Arsenale: the manifesto and the first node

In April 2026 the EU Commissioner for Defence told the European Parliament that Russia produces four million artillery shells a year and the Union two. That is the number that opened the first piece in this series. The one that interests me here is a different one, two years older: in 2024 Brussels claimed an ammunition production capacity roughly three times what industry estimated. It was not lying. It did not know.

This is obvious inside a factory and almost never obvious outside one. Europe’s defense industrial base is made of more than 2,500 small and medium enterprises that the Commission itself describes as organized “on a national basis, with limited cross-border cooperation”, and that the European Defence Agency calls “fragmented”; in Italy SMEs are eighty percent of the sector. Each of those companies knows how much it can produce, and knows it well. The aggregate number does not exist because nobody adds it up, and nobody adds it up because doing so would mean asking 2,500 companies to put their data in one place. Which is exactly what they will not do, and they are right.

Meanwhile the rest of the system behaves as if the map existed. Readiness 2030 puts 800 billion euros769566_EN.pdf) on the table. The EDIP regulation requires proof that non-EU components do not exceed thirty-five percent of the value, and somebody will have to compute that number, component by component, along supply chains whose second tier is often not even mapped. When the orders are firm, factories get built: Rheinmetall built one for 350 thousand shells a year in fifteen months. A mobilization plan without a map, though, is still a plan built on trust. Where that map has to live, and what happens when it lives inside another country’s infrastructure, is what we wrote about with the numbers in The map cannot live in Denver.

The Commission was not lying about European production capacity. It did not know it. Of the two, the second is worse.

Say “map” and the first thing that comes to mind is a census: a questionnaire, a deadline, a file. It has been tried, and it does not hold, for three reasons that are not opinions. The first is time: a census is true the day it closes and stale the day after, while a supply chain changes with every non-conformity and every supplier delay. The second is verifiability: in a questionnaire a number is a statement, and nobody can trace it back to the rows it came from. The third is trust: a company filling in a questionnaire is handing over data, a company exposing a status is handing over nothing.

The shape of the system follows from those three constraints, not from an architectural preference. You need something that sits inside the plant, wired to the ERP and to quality, that computes the answer on site and lets only the answer out. The supplier sees the orders. The buyer sees the status: capacity available yes or no, with the band, overall on-time delivery, how many single-source dependencies and from which countries. The authority sees aggregates. Neither sees the other’s data, and every external request is written into the log of the company that received it. This is not a product detail: it is the only form in which a business owner agrees to be part of the map. Oliver Wyman, which looked at more than 600 second- and third-tier suppliers, writes that most supply chain departments “do reactive problem-solving”. Not because they are careless: because nobody shows them the whole picture.

And here is the point. If that thing exists, the sum of the nodes is the map of a country’s production capacity: how many companies, making what, how much, dependent on which foreign components, and where the chain breaks if one supplier fails. That is strategic information before it is a software product, which is why it has to stay in the country, in nodes owned by the companies, with a public specification of what a node exposes and what it never does, and with code somebody can inspect.

This is not a position against anyone. In the United States they are doing exactly this, and it works: L3Harris reports using Palantir’s platform to find component shortages in night-vision goggle production and to rebuild the provenance of every part by reading millions of shipping documents. That is the proof the thing can be done, and it should be said. It is also the reason Europe cannot simply buy it: a map of European production capacity living inside another country’s infrastructure is convenient as long as interests line up. The rest of the time it is a dependency, and dependencies are counted, not debated.

A map living inside another country’s infrastructure is convenient as long as interests line up. The rest of the time it is a dependency.

So far this is a manifesto, and there are enough of those around. That is why, alongside the Arsenale manifesto, we are publishing something you can open and poke at: the node running on Vertesa, no sign-up.

Vertesa S.p.A. is a factory that exists only on paper, with synthetic data and real problems planted inside on purpose. The node has no free-form chat: it answers four fixed questions. How much can we produce in ninety days, by product family. What is the binding constraint. Who are we single-sourced from, and from which country. What do we not know. Every number is clickable and leads to the rows it came from, with the formula spelled out underneath.

The part that matters for this article, though, is a selector at the top: “View as: Vertesa / Buyer / Authority”. Switch the view, and the same company shows different things.

The constraint that has a first and last name in the company view becomes “one single-source dependency”, with the country and nothing else, in the buyer view. Prices, margins, client names, and drawings never appear in any external view. That is the difference between writing “your data stays yours” on a slide and showing it in three seconds. For someone reading from a ministry or an industrial district, that is the moment “federated” stops being a conference word and becomes a line in a specification.

The page I think is worth more than the other three is called “What we do not know”. The node lists whatever stops it from answering with certainty, and it finds it in the data by itself: a supplier that appears under three different spellings, dozens of deliveries with no actual date and no status saying whether they are open or simply unrecorded, scrap recorded in two units of measure in the same column, no stock figures at all. For each one there is the effect on the answer, in plain words: that supplier’s on-time delivery ranges between sixty-one and eighty-four percent depending on how the blanks are read. And next to it, the decision somebody in the company has to take.

Then there is the rule that holds everything else up. The system does not decide. It proposes the normalization, and the proposal stays a proposal until a person with a name and a role approves it. The approval goes into the log together with the question, the answer, the rows used, the version of the data, and the version of the rules it was produced with. The log exports.

It sounds like bureaucracy and it is in fact the operational difference. A chat output does not go into a quality system: it has no source, no version, no signature, and on audit day there is nothing to show. A node output does, because it is born in the shape those companies already work in. Anyone running a certified SME understands this sentence before all the others.

A chat output does not go into a quality system: no source, no version, no signature.

The name is not decoration. The first industrial production line in history was the Arsenal of Venice, three centuries before Ford. The Republic’s deterrent was not the fleet: it was the ability to build it, and the fact that whoever decided knew how fast it could be built. It is the same distinction we have today between having the money and knowing how much you can produce.

The Italian part is not a flourish either. Italy’s Ministry of Defence strategy “AI and Defence 2026” calls for modular, open systems without lock-in, and for an ecosystem of mid-sized firms alongside the large contractors. We read that as a technical requirement, not as a statement of principle. Modular means the AI model inside the node is replaceable today and in five years, because the model is a commodity and the map is not. Open means the specification of what a node exposes and what it never does is public from today, at version 0.1, so anyone adopting it strengthens the network, whoever they are. No lock-in means the node shuts down when the company decides to shut it down, and the know-how stays with the company.

The limits, to be clear, because in this sector they have to be said out loud. We do the factory, not the battlefield: no targeting, no weapon systems, no operational decisions. Capacity, supply chain, data, rules. And no invented numbers: if a figure is missing, the cell stays empty with the explanation next to it.

If you run an SME producing for defense, the move costs three minutes: open the demo and look at it thinking about your data, not ours. The question is not whether Vertesa’s number is right. It is whether your company could produce the same number with the same traceability, and how long it would take. If you like the answer, you already have almost everything in house. If you do not, you know exactly which piece of data to start from. And if you want to be one of the first real nodes, write to us: we are not asking for data, we are asking for an hour to understand how it works at your place.

If instead you work in a ministry, an agency, or an industrial district, the move is even simpler and you can make it today. Send your supply chain a single question: how much can you produce in ninety days, by product family, and which foreign components do you depend on. Then time it. How long the answer takes to arrive, and the shape it arrives in, are the exact measure of the problem this piece is about. If it comes back in three weeks across twenty different spreadsheets, you have just reproduced, in miniature, the wrong number from 2024.

The full argument is in the manifesto. The one underneath it, on why a company’s operating brain should not live on rent, is in data sovereignty. This piece has only one: the map of production capacity is infrastructure, and infrastructure is not something you rent.

1. [1]EU Perspectives, “Kubilius: we ask defence industry to speed up”, April 2026. euperspectives.eu
2. [2]RFE/RL, investigation into Europe’s ammunition supply chain. www.rferl.org
3. [3]European Commission, Defence SMEs. defence-industry-space.ec.europa.eu
4. [4]EESC and European Defence Agency, “No more national preference”, February 2025. www.eesc.europa.eu
5. [5]L’Imprenditore, CTNA, “Aerospace: a growing supply chain, 80% of the sector is SMEs”. www.limprenditore.com
6. [6]European Parliamentary Research Service (EPRS), briefing on European rearmament, 2025 (PDF). www.europarl.europa.eu
7. [7]Council of the European Union, final approval of the EDIP regulation, December 8, 2025. www.consilium.europa.eu
8. [8]Rheinmetall, opening of the Unterlüß plant, September 5, 2025. www.rheinmetall.com
9. [9]L3Harris, “Palantir and L3Harris: reindustrializing defense through AI”, December 2025. www.l3harris.com
10. [10]Forbes, Oliver Wyman, “European defense buildup may cause supply chain delays and shortages”, June 25, 2025. www.forbes.com
11. [11]Agenda Digitale, Italy’s Ministry of Defence strategy “AI and Defence 2026”. www.agendadigitale.eu