
CERN qualifies a cleaning procedure without an RGA. Standard coupons are contaminated on purpose, cleaned by the candidate recipe, and read by XPS within thirty minutes of leaving the air. Carbon at or under 40 atomic percent on stainless steel passes. The number is an atomic fraction inside the first few nanometres, roughly one monolayer of adventitious carbon, and the recipe behind it has been public for twenty years: detergent at 50 to 60 °C under ultrasound, demineralised water rinse below 5 microsiemens per centimetre, oven at 80 to 100 °C or filtered nitrogen. The first check happens on the shop floor, by eye. The rinse water either wets the steel in a continuous sheet or breaks into drops. A sheet that breaks fails the part before any instrument sees it.
LIGO is the one house that reads both axes in both units. Its parts are bought in aerospace units and qualified in vacuum units. After cleaning and a vacuum bake, the part cools and an RGA reads it. The rule fits on a page. The 43 peak stays at or below a tenth of the 44 peak. Everything above 44 stays at or below a hundredth of it. The calibrated sum of five masses, 41, 43, 53, 55 and 57, the masses the laboratory found indicative of every heavy hydrocarbon, stays below 4e-10 torr litre per second for a suspension structure, and a small load reads background, near 2e-12. No subtraction of an empty chamber. An argon and krypton leak calibrates the scan. Written in 2008 and still in force.
Who does the gesture splits the nine trades into two families.
Space, oxygen and semiconductor chamber parts write the number and buy the gesture. A verification industry stands between them and their suppliers, with balances, particle counters, plasma mass spectrometers and photographs of filters. A fab's chamber liners, shields and showerheads leave the building and come back from a parts cleaning contractor, with extractable metals reported in nanograms per square centimetre and particles counted in the extraction liquid down to 30 nanometres. The acceptance values stay between the tool maker and the contractor. I found none of them in public. The oxygen trade's numbers are printed in a standard anyone can buy.
Accelerators, SRF, the wafer fab, the quantum lab and LIGO keep the gesture in house and let the process itself verify it. The niobium cavity is rinsed at a hundred bar of ultrapure water, assembled in an ISO 4 room, and judged in the vertical test, where field emission from a micron particle shows up as radiation and a quality factor that collapses. Two SRF authors wrote in 2012 that the rinse protocols had evolved by loose empiricism rather than quantitative process development, and that pressure, nozzle, distance and sweep rate differed from one laboratory to the next. The judgement is severe and reliable. The recipe is local.
The frontier buys. A fusion startup or a quantum hardware company owns no surface treatment workshop. It orders chambers, bellows and feedthroughs from the same catalogues as everyone else, and its purchase order carries the two words. ITER wrote its numbers in 2019: torus components at 1e-9 Pa·m³/s/m² of impurities, cutting fluids under 200 ppm each of halogens, phosphorus and sulfur, halogenated solvents forbidden at every stage, handling areas at five million particles above half a micron per cubic metre, a class between ISO 8 and ISO 9, ten thousand times looser than the SRF room. The ITER vacuum group runs outgassing tests on candidate materials almost continuously, and its group leader has noted that a material fit for a fab is not automatically fit for a tokamak, and the reverse. I have not found an equivalent public document from a private fusion company.
Quantum inverts the quantity. In a niobium resonator on silicon, the loss that limits coherence lives in a few nanometres of oxide at the metal-air and substrate-air interfaces. A Berkeley team measured a median loss tangent of 1.07 parts per million at the single photon level, removed the substrate oxide by selective etching and read 0.48, thinned the niobium oxide from 4.8 to 1.6 nanometres and read 0.19. The quality factor rose from 0.9 million to 5.3 million. The contaminant was the material's own skin. Clean, on that bench, means fewer than zero monolayers of oxide.
The gesture itself changed in 2025. Two solvent families carried precision degreasing for sixty years, the chlorinated and the fluorinated. On 17 December 2024 the US EPA finalised a rule prohibiting all uses of trichloroethylene, most of them from 15 September 2025, with the exemption dates postponed six times since and a final rule on extensions on the agenda for October 2026. 3M had announced in December 2022 its exit from PFAS manufacturing, and the last order date for its Novec hydrofluoroethers fell on 31 March 2025. Production stopped at the end of that year. The vapour degreaser lost both fluids in twelve months. The aqueous recipe CERN published for accelerator parts, detergent, ultrasound, demineralised rinse, hot air, has become the default of a trade that used to look down on water. A ring laser gyroscope team wrote in June 2026 that cleaning had to enter its next procurement as an explicit specification, with certificates and a record of every electropolishing, ultrasonic and bake step, after a titanium cavity taught it the lesson the hard way.
The unit exists. ISO 14644-10 grades surface cleanliness by chemical concentration on a logarithmic scale: grade N stands for ten to the power N grams per square metre, from 0 down to minus 12, in steps of a tenth. ISO 14644-9 does the same for particles. Both were revised in 2022. ISO 14644-13, the guide to the cleaning methods that reach those grades, came out in its second edition on 25 February 2026. On that scale the oxygen part sits at minus 0.3, the space part at minus 2, the LIGO part near minus 4, the wafer's organics near minus 5, its metals at minus 8. One number, one scale, thirteen decades, and a method guide six months old. The European space standard cites the series. The ITER Vacuum Handbook cites neither part 9 nor part 10. The LIGO part-qualification specifications cite FED-STD-209 and IEST-STD-CC1246D instead. The CERN school's cleaning lecture cites neither. The vacuum trade has a surface unit on offer and keeps its flux.
The flux can be translated. One monolayer per year is 1e-12 mbar·L/s/cm², and LIGO's five masses give it a species. A part is UHV clean when its purchase order says how many monolayers of what may leave it per year, at what temperature, read by which instrument, against which empty-chamber scan. The oxygen contractor already writes his number. The satellite engineer already writes hers. The startup buying its first chamber can write one too, and the supplier who cannot meet it will say so before the flange closes.
Four bets, dated, so that this essay can be wrong in public. By the end of 2028, at least one private fusion company publishes a supplier specification with a numeric cleanliness acceptance, an outgassing rate, a residue level or an ISO 14644-10 grade. By the end of 2030, no big-science vacuum handbook cites the surface grades of ISO 14644-10. By the end of 2026, the EPA's rule on compliance-date extensions leaves the prohibition of trichloroethylene in open-top vapour degreasing in force. By the end of 2028, a superconducting-qubit foundry or a substrate supplier publishes an incoming acceptance value for surface oxide thickness in nanometres, read by XPS.
A clean part is a part whose number is written down.
## points of rendezvous
Dates this essay can be checked against:
- **October 15, 2025.** ECSS-Q-ST-70-01C Rev.1 published.
- **December 31, 2025.** End of 3M PFAS manufacturing, Novec fluids included.
- **February 25, 2026.** ISO 14644-13:2026 published, edition 2.
- **May 18, 2026.** Sixth postponement date of the TCE section 6(g) exemption provisions.
- **October 2026.** EPA final rule on compliance-date extensions for TCE, PCE, methylene chloride and carbon tetrachloride.
- **December 18, 2026.** TCE disposal-to-wastewater prohibition compliance date for processors.
- **December 2026.** ECHA SEAC opinion on the universal PFAS restriction.
- **2027.** ISO 14644-9:2022 and ISO 14644-10:2022 reach the five-year systematic review window.
- **September 15, 2028.** TCE as a processing aid in nuclear fuel manufacture, prohibition compliance date.
## Sources
- [A] [S1] IEST-STD-CC1246E (2013), Product Cleanliness Levels, Applications, Requirements, and Determination, Institute of Environmental Sciences and Technology. Read through E. N. Borson, IEST-STD-CC1246D: Product Cleanliness Levels and Contamination Control Program, CleanRooms, August 2005 (sst.semiconductor-digest.com), and Astro Pak, Precision Cleaning Standards (astropak.com/precision-cleaning-standards, 2026-04-08). Takes: R1 at 1 mg/0.1 m², smallest NVR level 10 ng/0.1 m² since revision D, particle level 50 bins, emphasis on performance-relevant contaminants. Rating [A] for the levels, [B] for the bins.
- [A] [S2] ECSS-Q-ST-70-01C Rev.1 (15 October 2025), Space product assurance, Cleanliness and contamination control, European Cooperation for Space Standardization, ecss.nl. Cancels ECSS-Q-ST-70-01C (15 November 2008). Takes: revision date, tables on ISO 14644-1:2015 classes and IEST-STD-CC1246E correlation (DIR1 table of contents, 18 November 2024). Rating [A].
- [B] [S3] CGA G-4.1 (2018), Cleaning Equipment for Oxygen Service, Compressed Gas Association. Read through Astro Pak (section 9.1, 220 mg/m² baseline) and Penflex Engineering Bulletin 148, Oxygen Service Cleaning (older acceptance about 500 mg/m², particle clause). Rating [B].
- [B] [S4] ASTM G93/G93M-19, Standard Guide for Cleanliness Levels and Cleaning Methods for Materials and Equipment Used in Oxygen-Enriched Environments, ASTM International. Takes: risk-based levels A to D, no baseline. Rating [B], values not read. — https://www.astm.org/g0093_g0093m-19.html
- [A] [S5] MSFC-SPEC-164E, NASA Marshall Space Flight Center, cleanliness specification for oxygen, fuel and pneumatic systems, standards.nasa.gov. Takes: section 3.7.2, oxygen ground test systems below 5,000 psig, Level B 5 mg/0.1 m² approval trigger, 20 mg/0.1 m² re-clean. Rating [A] for the values, [B] for the title. — https://standards.nasa.gov/standard/MSFC/MSFC-SPEC-164
- [A] [S6] R. Pearce, L. Worth, ITER Vacuum Handbook, ITR-19-004, ITER Organization, 19 November 2019, CC BY-NC-ND 3.0 IGO, iter.org. Takes: Table 5-1 outgassing rates by VQC, Section 6.1 cutting fluids, Section 24.1 cleaning and forbidden solvents, Table 24-1 handling areas at 5e6 particles above 0.5 µm per m³, Table 26-1 bake temperatures, Appendix 13 (ITER_D_2ELUQH), Appendix 17 (ITER_D_2EXDST). Rating [A]. — https://www.iter.org/sites/default/files/media/2024-04/iter_vacuum_handbook.pdf
- [A] [S7] M. Taborelli, Cleaning and Surface Properties, Proceedings of the 2017 CERN Accelerator School on Vacuum for Particle Accelerators, Glumslöv, arXiv:2006.01585. Slides CAS-Lund, June 2017, indico.cern.ch. Earlier version CAS 2006, CERN-2007-003, p. 321. Takes: detergent recipe, wetting check, XPS acceptance 40 at% C, contaminant classes. Rating [A]. — https://arxiv.org/abs/2006.01585
- [A] [S8] P. Chiggiato, Outgassing properties of vacuum materials for particle accelerators, arXiv:2006.07124, CAS 2017. Takes: Table 4 hydrogen outgassing after bake, warning on unbaked values without pumping time. Rating [A]. — https://arxiv.org/abs/2006.07124
- [A] [S9] D. Coyne, Qualifying Parts for LIGO UHV Service, LIGO-E1000088-v1, 19 March 2010, dcc.ligo.org. Takes: NVR A/50 and particulate Level 50 on piece parts, Class 100 per FED-STD-209, RGA preferred, FTIR at JPL for large parts. Rating [A]. — https://dcc.ligo.org/LIGO-E1000088/public
- [A] [S10] LIGO-E960022, LIGO Clean and Bake Methods and Procedures, current version v26, 11 December 2024, dcc.ligo.org. LIGO-E1000088-v1, Qualifying Parts for LIGO UHV Service, the specification that cites FED-STD-209 and IEST-STD-CC1246D. LIGO-E0900480, FTIR Testing to Qualify Parts for LIGO UHV Service. LIGO-E960050-B, Vacuum Compatible Materials List. Takes: method, hydrocarbon fragment comparison against qualification baseline. Rating [A] for method.
- [A] [S11] C. E. Reece, G. Ciovati, Superconducting RF Technology R&D for Future Accelerator Applications, arXiv:1208.1978 (2012), section 4.4. Takes: HPR above 100 bar as final step, protocols evolved by loose empiricism, field emission free Nb to at least 150 MV/m. Rating [A]. — https://arxiv.org/abs/1208.1978
- [A] [S12] H. Padamsee, History of gradient advances in SRF, arXiv:2004.06720. Takes: emitters are micron and sub-micron particles, Saito KEK 1994 at 100 bar, Class 100 or better. Rating [A]. — https://arxiv.org/abs/2004.06720
- [A] [S13] Key directions for research and development of superconducting radio frequency cavities, arXiv:2204.01178, section 8. HIM cleanroom for SRF activities, SRF2019 THP101, jacow.org. UKRI, Facility puts pressure on for cavity testing, ukri.org. Jefferson Lab SRF R&D page, jlab.org/accelerator/srf/rddevs. Takes: ISO 4 assembly, 100 bar, 12 h cycle. Rating [A] for the papers, [B] for the blogs.
- [B] [S14] ITRS surface molecular contamination deposition limits reproduced in UC Berkeley Microlab seminar slides, microlab.berkeley.edu, chia1.pdf, 2005 values. IRDS 2024 Yield Enhancement chapter, irds.ieee.org. ITRS 2.0 2015 Emerging Research Materials chapter, section 5.4, semiconductors.org. Takes: metals 1e10 atoms/cm² after 2009, organics 2 ng/cm², monolayer 1e15 atoms/cm² on Si(100), defectivity under 0.01 printable defects/cm², metal ions under 1 ppb. Rating [B] for the 2005 limits, [A] for the chapters.
- [A] [S15] US 6,810,887, Method for cleaning semiconductor fabrication equipment parts, uspto.gov. WO 2017/180304 A1 and US 10,453,709, Applied Materials, liquid particle counting of semiconductor component parts. Takes: extraction ICP-MS in atoms/cm² or ng/cm², TOC and GC-MS, in-line LPC to 30 nm. Rating [A] for methods, acceptance values not public.
- [A] [S16] M. V. P. Altoé et al., Localization and Mitigation of Loss in Niobium Superconducting Circuits, PRX Quantum 3, 020312 (2022), doi:10.1103/PRXQuantum.3.020312. Numeric sequence from the APS March Meeting 2022 abstract Q41.3. Takes: 1.07 to 0.48 to 0.19 ppm, oxide 4.8 to 1.6 nm, Q 0.93e6 to 5.26e6. Rating [A]. — https://doi.org/10.1103/PRXQuantum.3.020312
- [B] [S17] ISO 14644-1:2015, Cleanrooms and associated controlled environments, Part 1: Classification of air cleanliness by particle concentration. Takes: ISO 4 at 352, ISO 5 at 3,520, ISO 8 at 3,520,000 particles per m³ at 0.5 µm. Rating [B], from memory of the table.
- [B] [S18] J. F. O'Hanlon, A User's Guide to Vacuum Technology, 3rd edition, Wiley, 2003. Takes: monolayer of order 1e15 molecules/cm². Rating [B].
- [B] [S19] ECSS Q-70 cleanliness and contamination control training course, October 2023, definition of contaminant quoted from ECSS-Q-ST-70-01, list of ISO 14644-9 and 14644-10 as related standards. Rating [B].
- [B] [S20] Astro Pak, Cleanliness Verification Testing Services, astropak.com, 2024-11-18. Takes: instruments of the verification industry, levels reported to A/10 and A/20. Vendor page. Rating [B].
- [B] [S21] ISO 14644-10:2022, Cleanrooms and associated controlled environments, Part 10: Assessment of surface cleanliness for chemical contamination. Supersedes ISO 14644-10:2013. Read through the CEN adoption summary, EN ISO 14644-10:2022, standards.iteh.ai. Takes: ISO-SCC descriptor, grade N as log10 of g/m², thirteen grades 0 to minus 12, increments of 0.1. Rating [B], standard not read.
- [A] [S22] ISO 14644-9:2022, Cleanrooms and associated controlled environments, Part 9: Assessment of surface cleanliness for particle concentration, iso.org/standard/76889. Sample pages of ISO 14644-9:2012 and 2022, standards.iteh.ai. Takes: SCP grades 1 to 8, particles 0.05 to 500 µm, N as exponent of base 10 of the concentration per m² at the 1 µm reference. Rating [A] for scope, [B] for the grade formula.
- [A] [S23] ISO 14644-13:2026, Cleanrooms and associated controlled environments, Part 13: Cleaning of surfaces to achieve defined levels of cleanliness in terms of particle and chemical concentration, edition 2, published 2026-02-25, replaces ISO 14644-13:2017, iso.org/standard/91614. Rating [A].
- [A] [S24] LIGO-E080177-v2, RGA Test Qualification of components for the LIGO UHV, dcc.ligo.org. Takes: rules a to e, base pressure at or below 1e-6 torr, no empty-chamber subtraction, argon and krypton calibration leak, masses 41, 43, 53, 55, 57 indicative of high molecular weight hydrocarbons. Rating [A]. — https://dcc.ligo.org/LIGO-E080177/public
- [A] [S25] US EPA, trichloroethylene risk management under TSCA: final rule 17 December 2024, majority of uses prohibited 15 September 2025. EPA, Interim Final Rule on Compliance Date Extensions, 12 September 2025 (nuclear fuel processing aid to 15 September 2028, wastewater disposal to 18 December 2026). Federal Register, Extension of Postponement of Effectiveness for Certain Provisions of Trichloroethylene, 90 FR 51027 (14 November 2025), 91 FR 7401 (18 February 2026), notice of 5 May 2026. Harvard EELP tracker, 17 February 2026, sixth postponement to 18 May 2026. Bergeson & Campbell, EPA's 2026 Unified Agenda Includes TSCA Rulemakings, 13 July 2026 (NPRM July 2026, final rule October 2026). Rating [A] for EPA and Federal Register, [B] for the law firm summary.
- [A] [S26] 3M, 3M to Exit PFAS Manufacturing by the End of 2025, press release, 20 December 2022, news.3m.com. Best Technology and Enviro Tech International distributor pages: Novec last order date 31 March 2025, manufacturing ceased end of 2025. Rating [A] for the press release, [B] for the distributor pages.
- [A] [S27] ITER Organization, The art and science of cleanliness in a fusion reactor, iter.org, 3 July 2024. Takes: continuous outgassing tests, L. Worth on fab materials and fusion materials. Rating [A]. — https://www.iter.org/node/20687/art-and-science-cleanliness-fusion-reactor
- [A] [S28] First results of a high sensitivity and transportable Ring Laser Gyroscope, arXiv:2606.02594 (June 2026). Takes: cleaning as a design requirement, explicit cleaning specifications and process records in procurement, hydrocarbon risk under RF plasma. Rating [A]. — https://arxiv.org/abs/2606.02594
- [A] [S29] Vacuum Acceptance Tests for Particle Accelerator Equipment, CERN Accelerator School, arXiv:2006.10367. Takes: all components UHV cleaned before assembly, acceptance criteria defined per machine. Rating [A]. — https://arxiv.org/abs/2006.10367
- [A] [S30] Surface Finishing and Coatings for Accelerator Vacuum Applications, arXiv:2506.23691 (2025). Takes: cleaning as the most basic and underestimated surface finishing, detergent bath or solvent, dry plasma rarely. Rating [A]. — https://arxiv.org/abs/2506.23691
Format of the drawer: label, full reference, what the essay takes from it, rating. Ratings on the Stack scale: [A] primary document read in this pass, [B] secondary source, vendor page or memory of a table, [C] derived from stated assumptions. Access date 2026-08-23 unless stated.
---
# The Warhead Dividend
> Helium-3 was born in warheads, spent on borders, and is now hoarded for qubits. A field note on the strangest supply chain in cryogenics, with four dated bets.
Date: 2026-08-09
Canonical: https://pierreribault.com/essays/the-warhead-dividend/
import ProsePlate from '../../components/ProsePlate.astro';
import fig1Map from '../../assets/essays/warhead-dividend-fig1-map.jpg';
import fig2SupplyDemand from '../../assets/essays/warhead-dividend-fig2-supply-demand.jpg';
import fig3Price from '../../assets/essays/warhead-dividend-fig3-price.jpg';
import fig4HelionLoop from '../../assets/essays/warhead-dividend-fig4-helion-loop.jpg';
import fig5Migrations from '../../assets/essays/warhead-dividend-fig5-migrations.jpg';
The coldest object in most buildings that own one hangs from the ceiling like a chandelier: gold-plated discs stacked by diameter, coaxial lines falling between them like vines, and at the bottom, colder than interstellar space, a stage a few centimeters across where the qubits live. The working fluid of that last stage is helium-3. A large dilution refrigerator carries a charge of roughly forty liters of the gas. At standard conditions that is about five grams of matter, the mass of a sheet of paper, and at current purified prices it is a six-figure sum. The machine circulates the same charge, year after year, through a sealed loop of pumps, traps, and heat exchangers.
Every gram of it began life inside a nuclear weapon.
Modern warheads are boosted with tritium, and tritium decays with a half-life of 12.3 years, so the reservoirs come back on a schedule to be purged and refilled. The gas skimmed off during that maintenance, at the Savannah River Site, is decayed tritium: helium-3. It is the only domestic source the United States has ever had, auctioned in small lots through the federal isotope program. The deterrent pays out a dividend, a few thousand liters a year, and the strangest supply chain in cryogenics starts there.
A word on why this essay lives on this site. The vacuum hardware stays in the background this time, and I want to name that rather than smuggle it past you. But helium is my trade's own gas: every weld I have ever qualified was sprayed with it, every leak I have ever found announced itself as mass 3 or 4 on a spectrum. The machines that drink this isotope are cryogenic ultrahigh-vacuum systems built by people I would recognize at a conference. And the story that follows is, from its first liter to its last, a story about closed systems. Those are the only stories I know how to read properly, so I am going to read this one.
## the gas nobody makes
Nobody manufactures helium-3. It is harvested, and the harvest is small. On the supply side of the ledger there are exactly three working entries. The American one: warhead maintenance at Savannah River, federal draw now held under six thousand liters a year after the rationing reforms. The Canadian one: Ontario's CANDU reactors produce tritium as a byproduct of heavy-water operation, and Darlington holds one of the largest civilian tritium inventories on Earth; Laurentis, an Ontario Power Generation subsidiary, extracts the helium-3 that accumulates in the storage containers, and Air Liquide distributes it. The Russian one: real, and since 2022 effectively closed to Western buyers.
Add it up and the world produces somewhere between twenty-two and forty thousand liters a year, depending on whose survey you trust, against demand estimates of forty to sixty thousand. The American stockpile sat near ninety thousand liters at last public count, roughly double its level at the bottom of the 2010 crisis. The supply curve is a radioactive decay law applied to a warhead inventory whose size is set by arms-control treaties. Disarmament is the upstream industry. Every liter in every fridge is the deterrent's exhaust, metered by a half-life nobody can negotiate with.
## the year the gas ran out
We have run this experiment before. After 2001, the United States decided to scan cargo at its borders for smuggled nuclear material, and the neutron detector of choice was a helium-3 proportional counter. The build-out consumed the gas at a rate that peaked near seventy-nine thousand liters in a single year, quietly emptied most of the strategic stockpile, and forced a secret interagency rationing scheme before Congress ever heard the isotope's name. The GAO wrote the postmortem in 2011.
No new source appeared. The largest helium-3 consumer on Earth redesigned its detectors around boron-10 lined tubes and stopped needing the gas at all. Federal demand collapsed by an order of magnitude and the stockpile began, slowly, to refill.
So the recorded history of helium-3 has a shape, and it is worth stating as a law: demand has always yielded first. Supply has never once risen to meet it.
## arithmetic at the bottom of the tank
The forecasts now circulating for the quantum era do not survive their own units, and the failure is instructive. Helium-3 quantities arrive in liters of gas, liters of liquid, grams, and kilograms, often in the same paragraph, and the conversions are brutal: one liter of gas at standard conditions weighs 0.13 grams, so a kilogram is about seven and a half thousand liters. The headline figures of three to four hundred kilograms of annual quantum demand convert to two or three million liters, which at roughly forty liters per large machine means tens of thousands of new dilution refrigerators built every single year. The market's own per-machine gram counts put it near eighty thousand. The installed base, accumulated over two decades, is a few thousand; the largest manufacturer claims some eighteen hundred systems deployed, total.
The price record tells the same story in dollars. Federal auctions in the 2000s cleared at forty to eighty-five dollars a liter. The bulk market of the 2020s runs one thousand nine hundred to two thousand six hundred. Purified to dilution grade, above ten thousand, at times nearer twenty. Three decades of price climbed in twenty years. And here is my favorite line item: the lunar-mining pitch quotes helium-3 at twenty million dollars a kilogram. Divide by seven and a half thousand liters and the moon, priced honestly, lands at about two thousand seven hundred dollars a liter, squarely on the bulk market it claims to disrupt.
## stock, not flow
There is a deeper error under the arithmetic, and it takes a practitioner to feel it in the hands. A dilution refrigerator does not consume helium-3. It seals it. The charge is an inventory, circulating in a closed gas-handling system where recovery is a religion and venting a six-figure accident. Losses exist, through permeation, through cold-trap maintenance, through the occasional human mistake, and they are small in a well-run circuit. Demand from quantum computing is therefore a stock: filled once per machine, topped up against losses, returned when the machine retires. The forecasts sell an annual flow. The machine fills a stock, once.
That distinction rewrites the market. Flow demand scales with computation performed; stock demand scales with machines built, and machines are built slowly. It also means the existing fleet is a reservoir: eighteen hundred fridges at up to forty liters apiece is on the order of seventy thousand liters already above ground, sealed, recoverable, and eventually resalable. Anyone who has chased leaks for a living reads the forecasts and sees the confusion at a glance, because our whole trade is the discipline of keeping a stock from becoming a flow.
## three claimants to the first fresh liter
Which brings us to the frontier, because for the first time in the isotope's history, three different industries claim they are about to make new helium-3 rather than harvest old bombs.
The loudest claimant is the Moon. Interlune, a Seattle company founded by lunar-industry veterans, has assembled about half a billion dollars in agreements, including a deal with the largest fridge maker for up to ten thousand liters a year across 2028 to 2037, and a camera riding to the lunar surface late this year to measure regolith concentrations, ahead of an excavation payload NASA has funded toward readiness in 2027. Interlune is no longer alone: in early August a second lunar company, Black Moon Energy, signed a prospective helium-3 supply agreement with the fusion firm TAE Technologies, a robotic survey mission promised within five years. I will offer a dated bet on all of it below. But the detail worth the entire press kit sits in their own disclosures: the first deliveries are set to come from terrestrial sources. Interlune's Cold Capture program distills helium-3 out of ordinary Grade-A helium at liquefaction plants, hit ninety-nine percent purity this summer, and carries a defense contract to industrialize. The company sells the Moon and ships from a distillery, and the distillery is the better business. Prospectors have taken the hint: a Minnesota helium well spent the winter reporting the richest helium-3 concentration yet published for a terrestrial gas reservoir, about a dozen parts per billion, federally verified, and still, at heart, a feedstock for the same distillation.
The quietest claimant is fusion. Helion's reactors are designed to burn deuterium and helium-3, a reaction whose products are all charged, which is what lets them pull electricity straight off the expanding plasma instead of boiling water. The fuel problem solves itself in the loop: the deuterium-deuterium side reactions breed helium-3, directly and through tritium that decays into more of it, and the company recovers and purifies the isotope in-house. Their founder has said they make it efficiently enough to consider selling the surplus. Their first commercial plant is under construction against a 2028 delivery contract. If the loop closes, fusion becomes a net source of helium-3 before it becomes a net source of electrons.
And the third claimant is absence. A Munich company sells millikelvin cooling with no helium-3 at all, adiabatic demagnetization instead of dilution, and openly doubts that an industry should be built on this gas. If quantum hardware ever standardizes around such platforms, the story ends the way 2010 ended: the detector redesigned, the gas unneeded. The law above says to take that possibility more seriously than any supply forecast.
## the sting
One more thing happened while I was writing this. In March, strikes on Qatar's Ras Laffan complex put roughly a third of the world's helium supply under force majeure, with repairs measured in years; Russia, the swing producer, has its helium under export controls into late 2027; spot prices doubled. Ordinary helium-4, the abundant isotope, the gas we spray on welds without thinking, became the most contested molecule in the industrial gas trade. Now look one paragraph up: the most credible new source of helium-3 is a distillation column bolted to the world's helium-4 liquefaction stream. The rare isotope has moored itself to the abundant one at the precise moment the abundant one stopped being abundant. Nothing in this supply chain stands on its own floor.
## coda
The liter of gas in the chandelier has lived four lives, or will have soon. In 1965 it was tritium in a warhead reservoir, waiting for a war that never came. In 2010 its siblings were spent at the borders, sniffing cargo for the same plutonium the warheads carried. Today it circulates below ten millikelvin, sealed, hoarded, priced like treasure, holding a quantum processor below the temperature of deep space. And by the early 2030s, if one company in Everett is right, the machine that burns it will breed it, and the strangest supply chain in cryogenics will quietly close its own loop. Born of disarmament, spent on borders, hoarded for qubits, and finally home-made. It ships in a lecture bottle you could close one hand around: the charge of one fridge, the price of a house.
## points of rendezvous
Dates this essay can be checked against:
- **Late 2026.** Interlune's Crescent Moon camera flies on Astrobotic's Griffin-1 lander.
- **Fall 2027.** The Prospect Moon excavation payload is due ready for lander integration.
- **End 2027.** Russia's helium export controls expire, or are renewed.
- **2028.** Helion's Orion plant faces its Microsoft delivery deadline; first Interlune deliveries to Bluefors come due.
- **2029 to 2031.** Ras Laffan's helium capacity is due back, on the announced repair horizon.
## four bets
Falsifiable, dated, and signed:
- **B1.** On January 1, 2030, more than ninety percent of commercially sold helium-3 will still trace to tritium decay, arsenals plus CANDU, and zero lunar liters will have been delivered to any customer.
- **B2.** The first new source to ship meaningful volume, say a thousand liters in a year, will be terrestrial distillation of the helium stream. The Moon will not be first, on any timeline I would sign.
- **B3.** If Orion runs on deuterium and helium-3 by its 2028 commitment, Helion will sell surplus helium-3 commercially before it has sold a full year of contracted electricity.
- **B4.** Real quantum-driven demand stays below ten thousand liters a year through 2030, an order of magnitude under the headline forecasts, because the machine fills a stock and the forecasts price a flow.
## the six taps
| tap | mechanism | status 2026 | scale |
|---|---|---|---|
| Savannah River, US | warhead tritium decay | operating, sole US source | few kL/yr |
| Darlington, Canada | CANDU tritium stores, via Laurentis and Air Liquide | operating | few kL/yr |
| Russia | tritium decay, Rosatom | closed to Western buyers since 2022 | opaque |
| Terrestrial distillation | He-3 skimmed from Grade-A helium (Interlune Cold Capture) | 99 % purity shown, industrializing | claimed: could triple US supply |
| Lunar regolith | solar-wind He-3 mined from the Moon | camera flies late 2026 | promised, 2028 onward |
| Fusion breeding | D-D side reactions in-loop (Helion) | Polaris operating, Orion building | surplus offered, unquantified |
## Sources
- [A] Helium-3 as a warhead byproduct, Savannah River the sole US source, and the federal draw held under 6,000 liters a year after the mitigation reforms: DOE Isotope Program, National Isotope Development Center, Supply and Demand of Helium-3. — https://www.isotopes.gov/Supply-and-Demand-of-Helium-3
- [A] The 2008 to 2011 shortage and the boron-10 detector redesign that ended it: US GAO, Neutron Detectors, Alternatives to Using Helium-3, GAO-11-753 (September 2011). — https://www.gao.gov/products/gao-11-753
- [A] The DOE isotope-management postmortem, the historical auction price of 40 to 85 dollars a liter, and the allocation and pricing tables: US GAO, Managing Critical Isotopes, GAO-11-472 (May 2011). — https://www.gao.gov/products/gao-11-472
- [A] Demand history from about 8,000 liters a year before 2001 to about 80,000 liters in 2008, the stockpile falling from 235,000 liters (2001) to 50,000 liters (2010), Ontario Power Generation tritium yielding roughly 130,000 liters of helium-3 over ten years, and the near-lossless reuse of a charge: Shea and Morgan, The Helium-3 Shortage: Supply, Demand, and Options for Congress, CRS R41419 (October 2011). — https://www.everycrsreport.com/reports/R41419.html
- [B] The US stockpile near 90,000 liters per a 2021 DOE document, and Regnat of kiutra on dilution as an industrial base: Savitsky, As helium-3 runs scarce, researchers seek new ways to chill quantum computers, Science (AAAS). — https://www.science.org/content/article/helium-3-runs-scarce-researchers-seek-new-ways-chill-quantum-computers
- [B] Darlington tritium stores, Laurentis extraction, and Air Liquide purification and distribution as the first non-military source: Laurentis Energy Partners and Air Liquide press releases (December 2021). — https://www.laurentisenergy.com/releases/laurentis-energy-partners-signs-long-term-commercial-agreement-with-air-liquide-for-the-distribution-of-helium-3/
- [B] The XLD1000sl helium-3 charge at an optimal ratio of 40 liters helium-3 to 180 liters helium-4, and a dilution-refrigerator base temperature below 10 mK: Bluefors, Dilution Refrigerator Measurement Systems product documentation. — https://bluefors.com/products/dilution-refrigerator-measurement-systems/
- [C] The bulk market at 1,900 to 2,600 dollars a liter, purified dilution grade above 10,000, the roughly six-figure charge, the unit conversions, an installed base above 1,800 systems, and a structural deficit of 20,000 to 40,000 liters a year: PostQuantum, Helium-3 in Quantum Computing, and Quantum Cryogenic Infrastructure and Helium-3 (May 2026). — https://postquantum.com/quantum-systems-integration/helium-3-quantum-computing/
- [C] The roughly 79,000-liter peak-year framing, the demand-side reading of GAO-11-753, and world production near 40,000 liters a year from USGS-derived figures: PostQuantum, Helium-3 Alternatives for Quantum Computing, Assessed (June 2026). — https://postquantum.com/quantum-computing/helium-3-alternatives/
- [B] Interlune Cold Capture at 99 percent purity from Grade-A helium, an AFWERX SBIR Direct-to-Phase-II award (November 2025), nearly 500 million dollars in binding purchase agreements led by Maybell Quantum and Bluefors, and the claim of tripling US supply: Interlune press release (July 20, 2026). — https://www.prnewswire.com/news-releases/interlune-produces-pure-helium-3-from-domestic-helium-using-novel-cryogenic-technology-302829735.html
- [B] The Bluefors purchase of up to 10,000 liters a year, with deliveries across 2028 to 2037: Bluefors and Interlune press release (September 16, 2025). — https://www.prnewswire.com/news-releases/bluefors-to-source-helium-3-from-the-moon-with-interlune-to-power-next-phase-of-quantum-industry-growth-302557532.html
- [B] The NASA SBIR Phase III award of 6.9 million dollars for the Prospect Moon extraction payload, the Crescent Moon camera on Astrolab's FLIP rover aboard Astrobotic's Griffin-1 in late 2026, roughly 500 million dollars in contracts including DOE, and first shipments likely terrestrial via Cold Capture: SpaceNews (May 2026) and GeekWire (July 2026). — https://spacenews.com/interlune-wins-nasa-contract-for-helium-3-extraction-payload/
- [B] The Helion fuel loop breeding helium-3 from deuterium side reactions and reusing it in-house at high efficiency, Kirtley open to selling surplus, Polaris running deuterium-tritium at 150 million degrees, and Orion at 50 megawatts under construction for the 2028 Microsoft commitment: TechCrunch (February 13, 2026) and POWER Magazine (February 2026). — https://techcrunch.com/2026/02/13/fusion-startup-helion-hits-blistering-temps-as-it-races-toward-2028-deadline/
- [B] The March 2026 Ras Laffan strikes removing roughly a third of world helium supply under force majeure, a three-to-five-year repair horizon, Russian helium export controls running to end-2027, and doubled spot prices: gasworld (April 2026), WestAir (June 2026), and The Moscow Times (April 14, 2026). — https://www.gasworld.com/story/russia-places-temporary-export-controls-on-helium/2246455.article/
- [B] The TAE and Black Moon Energy prospective helium-3 supply and commercialization agreement, and BMEC's robotic lunar delineation mission within five years: TAE Technologies press release (August 5, 2026). — https://tae.com/tae-technologies-and-black-moon-energy-corporation-enter-into-strategic-helium-3-supply-and-commercialization-agreement-for-fusion-power/
- [B] Helium-3 at 11.2 to 11.9 parts per billion in raw gas at the Topaz project, Minnesota, verified by the USGS Noble Gas Laboratory and Lawrence Livermore National Laboratory: Pulsar Helium news release (January 19, 2026). — https://s203.q4cdn.com/212931576/files/doc_news/U-S--Federal-Government-Laboratories-Verify-Helium-3-Results-at-Pulsar-Heliums-Topaz-Project-in-the-USA-2026.pdf
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary. Links verified August 2026.
---
# The Banned Shelf
> For a century, the vacuum tube could order any material it wanted. The shelf is closing now, and the frontier has just walked back in to place the largest order since radar.
Date: 2026-07-30
Canonical: https://pierreribault.com/essays/the-banned-shelf/
import PlateTimeline from '../../components/PlateTimeline.astro';
import PlatePowerFreq from '../../components/PlatePowerFreq.astro';
Take the parts list of a radar magnetron, any of them, drawn around 1955. Read it against the three shelves of The Matter, workhorse, banned, future, and stamp each line with where its material stands in 2026.
Anode block, oxygen-free copper: workhorse. Cathode, barium and strontium oxides on a nickel sleeve: workhorse, with an asterisk this essay will spend. Heater, tungsten: workhorse, export-licensed since February 2025. End shields, molybdenum: workhorse, licensed under the same decree. Magnet, Alnico, cobalt-rich: workhorse, the cobalt carrying a carcinogen classification since 2021. Seal rings, Kovar, seventeen percent cobalt: same element, same file. Braze on the cooling fittings, silver alloy, cadmium-bearing: banned since 2011. Process sheet, step one, vapor degrease in trichloroethylene: authorization holders only since 2016. Step four, chromic bright dip: authorization holders only since 2017.
I have held drawings like this at the bench, in a French tube house older than the transistor, and read the margin notes the way you read the menu of a restaurant that has closed. The list survives. The shelf under it has moved.
---
The functions of a tube have held still since Fleming: emit electrons, hold vacuum and join dissimilar parts, let the power out through a window, dump the heat that remains. Everything else rotated, and for a hundred years the rotation ran in one direction: more elements, more recipes, more of the periodic table conscripted each decade.
Emission first. Wehnelt showed in 1904 that barium and strontium oxides emit at a fraction of tungsten's temperature. Coolidge made tungsten ductile in 1909, and by 1911 the industry had abandoned every rival filament process. Thoriated tungsten entered production tubes in 1925, with emission up to fifty times that of the bare metal. The 1950s brought the dispenser cathode, barium chemistry stored in a porous tungsten plug, and its M-type refinement still wears a sputtered film of osmium and ruthenium: a coating drawn from a metal produced at well under a tonne a year, worldwide, applied to the most consumable component in the device.
Joining next. Glass met metal through Dumet wire, then through Kovar in the 1930s, an iron-nickel alloy holding seventeen percent cobalt so its expansion tracks hard glass. The 1950s added molybdenum-manganese metallization, which let alumina take over from glass as the envelope, and silver-copper brazes to close the joints. Windows ran from mica to glass to beryllium for X-rays, to beryllium oxide where the heat climbed, to sapphire, and finally to diamond. Heat went into oxygen-free copper, pyrolytic graphite, tungsten-copper.
None of it was rationed. Kohl's 1967 handbook and Rosebury's 1965 shop manual read today like the inventory of an unlocked building: thorium in the cathode chapter, mercury in the pumps, cadmium at the braze bench, trichloroethylene in every cleaning sequence, chromic acid at the bright dip. Rosebury alone lists page after page of preparations that a European shop can no longer run in 2026 without an authorization file. Both books were written as references. They have aged into records.
---
Regulation closed the shelf first, and it closed it politely. Cadmium left the braze bench in 2011, when Europe capped it at a hundredth of a percent in filler metals. Trichloroethylene reached its sunset in April 2016; a solvent that moved fifty thousand tonnes a year through European industry in 2010 moves about twelve hundred today, a collapse of ninety-five percent inside a decade, and the survivors operate under individual authorizations with expiry dates. Chromic acid followed in September 2017. Cobalt, seventeen percent of every Kovar ring, has carried a European classification as a category 1B carcinogen since 2021: the sealed part is untouched, the powder, the grinding dust and the braze fume are another matter, and the exposure limits keep descending. Beryllium stayed legal and lost a factor of ten off its American workplace limit in 2017. Thorium retreats under radioactive-materials rules. Lead survives in tube glass on a lease: the RoHS exemption that covers it was restructured in 2025 and now expires in mid-2027 unless renewed. Regulation retires one shelf item at a time, and every single retirement looks survivable. The sum is a different shop.
Supply closed faster, and with dates attached. On December 3, 2024, China prohibited, in principle, the export of gallium, germanium, antimony and superhard materials to the United States, and barred dual-use items to American military users outright. Eleven months later the materials clause was suspended, as part of a trade truce, until November 27, 2026. The military clause stayed. Access now comes with a printed expiration, like the lease on the lead glass. On February 4, 2025, a second decree placed tungsten, tellurium, bismuth, molybdenum and indium under export licenses: the two framework metals of the tube, the heater wire and the end shield, in one document. The regime held through the truce and tightened in January 2026, when tungsten exports were centralized on fifteen designated companies and ammonium paratungstate crossed eleven hundred dollars per metric ton unit, roughly triple its level two years earlier. The United States has mined no tungsten since 2015; China holds about eighty percent of world supply. Behind the framework metals stand the trace ones. Rhenium, the heater alloy's stiffener, amounts to eighty-one tonnes a year worldwide, and superalloy turbine blades take eighty percent of it before a tube shop sees a spool. Osmium, sputtered onto every M-type cathode, is produced at well under a tonne a year. Scandium, carrier of the scandate cathode that has been the future for fifty years, runs near forty tonnes a year with no primary mine on Earth and a license regime on the largest producer.
The third closure makes no announcements. In 2023, a United States Senate bill proposed restricting traveling-wave-tube sourcing to the national technology and industrial base; the provision was not adopted, and the fact that someone drafted it says enough about how thin the base had become. Space-qualified traveling-wave tubes come, for the most part, from two Western houses. Below the companies sit the hands. Ceramic-to-metal brazing, cathode impregnation, glass lathe work: these are taught at the bench, over years, by people who are retiring faster than they are being replaced. I have stood in shops where one person held the whole braze schedule in their head, and I have watched such a person leave. No decree records this kind of closure. The parking lot does.
And yet the future shelf is filling, mostly with returns. Diamond entered a traveling-wave tube in 1988, natural stones standing off a 28 GHz circuit in a NASA program; the device worked and was never replicated, because the stones cost too much. Aluminum nitride now sits where beryllium oxide sat, carrying the heat without the toxicity file. C-103, the niobium alloy of rocket nozzles, arrives on the vacuum bench. Oxygen-free copper returns as powder, printed into monolithic bodies. The future shelf is mostly the past, coming back under a cleaner name.
---
The customers are coming back, and they are the largest since radar. ITER's heating system calls for gyrotrons at 170 GHz delivering a megawatt each, continuously, through diamond; the machines being drawn behind it ask for multi-frequency sources at 136, 170, 204 and 238 GHz. Private fusion companies are ordering from the same short list of gyrotron houses. EUV lithography runs thirty kilowatts of CO2 laser light through diamond windows on its way to a tin plasma, and every advanced fab on Earth sits downstream of that path. Satellite constellations still fly traveling-wave tubes where efficiency and radiation tolerance rule. The terahertz band waits for vacuum devices because solid state runs out of watts there. Accelerators never left.
All of them arrive at the shelf described above, the one with lease dates printed on it. So the design logic inverts. For a century, a tube engineer chose the best material and the buyer found a way to pay. The coming decades will run the sequence in the other order: first what is legal, then what is purchasable, then what performs. Call it availability-first design. Three bets follow from it, each with a condition that can prove me wrong.
First, cathodes. The first mainstream tube family to abandon the thermionic dispenser will do it over barium supply and compliance, with lifetime parity still unproven. Watch the launch documents: if the first series tube with a cold or photoemissive source justifies itself on performance while the barium chain stands intact, I lose this one. Horizon 2040.
Second, joints. Printed monolithic bodies will remove most brazed joints from production tubes by 2040; the braze survives at windows and feedthroughs, and the guild shrinks exactly when fusion's balance of plant starts ordering feedthroughs by the thousand. Count the joints per tube generation in the teardowns and the patents: if the count holds flat through 2040, I lose this one too.
Third, windows. Diamond disk capacity becomes the schedule-limiting item of at least one major electron-cyclotron heating program before 2035. The falsifier is public: a program schedule that slips, with the window supply named, before that date. Absent one, the bet fails.
---
One more name is moving toward the shelf, and it is already in the roughing pump. The universal PFAS restriction now finishing its scientific evaluation at ECHA covers more than ten thousand substances, fluoropolymers included: the perfluoropolyether oil in the pump, the FKM elastomer in every centering ring. The committees hand their opinions to the Commission around the turn of the year; a decision comes in 2027 at the earliest, application no sooner than 2029, and the industry's defense has taken the shape of the era: seventy-four derogations, dated. The next resident of the banned shelf will arrive with its lease already signed.
---
At the top of a fusion gyrotron sits a disc of chemical-vapor-deposited diamond, 119 millimeters across, 2.22 millimeters thick, about 87 grams. It is grown over weeks in a plasma reactor, and a single disc runs into six figures, which puts the gram somewhere an order of magnitude above gold. Match its price in gold and the pile is smaller than you might expect: gold is five and a half times denser than diamond. Every megawatt of electron-cyclotron heating on ITER will pass through carbon like this, because no other window material holds at that power. Ninety grams of carbon now hold the door between a burning plasma and the room.
## Sources
- [A] Commission Regulation (EU) No 494/2011, amending REACH Annex XVII entry 23: cadmium restricted to 0.01 percent by weight in brazing fillers, in force late 2011. On the ECHA Authorisation List (REACH Annex XIV), trichloroethylene reached its sunset on 21 April 2016 (EU use fell from about 50,000 tonnes in 2010 to about 1,200, per an ECHA impact case study, 2022); chromium trioxide followed on 21 September 2017. — https://eur-lex.europa.eu/eli/reg/2011/494/oj
- [A] Commission Delegated Regulation (EU) 2020/217, the 14th ATP to CLP: cobalt metal classified Carc. 1B, Muta. 2 and Repr. 1B, applicable from 1 October 2021, generic concentration limit 0.1 percent. OSHA final rule, Occupational Exposure to Beryllium, 82 FR 2470, 9 January 2017: permissible exposure limit lowered tenfold to 0.2 micrograms per cubic metre over an 8 hour TWA. — https://eur-lex.europa.eu/eli/reg_del/2020/217/oj
- [A] RoHS Directive 2011/65/EU, Annex III exemption 7(c) (lead in glass and ceramic of electrical and electronic components), restructured by Commission Delegated Directive (EU) 2025/2363 (published 21 November 2025), with revised sub exemptions and expiry dates in 2027.
- [A] MOFCOM Announcement No. 46 of 2024 (3 December 2024): export ban on gallium, germanium, antimony and superhard materials to the United States, and a bar on dual use exports to US military end users; the materials clause was suspended on 9 November 2025 until 27 November 2026, while the military clause remained. MOFCOM Announcement No. 10 of 2025 (4 February 2025): export licensing on tungsten, tellurium, bismuth, molybdenum and indium. In January 2026, tungsten exports were centralised on fifteen designated companies. — https://www.csis.org/analysis/china-imposes-its-most-stringent-critical-minerals-export-restrictions-yet-amidst
- [C] Exiger, Critical Minerals Export Controls: a running tracker of the 2024 to 2026 Chinese measures and their scope. — https://www.exiger.com/perspectives/critical-minerals-export-controls/
- [A] USGS Mineral Commodity Summaries, tungsten: no US mine production since 2015, China near 80 percent of world supply. Trade press, January 2026: ammonium paratungstate above 1,100 dollars per metric ton unit, near triple its level two years earlier. — https://pubs.usgs.gov/periodicals/mcs2026/
- [A] USGS Mineral Commodity Summaries: rhenium world production 81 tonnes in 2025, a by product of porphyry copper molybdenum roasting, about 80 percent taken by superalloys; osmium reported only within the platinum group, world output well under one tonne a year; scandium near 40 tonnes a year, exclusively a by product, with no primary mine in operation. — https://pubs.usgs.gov/periodicals/mcs2026/
- [A] A. Wehnelt reported oxide coated cathodes in 1904 (Annalen der Physik, 4th series, volume 14). W. D. Coolidge made tungsten ductile in 1909, rival filament processes abandoned by 1911, and thoriated tungsten entered production tubes in 1925 (Engineering and Technology History Wiki, Ductile Tungsten). Shop references: W. H. Kohl, Handbook of Materials and Techniques for Vacuum Devices (Reinhold, 1967); F. Rosebury, Handbook of Electron Tube and Vacuum Techniques (Addison Wesley, 1965, AIP reprint 1993).
- [A] Congressional Research Service, Insight IN12221, on FY2024 NDAA defense industrial base provisions: S.2226 Section 835, restricting traveling wave tube sourcing to the national technology and industrial base, was not adopted. — https://crsreports.congress.gov/product/pdf/IN/IN12221
- [A] US patent 6,917,162 (citing NASA CR-182183): a 28 GHz TunneLadder traveling wave tube built with natural diamond supports in 1988, not replicated on cost grounds. — https://patents.google.com/patent/US6917162
- [A] ITER class electron cyclotron heating uses gyrotrons at 170 GHz, one megawatt continuous, through CVD diamond windows; multi frequency concepts at 136, 170, 204 and 238 GHz are reviewed for machines beyond ITER (EUROfusion DEMO gyrotron review, Fusion Engineering and Design, 2017). Diamond output windows carry megawatt power where sapphire limited near 150 kW, and serve the CO2 laser chain of EUV lithography (Calabazas Creek Research, DOE OSTI 1866817; Element Six). — https://www.sciencedirect.com/science/article/pii/S0920379617300686
- [A] Karlsruhe Institute of Technology, step-tunable megawatt gyrotron pre-prototype for DEMO with a diamond Brewster window. — https://library.oapen.org/handle/20.500.12657/79401
- [A] ECHA universal PFAS restriction proposal, submitted 13 January 2023, covering more than 10,000 substances with fluoropolymers in scope; the updated proposal of 20 August 2025 raised derogations from 26 to 74. RAC opinion March 2026; SEAC draft opinion 26 March 2026 (consultation closed 25 May 2026); opinions to the Commission near end 2026; application not before 2029. — https://echa.europa.eu/hot-topics/perfluoroalkyl-chemicals-pfas
- [A] The gyrotron window disc, 119 mm across and 2.22 mm thick, computes to near 87 grams at 3.52 grams per cubic centimetre. A diamond output window runs near 50,000 dollars per disc for direct coupled gyrotrons (Calabazas Creek Research, DOE OSTI 1866817). The disc sits between roughly 1,100 and 2,300 dollars per gram, an order of magnitude above gold, which was near 131 dollars per gram at end July 2026, spot. — https://www.osti.gov/biblio/1866817
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary. Links verified 2026-07-30.
---
# The Floor Under the Fab
> The chip war is fought over things everyone can name. Under all of them sits one layer nobody does.
Date: 2026-07-13
Canonical: https://pierreribault.com/essays/the-vacuum-lock/
Everyone has a map of the chip war now.
The map has famous places on it. EUV. ASML. Taiwan. The 3-nanometer node. Export controls. The names are public, the stakes are public, the fight is on the front page. People who have never seen a wafer can tell you that one Dutch company makes the machine that prints the smallest chips, and that the United States is trying to keep that machine out of China.
That map is real. It is also incomplete. Under every place on it sits a layer that almost no one names.
Vacuum.
Not the household word. The discipline. Pumping a chamber down through more than ten orders of magnitude of pressure, holding it there, keeping it clean, for years. A chip is not made in air. A chip is made in near-emptiness.
## The process is the vacuum
Walk the front end of a fab and count the steps that need a chamber pumped down. Almost all of them.
Deposition lays the thin films. Physical vapor deposition vaporizes a source inside a vacuum chamber and lets it condense on the wafer. Chemical vapor deposition feeds reactive gas at low pressure. Etching removes material with plasma, which only forms below atmospheric pressure. Ion implantation fires dopants down a vacuum beamline. Each layer of a modern chip is built and cut and doped this way, and a complex CMOS wafer runs the patterning cycle dozens of times before it is done.
- **[B]** HORIBA, semiconductor lithography process (a CMOS wafer can run the photolithographic cycle as many as 50 times), https://www.horiba.com/int/semiconductor/process/lithography/
Then there is the machine on everyone's map. EUV. The light it uses is absorbed by ordinary matter, so the whole instrument has to live in vacuum. ASML says it plainly.
> EUV light is absorbed by everything, even air. So the whole light path and everything the light interacts with, from source to wafer, must be in high vacuum.
- **[A]** ASML, EUV lithography systems, https://www.asml.com/en/products/euv-lithography-systems
Inside that source chamber, a high flow of hydrogen protects the collector mirror, and hydrogen is one of the hardest gases to pump. The pumps that do it are a product in their own right.
- **[B]** Pfeiffer Vacuum, vacuum solutions for EUV (high-flow dry pumps for hydrogen), https://www.pfeiffer-vacuum.com/global/en/markets/semiconductor/lithography/euv-lithography
So the headline is true but shallow. The bottleneck is not only the light. The bottleneck is the emptiness the light needs. No vacuum, no chip.
## Who makes the pumps
If vacuum is the floor under the fab, the next question is simple. Who builds the floor.
The market estimates disagree, and they disagree by a lot. One 2025 survey puts the top six pump makers at about 62 percent of the semiconductor vacuum pump market. Another puts a top five at 45 to 50 percent. A broader vacuum survey from 2023 puts a leading group at 25 to 30 percent. Three reports, three numbers. All of them are commercial, paywalled, and quiet about method.
- **[C]** Dataintelo, semiconductor vacuum pump market (top six ≈ 62%, 2025), https://dataintelo.com/report/semiconductor-vacuum-pump-market
- **[C]** Mordor Intelligence, vacuum pump market (top five ≈ 45–50%, 2025), https://www.mordorintelligence.com/industry-reports/vacuum-pump-market
- **[C]** GMInsights, vacuum pump market (leading group ≈ 25–30%, 2023), https://www.gminsights.com/industry-analysis/vacuum-pump-market
Do not lean on the percentage. Lean on the names. Because the names do not move.
Pull any of those rankings and the top of the list is the same short roster. Edwards, the dominant semiconductor vacuum brand, owned by Atlas Copco of Sweden, built in the UK. Pfeiffer, German, now inside Busch, German. Leybold, German. Ebara, Japanese. ULVAC, Japanese. Kashiyama, Osaka Vacuum, Canon Anelva, Japanese. The disagreement is about how to slice the share. The agreement is about the geography.
The ownership is on the record, in the companies' own filings.
- **[A]** Atlas Copco (Stockholm, Sweden), completion of the Edwards acquisition (Edwards headquartered in the UK), https://www.atlascopcogroup.com/en/media/press-releases/2014/140109-edwards
- **[A]** Pfeiffer Vacuum / Busch Group (Pfeiffer based in Asslar, Germany; Busch Group headquartered in Germany's Black Forest; Busch holds a majority stake), https://group.pfeiffer-vacuum.com/en/news-media/press-releases/three-strong-brands-form-the-global-busch-group/
- **[A]** Ebara Corporation, corporate outline (headquartered in Tokyo, Japan; semiconductor vacuum pumps and abatement), https://www.ebara.co.jp/en/about/corporate/about/outline/index.html
Germany. Japan. The UK and Sweden. That is the top tier.
Look at what is not in the top tier. No American maker leads it. No Chinese maker is near it. No Russian maker exists in it. The Chinese names that do appear sit in the fragmented tail of every survey, the long list after the short one.
This is the part worth sitting with. The country running the chip war does not make the floor the fab stands on. It buys it from allies. And the country the chip war is aimed at cannot make it either.
## The control regime already knows
The export-control machinery has noticed, even if it never says the word.
When the United States moved to cut China off from advanced chipmaking, it did not write rules about "vacuum." It wrote rules about tools. The October 2023 controls name the equipment by function. Etch. Deposition. Lithography. Ion implantation. Annealing. Metrology and inspection. Cleaning.
- **[A]** US Bureau of Industry and Security, Export Controls on Semiconductor Manufacturing Items, 88 Fed. Reg. 73424 (Oct. 25, 2023), https://www.federalregister.gov/documents/2023/10/25/2023-23049/export-controls-on-semiconductor-manufacturing-items
- **[A]** US BIS press release on the rule, https://www.bis.gov/press-release/commerce-strengthens-export-controls-restrict-chinas-capability-produce-advanced-semiconductors-military
Read that list again with a vacuum scientist's eye. Etch is a vacuum process. Deposition is a vacuum process. Advanced lithography is a vacuum process. Ion implantation is a vacuum process. The control list is, in large part, a list of machines that cannot run without a pumped chamber. The state has drawn a fence around the vacuum frontier of the fab. It just labeled the fence by process, not by physics.
That is the tell. The strategic layer is being defended already. It is being defended without being named.
## What this movement does not settle
Here is the limit of everything above. All of it is visible.
A pump has a serial number. A tool has a customs code. A factory has an address you can put under sanction. The lock described here is the lock you can see, count, and ship. It is the lock a government can act on.
It is not the deepest one.
You can buy a pump in a year. You can sanction a pump in a day. What you cannot do quickly is rebuild the forty years of hands that know how to balance a rotor, coat a surface, and chase a leak no instrument has found yet. That knowledge has no serial number. It does not show up on the export list. It is the real lock, and it is the one nobody is counting.
That is the next movement.
Movement II is [The Names Nobody Cites](/essays/the-names-nobody-cites/). Underneath both, the [world map](/actors/world/), actor by actor.
---
*Movement I of an open series on the geopolitics of vacuum. Sources are rated [A] standard or primary, [B] manufacturer, [C] secondary, the same scale used across The Stack. No internal lab data is used. This is open-source analysis of public material, not guidance.*
## Sources
- [A] O'Hanlon, J.F., A User's Guide to Vacuum Technology, 3rd ed., Wiley, 2003.
- [B] May, G.S. & Sze, S.M., Fundamentals of Semiconductor Fabrication, Wiley (load-lock and contamination-control chapters).
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary. In-house derivations, anchors in O'Hanlon.
---
# The Names Nobody Cites
> Two hundred and thirty-five actors on one map: the consolidation sits one layer below the headline names, the leverage one layer above.
Date: 2026-07-07
Canonical: https://pierreribault.com/essays/the-names-nobody-cites/
import Spec from '../../components/Spec.astro';
import U from '../../components/U.astro';
I placed every company and laboratory that touches ultra-high vacuum onto a single map, two hundred and thirty-five of them, each pinned to the city where it sits. I did it to see the field I work inside from above, the way I watch a bake from a screen instead of standing at the chamber. The map made an argument on its own, before I had written a word beneath it.
The first thing it says, it says by where the pins are not. They crowd into three places, the American coasts, a band of small countries along the Rhine and the Alps, and a tight arc across Japan, Taiwan and Korea. Below the equator there is almost nothing. The emptiness that every frontier machine is built around, that fabs and colliders and fusion reactors depend on, is held up by three regions and left alone by the rest of the planet. The vacuum has three poles.
Then it says something quieter, which took the ownership data to see. The pumps that hold the world's vacuum carry famous names, and those names are owned by a few holding companies you would never stencil on a chamber. One Swedish group, Atlas Copco, controls Edwards in England, Leybold in Germany, and Gamma Vacuum and Montana Instruments in the United States, four of the names I trust on my own stations, flying flags of three nations, all steered from a fourth. Busch, a German group, holds Pfeiffer. The consolidation is real, but notice where it is not. Nobody has rolled up the headliners. ASML stands alone, TSMC stands alone, the names in every article about chips are not the ones being quietly gathered. The gathering is one layer down, on the ion pumps and the cryostats and the gauges, the parts nobody cites. That is where ownership is actually changing hands, and it changes hands in the dark, because no one is looking at the components.
A handful of these actors do not merely make a thing, they gate it. The list is short, and every entry is a single point where one failure, or one export order, stops the line.
Japan is worth stopping on, because the map corrects a lazy idea. Japan makes the tools that make the chips, and holds some of them alone. Lasertec, in Yokohama, is the only commercial source of actinic patterned-mask inspection for EUV, the machine that checks the mask before ASML's light ever prints it, a monopoly standing upstream of the monopoly. DISCO holds the dicing and grinding of finished wafers to nearly the same degree. The country that gets filed away as one fab among many is, in tooling, a lock.
The flag on the pin is rarely the flag that owns it. Edwards sits in England and answers to Sweden. Nor-Cal sits in California and answers to Germany. Cryomech sits in New York and answers to Finland. When I recolor the map by who controls each actor instead of by where it stands, the picture rearranges, and the rearrangement is the whole point. The map of where the work is done and the map of who owns the work are not the same map. Almost everyone reads the first one. The second is the one that decides what happens when a border closes.
And there is a last move, and the ground under it is still shifting. The hardware that holds the vacuum is made across a scatter of small countries, each a national monopoly in one narrow thing, Dutch light, German and Swedish pumps, Swiss valves, Italian getters, Austrian metals, French helium. That sovereignty is real, it is physical, and it is hard to move. But the value is leaving it. The layer that is starting to matter sits above the pump: the data coming off it, the predictive service, the telemetry, the model that warns a fab a pump will fail before it does. That layer is sold by the equipment vendors themselves, but it runs on American clouds, a dependence one layer further up. The countries that make the machine do not hold the layer that is learning to own the machine. Sovereignty over the metal is being quietly refinanced through the infrastructure that watches it.
The map is a small and honest instance of a general law, that the thing you manufacture and the thing that holds the leverage drift apart over time, and the leverage settles on whoever sits one layer above. I can read it cleanly in the vacuum chain only because I drew that chain by hand, node by node, and the ownership refused to collapse into a slogan. It collapsed into this instead. The famous names are a distraction. The chain is held beneath them, by owners flying flags that are not their own, and it is being refinanced from above, in a layer that flies just one.
So the pins glow where the work is done, and the work is real, and the names on it are the ones everyone knows. They are only not the names that own it, and less and less the names that will keep what it earns. The chips have their headlines. The vacuum is held, and quietly bought, by the names nobody cites.
Under this map sits [a floor of its own](/essays/the-vacuum-lock/), the load-bearing vacuum beneath the fabs. The map itself is the [world map](/actors/world/), actor by actor.
## Sources
- [B] Ownership and consolidation, drawn from the acquirers' own announcements: Atlas Copco's acquisitions of Edwards, Leybold, Gamma Vacuum and Montana Instruments; Busch Group's ownership of Pfeiffer Vacuum; Pfeiffer's acquisition of Nor-Cal Products; Bluefors' acquisition of Cryomech. Cross-checked against each company's corporate and investor pages.
- [B] Chokepoint positions, from company disclosures and standard semiconductor supply-chain reporting: ASML as sole EUV lithography supplier; Lasertec as the only commercial actinic patterned-mask inspection tool for EUV; VAT in ultra-high-vacuum gate valves; SAES in getters; Plansee in refractory metals; DISCO in wafer dicing and grinding; the large cryoplant duopoly of Air Liquide and Linde Kryotechnik, the Swiss cryoplant arm of Linde; TSMC at over 90 percent of leading-edge logic.
- [B] The value migrating to a data layer above the hardware: the predictive and telemetry services now sold by the equipment vendors themselves, including Edwards' Semiconductor Intelligent Service, Applied Materials' and Lam Research's equipment-analytics platforms, all running on external cloud infrastructure.
- [C] Actor positions, sectors and control coding: the Mean Free Path Actors map and its actor-control layer, a hand-built synthesis of the above, joined actor by actor.
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary.
---
# No Dashboard Pumps Faster
> Nine bake stations on one screen. What it produces is a record of how a healthy pumpdown behaves, and an intuition the next operator can inherit.
Date: 2026-06-29
Canonical: https://pierreribault.com/essays/no-dashboard-pumps-faster/
import Spec from '../../components/Spec.astro';
import U from '../../components/U.astro';
There is a screen in our bake area that I built. It shows nine stations at once: total pressures, wall temperatures, the slow climb of a residual gas spectrum as a chamber gives up its water. I wrote every line of it. On a good week it tells me at a glance the health of a fleet I once had to walk to, controller by controller.
And it does not make a single pump pump faster.
The trade keeps a bright promise about screens like mine. Connect the pumps, gather the data, and the vacuum becomes a thing you can predict, manage, stay ahead of. The promise is not empty, and the people selling it are very good at what they do. But the bake takes what the bake takes. On the electron gun I work on now, we run for about a week. On a X-ray tube I built in an earlier life, the bake ran above for three weeks. The screen watches the water leave the walls. It does not negotiate with it. No measurement I have ever taken has shortened a pumpdown by an hour.
The reason the promise sells anyway is arithmetic, and the arithmetic is real. In a leading-edge fab a pump that dies unannounced can scrap a hundred wafers and cost into the millions of dollars a year, so the largest pump makers have turned decades of telemetry into models that call a failure before it lands, and the largest toolmakers wire tens of thousands of chambers into a single screen. At that scale an hour of warning is worth a fortune. The money is not the lie. The lie is in what we let ourselves believe the data is.
Prediction is honest about a pump. Bearings wear, seals harden, a motor draws a little more current each month. That is a trend, and a trend can be foreseen. But the failures that actually ruin a vacuum are not trends. A leak, a virtual leak trapped behind a blind weld, a surface that quietly went dirty: these are events. No history tells you the hour a weld will begin to weep. The screen cannot see them coming, because there is nothing, yet, to see. On an event, what the screen adds is a faster alarm and a better autopsy.
Which leaves the real question. If it does not pump, and it cannot foretell the failures that matter, what does the screen actually do?
Not divination. Memory.
For years the knowledge of what a healthy pumpdown looks like lived in a few people's hands. The shape the curve should take. The pressure that ought to have fallen by Thursday and did not. The particular wrongness of a bad bake, felt long before any number confirmed it. That knowledge was real and it was load-bearing, and it left the room when the people did. A retirement, a transfer, and a fleet forgets how it is supposed to behave.
The screen writes the operator down. It takes an intuition that lived in one head and turns it into a curve the next person inherits. The generation before mine automated the gesture: the valve, the interlock, the sequence that used to be a hand on a knob. My generation documents the gesture, at the scale of a fleet, so it can be compared, audited, and remembered. That is a smaller claim than the promise makes. It is also the true one.
There is one place where this stops being a convenience and becomes a law of the building. In a fusion reactor, once the machine is running, no human will enter it again. The walls are activated, the fuel is deuterium and tritium, and the inside becomes a room no person can walk into. The vacuum system there is not modest, hundreds of pumps and some of line, watched without pause by software. That watching is there because there is no one to send. The exact move I make on a screen for nine bake stations, a reactor is forced to make for a machine no one can ever touch again. The scale is absurd. The gesture is identical.
Anywhere the people who can read a process by feel will one day leave the room, the instrument that wrote down what they knew is the only thing that stays. The dashboard makes the knowing survive the knower. That is a humbler promise than the one on the slide, and it is the one worth keeping.
So build the screen. It is worth building, and I have given mine years. But be honest about what it is. It does not pump. It remembers. The wall, the one where empty space pushes back, has not moved, and it has no API.
## Sources
- [A] Outgassing as the rate-limiting load in ultra-high vacuum, water as the dominant species at room temperature, and the bakeout that drives it off: O'Hanlon, A User's Guide to Vacuum Technology (Wiley); Redhead, Hobson & Kornelsen, The Physical Basis of Ultrahigh Vacuum.
- [B] Predictive maintenance on semiconductor vacuum pumps, the cost of an unplanned pump failure in a fab, and fleet-scale models trained on pump telemetry: Edwards Vacuum, Semiconductor Intelligent Service and predictive-maintenance case material; vacuum-pump predictive-maintenance papers in the SEMI Advanced Semiconductor Manufacturing Conference proceedings.
- [B] Tens of thousands of process chambers connected to a single analytics platform for diagnostics and predictive maintenance: Applied Materials AIx platform disclosures; Lam Research Equipment Intelligence.
- [A] Forecastable wear versus discrete failure events, the leak, the virtual leak, and surface contamination: O'Hanlon, A User's Guide to Vacuum Technology (Wiley).
- [A] The ITER vacuum system, the inaccessibility of the activated, tritium-fuelled machine in operation, and the resulting reliance on remote monitoring and handling: ITER Organization vacuum-system and remote-handling descriptions; ITER Vacuum Handbook.
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary.
---
# Tight Enough
> Two brazed joints, the same certified leak rate, opposite void networks under the objective. What decides is whether the voids percolate.
Date: 2026-06-17
Canonical: https://pierreribault.com/essays/tight-enough/
import Spec from '../../components/Spec.astro';
import Equation from '../../components/Equation.astro';
import Callout from '../../components/Callout.astro';
import U from '../../components/U.astro';
One morning I had two brazed joints under the microscope, side by side. Ceramic sealed to metal with an active braze, the kind that has to hold ultra-high vacuum and never let the outside in. Both had passed the same helium leak test. Both carried the same stamp on the traveller, the same number under the same bar. Leak tight.
Under the objective they were nothing alike. One was a dense, tangled lattice, the braze and its reaction product woven into a mesh with no way through. The other, at the same magnification, showed a channel. A clear path running where the ceramic met the metal, fine but continuous. Same certificate. Opposite insides.
The field keeps a tidy definition of tight. A leak rate, a figure in Torr·L/s, measured below some threshold, written down as a pass. It is a single number, and a single number is easy to trust, because it sorts. You can always say one seal leaks less than another, rank them, set the value in a cell and sign. But the stamp hides what the microscope shows. Two joints under the same stamp can be built on opposite topologies, and the difference is in how the braze is connected.
One number, stamped on both joints. The networks behind it are opposites.
So start where the field starts, at the leak, and notice that the leak is only the easy case. A leak is a hole, a channel that is short and almost straight, and gas pours through it the way water finds a crack. We have a word for how twisted a path is. Tortuosity. A clean hole has a tortuosity near one. Everything else, every real seal, lives somewhere on the long road between that hole and solid matter that has no hole at all.
A joint is a network. I have spent years reading the network that titanium builds inside an active braze, the reaction product it grows against the ceramic, and I can tell you that whether a joint seals is governed by connectivity, and barely at all by how much metal is in it. Picture the empty space in the joint, the pores and the microcracks, as a network of its own. As that network thins out, there comes a point, sharp and sudden, where no continuous path crosses from one face to the other. Below it, gas has nowhere to travel and the joint holds. Above it, a single thread of connected voids spans the gap and the joint leaks. Physicists who study disordered media have a name for that point. The percolation threshold: a geometric phase transition, a switch between a network that crosses and a network that does not. More braze can sit above the threshold and leak. Less can sit below it and hold. Tightness is a property of the network.
There is not one network in the joint. There are two, laid over each other. The first is the network of voids I just described, the gas-phase path, governed by percolation, the one that decides whether a molecule of helium can find its way through open space. The second is the network the titanium itself draws, the reaction product and the boundaries between phases, and this network is always connected, by construction, because it is the thing that bonds the ceramic to the metal. And a connected solid network can be a highway. Atoms cross a solid without an open pore: they dissolve into it and move along the interfaces, the grain boundaries, the interphases, faster there than through the bulk. So the same titanium mesh that seals the joint against open flow can, at the same time, short-circuit transport at the atomic scale. Density and tortuosity cut both ways, and which way depends on the gas you are asking about and the regime it travels in. A joint that is tight to helium at the detector can be a fast path for hydrogen, or for permeation creeping along the very seam that makes it strong.
Push that to its end and the two questions become one. When the void network is pushed below its threshold, when there is no open path left at all, gas crosses through the solid instead. It dissolves into the lattice and diffuses through the spaces between atoms, the interstitial network of the crystal itself. That is permeation, the same network problem at the smallest scale there is. The wall is just the densest, most tortuous network of all, the last one, the one made of atoms. A leak, a porous seal and a flawless wall sit on one continuous axis, and tortuosity is the single variable that runs its length, from the open hole to the lattice.
Aluminium nitride and alumina are a natural experiment. Take the same silver-copper-titanium braze and put it against each. Against alumina the titanium reacts with oxygen and grows oxides and titanates. Against nitride it reacts with nitrogen and grows titanium nitride. Different reaction product, different continuity, different network. And the thermal expansions do not match the same way, alumina near , nitride near , the braze near , so on cooling each builds a different field of residual stress, and stress is what seeds the cracks that become the void network. Change the ceramic and you change the mesh you get. The ceramic you choose is the network you inherit.
2O3", v: "Ti → oxides, titanates" },
{ k: "Against AlN", v: "Ti → titanium nitride" },
{ k: "CTE · alumina", v: "7–8", unit: "ppm/°C" },
{ k: "CTE · nitride", v: "4.5", unit: "ppm/°C" },
{ k: "CTE · braze", v: "≈ 19", unit: "ppm/°C" },
]}
/>
You cannot see tightness in a cross-section. A slice through the joint shows you a plane, and connectivity is a property of the volume. A void that looks sealed in the plane may be open just behind it, out of view. The one channel that percolates may pass through a corner the slice never touched. A two-dimensional section underestimates blockage every time, because it cannot see the connections that leave the plane, and it can call a joint tight that is not. Only the full volume sees the real path, or the integrated flux of a leak test, which adds up every route at once whether you can picture it or not. You see a plane, or you see a number, and you infer the rest. The measurement is itself an act of sealing off, a choice to sample a plane or to integrate a volume. The tomography I run is the instrument that settles what metallography could only suggest.
The difference is about what kind of thing tightness is. A leak rate is a point on a line, and a line lets you rank. A network refuses that. Two joints can fail in ways that do not compare, one with an open pore, one with a fast interface, and there is no honest sense in which one of them is simply tighter. There is no tight in the abstract. There is only tight enough, for this gas, for this long. The stamp on the traveller is the most honest version of the trap, because it tells you, exactly and truthfully, the one thing that was never quite the question. And the trap is not the brazer's alone. Anywhere a single number is trusted to stand in for a structure, a qualification that passed, a supplier that scored, a coverage that was met, the same thing hides in the same place. The work is always the same, to read the network the number cannot show.
And this same network sits under rooms I have never stood in. A fusion reactor cannot seal its tritium. Tritium is small and it permeates hot metal walls, and there is no wall thick enough or pure enough to stop it, only to slow it. So the field grows a barrier coating instead, an oxide a micron thick, alumina or erbium oxide, chosen because it raises the tortuosity of the path through the wall and drops the permeability by a factor of to . Then it licenses the flux that remains and keeps the books on it. Containment, at the scale where a fuel is also a radiological hazard, is tortuosity engineering plus accounting. And note the company that coating keeps, because the same machine carries another thing called a barrier, the thin titanium nitride on its radio-frequency surfaces, which stops nothing from permeating and instead lowers the secondary electron yield. One barrier of network, one barrier of surface, two physics under one word, in one chamber.
A barrier is a network you hold below its threshold, and a flux you agree to account for. A barrier is a verb. And once you see it as a verb you see it everywhere the same shape repeats, because knowledge percolates too. An embargo raises the tortuosity of a field of expertise. You watch what crosses anyway. The lock that cannot be counted, the one no export list can name, is simply the one whose tortuosity is high enough that the flux is slow. The titanium network in my joint and the lock under the fab are the same picture at two scales. Neither is a wall. Both are a network, and a rate.
So the stamp says tight, and the number is true, the same way a pressure is a rate and a clean surface is a distribution. Each time, the scalar written on the traveller is the honest answer to a question just beside the one that mattered. Things are made tortuous enough, for this gas, for this long, and the craft is knowing how much is enough.
## Sources
- [A] Percolation as a geometric phase transition and the spanning-cluster threshold in disordered media: Stauffer & Aharony, Introduction to Percolation Theory; Sahimi, Applications of Percolation Theory.
- [A] Effective transport through porous media scaling as porosity over tortuosity, and tortuosity as the governing variable for conductance: standard transport-in-porous-media literature; O'Hanlon, A User's Guide to Vacuum Technology (Wiley) for the leak, virtual leak, and permeation distinction.
- [A] Permeation as dissolution and diffusion through the solid lattice, and short-circuit diffusion along grain boundaries and interphases: standard solid-state diffusion texts; permeation chapters in Redhead, Hobson & Kornelsen, The Physical Basis of Ultrahigh Vacuum.
- [A] Active brazing of alumina and aluminium nitride with Ag-Cu-Ti, the titanium reaction product (oxides and titanates on alumina, titanium nitride on nitride), and the role of CTE mismatch in residual stress and microcracking: active-brazing and ceramic-to-metal joining literature; CuSil-ABA and TICUSIL data.
- [A] Tritium permeation barriers in fusion, alumina and erbium oxide coatings, permeation reduction factors of 10³ to 10⁵, erbia chosen for compatibility with liquid lithium and Pb-Li breeders where alumina is attacked: Chikada et al., Nucl. Fusion 51 (2011); erbia coatings on EUROFER and RAFM steels (PRF up to 10⁵ at 873 K); EU DEMO breeding-blanket overview, Federici et al., Fusion Eng. Des. 141 (2019).
- [A] Titanium nitride and low secondary-electron-yield coatings on radio-frequency and superconducting surfaces as a distinct, non-permeation barrier: SRF and klystron-window anti-multipactor literature.
- [C] Helium as the gas that cannot be cryopumped at usual temperatures and cannot be excluded by permeation, and the pumping strategy that follows in cryogenic and quantum hardware: cryopumping and UHV practice.
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary.
---
# The Better Ceramic
> What a material costs once it leaves the datasheet.
Date: 2026-06-11
Canonical: https://pierreribault.com/essays/the-better-ceramic/
import Equation from '../../components/Equation.astro';
import Spec from '../../components/Spec.astro';
import U from '../../components/U.astro';
The purchase order said aluminum nitride. The datasheet had made the case: seven times the thermal conduction of alumina, an expansion coefficient that shakes hands with refractory metals, and just enough electrical leak to let implanted charge drain away instead of building toward an arc. On paper, the better ceramic.
The paper said nothing about the next six weeks. A design-of-experiments campaign to industrialize one active-brazed seal, the seal the alumina route gives away almost for free. A sandblast booth with its standard grit banned, because the standard grit is corundum, and corundum is the wrong oxide to hammer into a nitride. Incoming lots sorted piece by piece on measured properties, because with this material a datasheet describes a population rather than a promise.
We had chosen the better ceramic. This essay is about what the better ceramic costs, and why that cost, rather than the datasheet, is how materials are actually chosen in the vacuum world.
## The case on paper
The part was an insulator on an X-ray source, and an insulator in a tube is asked to do three jobs at once: hold off the voltage, pass the heat, and refuse to hoard charge. Alumina aces the first and fails the other two politely. It conducts heat at roughly , and it insulates so well that an electron implanted by backscatter has nowhere to go. It waits there, distorting the field, saving itself for the flashover.
Aluminum nitride answers all three. Thermal conduction of and up, into the 200s when the crystal is clean. A thermal expansion near that matches molybdenum, tungsten and silicon, the metals and substrates you actually want against it, where alumina's 7-plus pairs it off with Kovar and niobium instead. And, in the right grades, a bulk resistivity low enough to bleed implanted charge away before it matters. Three jobs, three wins. No contest, on paper.
## Where the ceramics live
From that one insulator, the whole tube world sorts itself into a handful of ceramic posts. Each post asks the material a different ruling question, and each has settled, over seventy years, on its own answer.
| The post | Where it lives | The ruling question | Today's answer | Tomorrow's bet |
| --- | --- | --- | --- | --- |
| The standoff | X-ray tubes, electron guns | volts held, charge refused | alumina | graded insulators, some printed |
| The window | klystrons, gyrotrons | RF loss and secondary electrons | pure alumina, then diamond | Brewster-cut diamond, bigger disks |
| The rod | TWT slow-wave circuits | heat and RF through one part | BeO, then pyrolytic BN, then diamond | no ceramic, all-metal microcircuits |
| The lossy load | wherever RF must die quietly | absorb the wave, pass the heat | BeO–SiC, now AlN–SiC | loss engineered in layers |
A standoff can hold half a megavolt across two brazed cones. A window survives megawatts because of a coating a few nanometres thick. A helix rod must carry heat and RF through the same sliver of solid. The posts could not be more different, and in every cell of that table, the deciding move turns out to be the same one. Not the bulk. The joint. Tomorrow's bets obey it already: a Brewster window is a harder braze, a graded insulator is a sintering campaign, and the rod's successor is no ceramic at all.
## The answer off the paper
Because here is what alumina brings that no datasheet has a column for: a secret organ. Technical alumina at 94 to 97 percent purity carries a silicate glassy phase between its grains, the "impurity" that lowers its grade. The moly-manganese metallization process, the one the entire tube industry's hermetic seals stand on, is fired above in wet hydrogen and *feeds* on that glass: the phase migrates, wets the molybdenum sponge, and locks a metal skin onto the ceramic that a nickel plate and an ordinary braze can then take over. The impurity is the adhesive. Which produces the inversion every newcomer trips on: the purer the alumina, the harder it is to join.
Around that organ, seventy years of habit. Alumina shrugs at water through every step, grinding, lapping, ultrasonic cleaning. Lot N behaves like lot N−1. Its expansion has a marriage of convenience with Kovar that the whole feedthrough catalogue is built on. None of it is on the datasheet. All of it is the reason the default exists.
## Three invoices
Aluminum nitride is a non-oxide. No glass between the grains, nothing for moly-manganese to feed on, so the joint moves to active brazing: a filler carrying a few percent of titanium that must be persuaded, in vacuum, to react with the nitride itself. The reaction has a window, temperature, time, atmosphere, the oxygen it finds on the surface, and the window is narrow. Narrow windows are what design-of-experiments campaigns are for. Six weeks of furnace runs bought us a process. The alumina line next door had inherited one.
The second invoice is the surface itself.
It reaches the workshop in strange places. After an arc had marked a part, refinishing it meant qualifying a dedicated blast medium, because the universal one, corundum, is aluminum *oxide*, and driving the oxide of aluminum into the nitride of aluminum salts the surface rather than restoring it.
The third invoice is hiding inside the headline number itself.
This site is named for the molecule's mean free path; the solid keeps one of its own. Alumina's conduction is modest but honest. AlN's is magnificent and hostage to lattice oxygen, which varies with the powder, the sintering aid, the run, so the receiving inspection grew a new step: sort the incoming pieces on measured physical properties, one by one. The datasheet, as a population.
## The law
Once you see the pattern, it stops being a story about two materials.
Beryllia ruled the heat-through-an-insulator niche for decades, and better, the rod and the load of choice. It is leaving the industry now, and not because a single property of the finished part failed. Its machining dust causes berylliosis; regulation followed; a *process* property ended a *bulk* king. Its succession, AlN in the rods, AlN–SiC in the loads, is the same bill being paid by a new generation.
And at the very top of the ladder, diamond. Best column in every table, the material that took gyrotron windows from sapphire's hundred-kilowatt ceiling to megawatts, continuous, and some of the first megawatt windows died anyway, of a lossy surface layer born during the braze. The finest bulk on Earth, killed at the joint, until the joint was learned.
From the bottom of the ladder to its top, the same law: you do not choose a ceramic. You choose the chemistry of its assembly.
## Budget a campaign, not a purchase order
The bill does get paid down, by volume. AlN heaters and electrostatic chucks are ordinary industrial articles today because the semiconductor industry ran those DOE campaigns years ago and amortized them across fabs. That is also why AlN's price keeps falling: the LED and power-electronics lines pull the powder. A material matures not when its properties improve but when someone else has already paid its process debt.
The next customers at the counter are the frontier machines. Pulsed-power fusion wants feedthroughs that hold heroic voltages; quantum hardware wants substrates and packages that survive 300 K to millikelvin. Their teams will shortlist ceramics off datasheets, as everyone does, and the shortlist will be right. Then the swap will arrive with its true contents: a new metallization to qualify, a new braze window to map, a new cleaning sheet, a new incoming inspection. The material is the cheapest line in that purchase order.
1200", unit: "°C" },
{ k: "Hydrolysis onset, AlN", v: "ambient", unit: "humidity" },
{ k: "One industrial active-braze seal", v: "six", unit: "weeks of DOE" },
]}
/>
The better ceramic is real. The numbers are real, and on the right post they are decisive. But a material is never its properties; it is its properties plus everything a factory must do to keep them intact through a joint, a surface, and a thousand serial parts. The datasheet prints the first half. The bill prints the rest. Read the datasheet, then budget for the bill.
## Sources
- [A] Mo–Mn process & alumina brazing: Chiggiato et al., Manufacturing and Assembly for Vacuum Technology, arXiv:2006.12072; glassy-phase mechanism: ceramic-solutions.com technical articles; high-purity metallization gradients: J. Asian Ceram. Soc. (2023).
- [A] Active brazing / TICUSIL (63Ag–35.25Cu–1.75Ti): US 6,663,982; single-step CuSil-ABA UHV joints: Mater. & Design 63 (2014).
- [A] AlN hydrolysis: Krnel & Kosmač, J. Eur. Ceram. Soc.; Fukumoto et al.; AlN vs BeO substitution: Kettner et al., Microwave Journal (2001).
- [A] AlN lattice-oxygen / thermal conductivity: Slack, J. Phys. Chem. Solids 34 (1973); Virkar et al., J. Am. Ceram. Soc. 72 (1989).
- [A] RF windows & TiN anti-multipactor: Michizono et al., Vacuum (KEK); CEPC klystron window, Chin. Phys. B 27 (2018); window microstructure & F-centres: Saito, IEEE TDEI.
- [A] Gyrotron diamond windows & braze-born loss layer: GA-A23723 (DIII-D); Heidinger et al., IEEE Trans. Plasma Sci. 30 (2002); ITER EDA window R&D.
- [A] TWT rods & diamond: US 4,153,859; US 4,683,400; Han et al., IEEE TED (diamond-film rods, +20–60%).
- [A] Lossy ceramics BeO–SiC → AlN–SiC: US 6,579,393; J. Mater. Res. (SPS AlN–SiC).
- [A] X-ray standoff & charge management: Thales US 11,538,604; Behling, Modern Diagnostic X-Ray Sources, CRC (2021).
- [A] Tomorrow column: large-area diamond Brewster windows for DEMO, Fusion Eng. Des. (2020); KIT step-tunable 1 MW gyrotron with diamond Brewster window, IEEE Trans. Electron Devices 61 (2014); graded & 3D-printed insulators against flashover: Li et al., High Voltage (2020), Wang et al., High Voltage (2025); all-metal microfabricated slow-wave circuits: Paoloni et al., J. Electromagn. Waves Appl. 34 (2020), NRL 220 GHz serpentine (Cook, Joye et al.); layered lossy dielectrics: US 6,579,393.
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary.
---
# The Arc
> The one failure a tube, an accelerator, and a fusion magnet all share, and the craft that keeps it away.
Date: 2026-06-02
Canonical: https://pierreribault.com/essays/the-arc/
import Equation from '../../components/Equation.astro';
import Spec from '../../components/Spec.astro';
import U from '../../components/U.astro';
Push enough voltage across a gap in vacuum and, for a moment, nothing happens. Empty space is the best insulator there is. Then a single point on the metal, far too small to see, gives way. In under a microsecond the gap fills with plasma and shorts. The field calls it an arc, and every high-voltage tube ever built has been a quiet fight to keep it from happening.
I have fought it from both ends of the scale. A X-ray tube I built in industry, sealed, the size of a loaf of bread. An electron gun on an accelerator I bake today. The tube made photons, the gun makes a beam, and a fusion magnet next door wants neither, only to hold its voltage. Three machines that share nothing on the surface, afraid of the same thing underneath.
## It is field emission, run away
It starts on the cathode, at one microscopic asperity, not in the gap. A sharp point multiplies the local field by a geometric factor, and the current emitted from that point climbs faster than anything else in the system. The current heats the tip, the tip melts and vaporizes, the metal vapor ionizes, and now there is a plasma where there was vacuum. The plasma shorts the gap. The arc sustains the very field that lit it.
## You don't calculate past it. You condition past it.
A fresh surface field-emits and breaks down at low voltage. You don't force it. You raise the voltage slowly, letting small breakdowns burn off the worst emitters and pump away the gas, and the surface cleans itself up until it holds. It is patience, not cleverness, the same lesson as the bakeout.
Conditioning tracks the number of pulses. A breakdown damages the surface rather than cleaning it, and every new high-voltage program still counts the wrong thing first. So you push pulses, and you stay below the breakdown rate you intend to run at. Chasing sparks is chasing your own tail.
## Two signatures you actually watch for
**Dark current** is the tell. Steady field emission below breakdown, the precursor that says a surface is getting ready to fail. You watch it the way you watch a gauge drift.
**The Paschen strike** is the one that bites beginners. Not at operating vacuum, where it is safe, and not at atmosphere, where it is safe, but in between, during pump-down or venting, crossing the pressure where a few hundred volts will arc in gas the rest of the curve shrugs off.
## Pulsed buys margin
Same surface, same metal. Run it DC and you hold near twice the Kilpatrick field. Pulse it under a millisecond and you reach near five times it. The reason is time: the arc has to heat and melt its way to a plasma, and a short pulse ends before it gets there. It is why a pulsed accelerator runs at gradients a DC gap would never survive, and why my DC X-ray tube had no such slack. Its only margin was a clean surface.
## The same problem makes the photons
The X-ray tube is the other end of my career and the same physics. Electrons leave a hot cathode, cross the gap under the tube voltage, and stop in a tungsten anode. Most of that energy, about ninety-six percent of it, becomes heat, which is why the anode is tungsten. A sliver comes out as X-rays: a continuous bremsstrahlung floor, with sharp characteristic lines on top, set by the metal. And the hardest photon the tube can make is fixed by one number alone.
## The RF-only failure, for completeness
One mechanism needs neither asperity nor gas. In an RF field an electron knocked off a wall can be driven back in phase to strike again and release more than one secondary. If the timing resonates and the yield exceeds one, the population multiplies, a multipactor. It loads cavities and cracks windows. It is a resonance rather than a cascade, so the cures are different: geometry, a surface that emits fewer secondaries, a bias field.
## What can bite you
| Failure | Where | Mechanism | Fix |
|---|---|---|---|
| Vacuum arc | cathode, any HV surface | field emission at an asperity, runaway to plasma | clean, conditioned surface |
| Flashover | gun insulator, ceramic window | triple-junction emission along a dielectric | shaped junctions, clean ceramics |
| Dark current | high-field surfaces | sub-breakdown field emission | low-β geometry, no sharp edges |
| RF breakdown | output cavity, structure | pulsed surface field over the limit | lower gradient, more conditioning pulses |
| Multipactor | windows, couplers | resonant secondary-electron multiplication | geometry, low secondary yield, bias |
| Paschen strike | whole tube, at pump or vent | gas avalanche near the pressure minimum | pump and vent fast through the minimum |
The voltage at which empty space gives up is not a number you can design your way under once and forget. It is a line you walk every time you bring a machine up, pulse by pulse, watching the dark current, crossing the Paschen floor fast, never chasing a spark. The X-ray tube taught this to one industry over forty years. The accelerator is teaching it again. The fusion magnet is about to learn. The arc was never divided between worlds. Neither is the craft of holding it off.
## Sources
- [A] Latham, R.V., High Voltage Vacuum Insulation, Academic Press.
- [A] Kilpatrick, W.D., Criterion for Vacuum Sparking Designed to Include Both rf and dc, Rev. Sci. Instrum. 28, 824 (1957). — https://doi.org/10.1063/1.1715731
- [B] Slade, P.G., The Vacuum Interrupter: Theory, Design, and Application, CRC Press.
- [A] Dyke, W.P. & Trolan, J.K., The Field Emission Initiated Vacuum Arc. I. Experiments on Arc Initiation, Phys. Rev. 91, 1043 (1953). — https://doi.org/10.1103/PhysRev.91.1043
- [A] Mesyats, G.A., Ecton mechanism of the vacuum arc cathode spot, IEEE Trans. Plasma Sci. — https://doi.org/10.1109/27.476469
- [A] Degiovanni, A., Wuensch, W., Giner Navarro, J., Comparison of the conditioning of high gradient accelerating structures, Phys. Rev. Accel. Beams 19, 032001 (2016). — https://doi.org/10.1103/PhysRevAccelBeams.19.032001
- [A] Loveless, A.M. & Garner, A.L., The Transition to Paschen's Law for Microscale Gas Breakdown at Subatmospheric Pressure, Sci. Rep. 9, 5669 (2019). — https://doi.org/10.1038/s41598-019-42111-2
- [A] Wangler, T.P., RF Linear Accelerators, 2nd ed., Wiley-VCH.
- [A] Grudiev, A., Calatroni, S., Wuensch, W., New local field quantity describing the high gradient limit of accelerating structures, Phys. Rev. ST Accel. Beams 12, 102001 (2009). — https://doi.org/10.1103/PhysRevSTAB.12.102001
- [A] Palaia, A. et al., Effects of rf breakdown on the beam in the Compact Linear Collider prototype accelerator structure, Phys. Rev. ST Accel. Beams 16, 081004 (2013). — https://doi.org/10.1103/PhysRevSTAB.16.081004
- [A] Duane, W. & Hunt, F.L., On X-ray wave-lengths, Phys. Rev. 6, 166 (1915). — https://doi.org/10.1103/PhysRev.6.166
- [A] Behling, R., Modern Diagnostic X-Ray Sources: Technology, Manufacturing, Reliability, CRC Press (2021).
- [A] Parodi, R.F., Multipacting, CERN Accelerator School 2010, arXiv:1112.2176. — https://arxiv.org/abs/1112.2176
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary. Links verified August 2026.
---
# Everything Outgasses
> A law that does not care whether you are building an X-ray tube, an accelerator, or a star.
Date: 2026-05-31
Canonical: https://pierreribault.com/essays/everything-outgasses/
import Spec from '../../components/Spec.astro';
import U from '../../components/U.astro';
Somewhere this year, a steel chamber is being sealed shut. It is enormous. Inside it, magnets cold enough to touch the temperature of deep space. The chamber has one job before anything else can happen. It has to hold a vacuum, a deep one, as clean as the surface of the Moon.
They pump it down. The gauge falls toward nothing. And then the chamber starts to breathe.
Not air. There is no leak. The walls themselves are letting go of gas, molecule by molecule, from steel that looked perfectly solid. The pressure stops falling. It even climbs. The vacuum they just built is being spoiled from the inside, by the chamber itself. The field has a plain word for this. Outgassing. Everything does it. Metal, glass, a fingerprint, a weld. Seal anything into a vacuum and it will quietly exhale for days.
I have met this wall three times, in three worlds that do not talk to each other. An X-ray tube in industry, a sealed object the size of your hand. A synchrotron, a ring of light, where the vacuum is the condition for the beam to exist at all. And a large accelerator, a free-electron laser, where I bake chambers today and read what comes off their walls. From a tube you can hold to a machine a kilometer long. Three scales, one problem. The vacuum did not care which one I was standing in.
That is the strange and beautiful thing. The physics underneath is the same everywhere. The cultures are not. A field that has spent forty years on small sealed tubes knows things in its bones that a young, fast-moving field has not yet had reason to learn. And the reverse is true too. Nobody owns this knowledge. It just gets paid for, again and again, in different rooms.
So here is what the vacuum teaches, every time, whatever you are building.
Water is the patient enemy. It clings to every surface as a thin invisible film, and at room temperature it leaves on its own slow schedule, far too slow. So you heat the whole chamber and drive the water off before it can leave during the work that matters. We call it a bakeout, and the cost of it tells you how stubborn the gas is.
On the electron gun I work on now, we bake at for about a week. On a X-ray tube I built earlier, the bake ran above for three weeks. Same gesture, wildly different price. The cleaner and the higher the voltage you need, the more you pay to get there. It is patience more than cleverness. You pay the debt early so it does not come due at the worst moment.
Materials have a memory. Whatever a metal absorbed during its life, it gives back later, under heat and vacuum. So you choose a material for what it will not say later. The wrong alloy, the wrong oil, a polymer that seemed harmless, and your clean vacuum is ruined by a decision made months earlier. Half the craft is refusing materials that will betray you down the line.
And what you cannot remove, you trap. The oldest trick in the sealed tube is the getter, a small patch of reactive metal that silently captures stray gas and holds it for the life of the device. It is a hundred years old. It powered the radio valves your grandparents listened to. And it never left. Open the packaging of a superconducting quantum chip today and you can find the same idea, a getter keeping a few cubic centimeters clean enough for a qubit to survive. A hundred-year-old vacuum-tube trick, holding up the most modern machine we know how to build. The same physics, reaching across a century.
None of this is secret. It is the ordinary discipline of anyone who works with empty space. But it was learned the hard way, over decades, mostly by people building things that have nothing to do with each other.
The tube wanted a clean image. The accelerator wants a clean beam. The fusion machine wants a clean plasma. Three goals that share nothing on the surface, sitting on one discipline underneath. A lesson one world paid for in full, another world can simply borrow. It does not have to be bought twice.
The vacuum does not know whether the chamber around it makes X-rays or makes a star. It exhales from the walls, it hides water in the corners, it remembers what the metal absorbed. Empty is empty.
I have had the rare luck of meeting it from both ends of the scale. That is the only reason I can say it out loud. The walls between these worlds are mostly habit. The physics underneath was never divided.
## Sources
- [A] Redhead, P.A., Recommended practices for measuring and reporting outgassing data, J. Vac. Sci. Technol. A 20, 1667 (2002). — https://doi.org/10.1116/1.1496783
- [A] Chiggiato, P., Outgassing, CERN Accelerator School: Vacuum for Particle Accelerators (CERN Yellow Reports).
- [A] O'Hanlon, J.F., A User's Guide to Vacuum Technology, 3rd ed., Wiley, 2003 (outgassing chapters).
Ratings follow the Stack scale: [A] primary or standard, [B] manufacturer, [C] secondary.
---