Palantir is the clearest example of a software company becoming part of how a state operates. Understanding how it got there is more useful than having an opinion about it.

Founded in 2003 with backing aligned to US intelligence interests, it started as a counterterrorism tool for fusing data that would not otherwise meet. Two decades later it sells to Five Eyes governments, NATO members, EU states and a long list of civil agencies. The technology matters less to that story than three deliberate choices.

Model the data before building the application

The ontology came first: entities, relationships and a disciplined model of what the organisation actually deals with. Government data sits in silos built by different people for different reasons over thirty years, and most integration projects fail by trying to move it. Palantir instead invested in the layer that lets messy sources be reasoned about together, then built pipelines, dashboards and machine learning on top of that substrate rather than directly on the sources.

Put engineers inside the customer

Forward-deployed engineers sit with the operators and analysts who will use the thing. It is an expensive way to sell software, and it solves the problem that kills most defence IT: the requirement nobody could articulate in a document. It also makes the product very difficult to remove, which is a commercial feature as much as a technical one.

Win the hardest customer first

Defence and intelligence were the anchor, not the adjacent market. Once a platform is trusted at the sensor-to-shooter level, in programmes of the JADC2 and Army Vantage type, selling into logistics or public health becomes a conversation about scope rather than credibility. Analysts have put the defence pipeline well into the billions (one assessment), and the civil work then funds capability that flows back the other way.

In the UK and the rest of Five Eyes this has taken the form of sovereign deployments and partner-integrated variants, alongside a continuing argument about a domestic sovereign foundry that would keep the capability without the dependency. Within NATO and the EU the platform has mostly been used as neutral ground: cross-border crisis monitoring, pandemic logistics, joint exercises. Being the fabric rather than the application is a strong position to hold.

The ecosystem supplies the domain knowledge

No single company understands every warfighting domain or classified environment, and Palantir does not pretend to. Primes, systems integrators and specialist SMEs build the ontologies and domain logic that run on the core. Palantir keeps the substrate; partners carry the expertise. That is how the deployment footprint grew faster than the headcount.

Where it is genuinely exposed

The war in Ukraine put the operational case beyond argument, and the markets have responded accordingly (commentary). The risks are structural rather than technical. Deep reliance on one vendor narrows a government’s room to manoeuvre, and both France and Germany have pushed for alternatives on exactly that basis. The forward-deployed model is costly and does not obviously scale without being diluted. And closed model layers sit awkwardly with the auditability that defence oversight increasingly demands.

What is worth copying

Start with one high-value use case and a small team that has a real problem, not an enterprise rollout. Put your own engineers alongside the users so that the vendor is never the only party who understands the system. Build a partner network for domain logic instead of trying to hire every specialism. Design for air-gapped operation from the start, because retrofitting isolation is close to impossible. Make judgements explainable, because somebody will eventually ask. And assume political pushback against dependency on a single large supplier, because it is already here.

The lesson is less about data platforms than about sequencing. Palantir delivered something useful to a sceptical customer, stayed close enough to keep delivering, and only then scaled. That order is the hard part, and it is the part most imitators skip.