Case study — identifying suppliers into ONE European product by OSINT

Product: Robert Bosch GmbH hydrogen fuel-cell power module (FCPM) — module assembly at Stuttgart-Feuerbach (volume production since 2023), PEM stack at Bamberg. Date: 2026-08-22. Effort: five parallel research passes across thirteen channel types, ~6 hours of agent time (~340k tokens). Purpose: test how deep outside-in OSINT can actually go on a single product, and record which channels earn their cost. This is a method case study as much as a findings one.

Why this product: built in Europe, uses genuinely chokepointed materials (platinum, iridium, plus rare-earth magnets in the balance-of-plant), and fuel cells generate unusual amounts of public signal — EU funding, licensing deals, research consortia.

1. What we found — the confirmed chain

Three external component/technology relationships confirmed from a primary source, plus four infrastructure/equipment suppliers and a fully-mapped internal chain.

Three rows are now live in `bosch.md` as named_counterparties — the dossier's first, taking the corpus from 17 to 18 dossiers with a supplier layer.

The single most important structural finding

Bosch is deliberately vertically integrated on this product. Compressor, blower, ECU, sensors and >50% of the production line are in-house; what it buys externally is exactly the two highest-IP stack materials (CCM, bipolar plates) plus the stack architecture licence. So for this product, the supplier map is mostly a plant map — and a method that only hunts for external vendors would have concluded, wrongly, that the chain is opaque.

The relationship that inverted — and why single-vintage reads are dangerous

PowerCell was the licensor in 2019 (Bosch pays to use its design). By 2023 the physical flow reversed: "Bosch manufactures and delivers tested, activated 100 kW fuel cell stacks to PowerCell's Gothenburg facility" — PowerCell became a customer. In June 2025 Bosch acquired further rights to adapt the stack for China.

A snapshot taken in any single year mislabels this relationship. It is filed as role: both with the inversion documented. This is the strongest argument in the whole exercise for the dated standing record — a one-shot AI query answers "who supplies Bosch's stack?" with "PowerCell," which has been wrong since 2023.

2. Channel-by-channel yield — the methodological payload

Thirteen channels tried, seven productive, and the failures are as informative as the hits — three failed for structural reasons we can now predict in advance rather than rediscover per company.

2b. The finding no supplier-hunting channel would ever produce

The works-council / union channel named zero suppliers — and produced the most commercially important fact in the whole exercise:

• Nikola, the named pilot customer, filed Chapter 11 in Feb 2025, with Bosch among the largest

creditors (~$13.3M claim). Bosch paused its Anderson, South Carolina fuel-cell expansion as a direct result.

• German works-council reporting (paywalled, Dec 2024) indicates **no orders for hydrogen drives

from truck manufacturers**.

• IG Metall Stuttgart material documents the fuel cell going from site-security hope to retrenchment;

Bosch is cutting 3,500 Feuerbach jobs by 2030 and shed 550 SOFC staff.

Why this matters methodologically: a supply-chain map of this product built purely from corporate PR would describe a healthy ramp. The union and regional-press channels show a product whose pilot customer is bankrupt and whose order book is reportedly empty. Exposure analysis that ignores demand-side collapse will misprice the risk — a company's supplier graph is only as load-bearing as the product's order book. This is a strong argument for pairing the supplier layer with a demand-reality check, and it is only reachable through non-corporate channels.

3. Coverage against the bill of materials — honest scorecard

≈6 of 9 BOM elements attributed — and critically, the platinum entry point (the material we most care about) is one of the confirmed ones.

4. What this proves, and what it doesn't

Proves: with concentrated effort on one product, outside-in OSINT reaches named, primary-sourced, component-level supplier attribution — including the critical-material entry point — for a company that publishes no supplier list, is privately held, and files no 10-K. Three of these relationships are now structured data in the corpus, one linking to an existing dossier (jmat-l).

Doesn't prove: that this scales cheaply. This took five parallel research passes on one product line of a company with ~32 named plants. Bosch has many product lines; the corpus has 2,527 companies. The honest unit economics: this depth is a premium, per-product deliverable, not a corpus-wide sweep.

Doesn't establish volumes or shares. Every row here has share_pct omitted, because no share is disclosed anywhere. We know who, not how much — which is exactly the boundary the MOFCOM 0.1% product must respect (we supply provenance, the manufacturer does the arithmetic).

5. Reusable playbook (ranked by value per hour)

1. Invert the search. Query candidate suppliers naming the OEM, not the OEM naming suppliers. Suppliers announce wins; OEMs don't. Run the BOM as your candidate list. 2. Go to the national funding database first. EnArgus/Förderkatalog named five companies with amounts and dates in one pass — and exposed the internal business-unit code, which is a high-precision pivot key. https://www.enargus.de/pub/bscw.cgi/?op=enargus.eps2&q=<term> 3. Search in the OEM's language. OMB Saleri and the Dana coverage surfaced only via German queries. English-only search systematically under-reads European chains. 4. Mine the start-of-production press kit. SOP releases name inter-plant flows that no annual report ever will — and they tell you what isn't bought externally. 5. Force every relationship into one bucket: equity / IP licence (with direction) / contract manufacturing (with direction) / component supply / customer. Trade press conflates all five. Both Bosch relationships here are routinely mis-reported. 6. Check three structural questions before spending effort: Does the target own a special-machinery arm? (kills the equipment channel) Is the business a separate legal entity? (kills registries) Does the product cross a US border? (kills customs) 7. Re-check trade codes annually. A dedicated CN code now exists — 8501 33 10, hydrogen fuel-cell DC generators 75–375 kW — which enables module-flow corroboration (e.g. DE→SE, consistent with Bosch's deliveries to PowerCell) though never company attribution.

6. Leads deliberately NOT filed

Recorded so nobody re-researches them, and so the discipline is visible:

• Freudenberg → GDL: 2022 "global automotive tier one" contract for GDLs, trucks/buses

application, German. Fits Bosch. Customer never named — not filed.

• Umicore / FUMATECH / Coperion / Gustav Klein / Deutsche Edelstahlwerke: all in *funded R&D

consortia with Bosch's fuel-cell unit (EnArgus, with grant numbers). A consortium partner is not* a serial supplier — these are one evidential tier below a supply contract. Filed here as context, not as named_counterparties.

• Syensqo: won a 2025 Bosch Global Supplier Award for specialty polymers, and makes Aquivion

PFSA membrane. The award text never mentions fuel cells — plausible, unproven.

• Ceres Power: Bosch held ~17.4% and licensed SOFC tech — different technology, exited 2025.

Routinely mis-reported as a fuel-cell supplier relationship. Explicitly excluded.

7. Follow-ups this generated

• Onboard Dana and PowerCell — both confirmed suppliers, neither has a dossier.
• Johnson Matthey (jmat-l, has an LEI) should get dossier_slug linkage so its own

platinum/iridium exposure becomes reachable from Bosch — the recursion the schema was built for.

• Bosch Global Supplier Award winner lists are published as machine-readable PDFs every two

years (49 winners in 2025) — a standing supplier census worth ingesting corpus-wide, though it never says which product a supplier serves.