The Procurement Layer: Where Europe’s AI Decisions Actually Happen

Dell now ships Palantir as data center infrastructure. Baden-Württemberg signed its Palantir contract before it had a law permitting the software’s use. The NHS wrote a break clause where a policy dec

On 18 May, at Dell Technologies World, Dell and Palantir announced that Palantir’s Foundry, Ontology, and AIP platform will ship as an integrated stack on the Dell AI Factory with NVIDIA, with Dell ObjectScale and PowerFlex providing the storage layer beneath Palantir’s Ontology services. The announcement was read two ways: as a commercial partnership extending Palantir’s reach into the enterprise data center, and as a sovereignty pitch aimed at regulated buyers in banking, healthcare, government, and critical infrastructure who cannot or will not move sensitive data to a public cloud.

Both readings are accurate. Neither is necessarily the consequential one. What changed on 18 May is procedural: the venue in which the decision to deploy Palantir gets made.

The mechanism

Call it procurement-layer governance. When contested software arrives inside a hardware vendor relationship, the decision to adopt it moves out of the channels where its political character is legible and into channels where it is not. A standalone Palantir contract is a policy decision. It attracts scrutiny from legal counsel, data protection officers, legislators, and the press, because the vendor’s name carries a documented history at the policy layer. A Palantir workload arriving as a validated architecture inside a Dell AI Factory refresh is an infrastructure decision. It is evaluated by IT procurement, vendor management, and hardware RFP processes, on the criteria those processes are built to apply: price, interoperability, support terms, reference architectures, existing vendor relationships.

Those criteria are not wrong. They are simply not the criteria under which Palantir has repeatedly been contested. Procurement channels have no line item for the question a legislature or a constitutional court would ask. Bundling does not resolve the contestation that has followed this vendor through the German courts, the NHS, and the European sovereignty debate. It relocates that contestation to a venue where it may never be raised.

This is not a claim about intent, and it does not require one. It is a claim about institutional design. Procurement processes were built to prevent corruption and ensure value for money, and they are reasonably good at both. They were never built to surface the political stakes of a software dependency, and bundling contested software into infrastructure ensures that they are the processes doing the evaluating.

Germany: the contract precedes the law

Germany has run the experiment repeatedly, and the sequencing is the finding. Hesse purchased Palantir’s Gotham in 2017 and operated it as hessenDATA, with the statutory basis written afterward to fit the established practice. In February 2023, the Federal Constitutional Court struck down the provisions authorizing automated police data analysis in Hesse and Hamburg, finding they violated the right to informational self-determination by allowing police to generate, in the court’s words, “with just one click” comprehensive profiles of persons and groups without sufficient thresholds for interference. The software came first. The constitutional question was answered six years later, and only because civil society litigated it into the venue where it could be asked.

The same judgment records a detail that renders the mechanism as contract language. Bavaria had signed a framework agreement with Palantir for its VeRA analysis platform structured so that other German states could adopt the product without conducting their own procurement procedures. The contract did not merely acquire software for one state. It built a bypass around the venue where fifteen other states would otherwise have had to deliberate.

Baden-Württemberg then improved on the sequence. In March 2025, the state interior ministry signed a five-year contract with Palantir worth roughly 25 million euros, with no exit clause, at a time when no legal basis for the software’s use existed; the ministry’s stated justification was a price-binding window that would otherwise lapse. The Landtag amended the police law in November 2025 to permit what had already been purchased, over the objections of the state data protection commissioner, who argued the draft did not meet the constitutional court’s requirements. Deployment begins this year. In the same session, the parliament passed a resolution calling for a transition to a European alternative by 2030. The order of operations is the argument of this post stated as a legislative calendar: contract, then law, then politics as a non-binding resolution appended at the end. The Society for Civil Liberties has meanwhile taken a constitutional complaint against Bavaria’s deployment to Karlsruhe, and the federal interior ministry introduced draft legislation in March to open the door for analysis software in the federal police. The pattern is not receding. It is federalizing.

The NHS: contestation on time, and what it costs

The UK, meanwhile, shows what happens when the political reading arrives on schedule instead. When NHS England awarded the Federated Data Platform contract in November 2023 to a group led by Palantir, with Accenture, PwC, NECS, and Carnall Farrar in support, worth up to £330 million over seven years, the vendor question had already surfaced publicly: doctors’ organizations, patient groups, and privacy campaigners contested the award, and trust became an operational variable in the rollout. The contestation has not closed. The contract’s first term ends in March 2027, ministers are reportedly weighing the break clause, and Parliament debated the platform in Westminster Hall this April. Meanwhile the independent evaluation NHS England commissioned to assess whether the platform delivers value runs to 2029, two years past the point at which the extension decision must be made. Even where the political question forced its way in, the procurement calendar, not the evidence calendar, sets when it gets answered.

Residency is not jurisdiction

The sovereignty pitch attached to the Dell partnership answers a narrower question than the one being asked. On-premises deployment answers data residency: the data stays in the building, on infrastructure the customer owns. It does not answer vendor jurisdiction. Palantir remains a US company subject to US law, including the CLOUD Act’s provisions on data held or accessible by US providers, and the operational dependency on the vendor’s software, updates, and support does not diminish because the servers are local.

This matters because the sovereignty question, as posed by European governments and regulators, is a jurisdiction question. The sovereignty answer, as packaged in the on-premises pitch, is a residency answer. The two pass each other. And the gap between them is invisible at the procurement layer, where “sovereign deployment” appears as a checked compliance box rather than a claim requiring interrogation. A procurement rubric can verify where the data sits. It cannot weigh what it means that the semantic layer organizing a government’s operational data answers to another government’s legal process.

“No alternative” as a procurement argument

In May, NATO’s Supreme Allied Commander Transformation, Admiral Pierre Vandier, told Politico that “as far as I know, today there is no real competitor for Palantir,” in the context of the alliance’s adoption of Palantir’s Maven Smart System. The claim collapses three assertions: that no European product matches Palantir’s technical capability, that none matches its operational integration with allied forces, and that none carries its accumulated credibility with defense buyers. These have separate truth values, and “no alternative” is not typically deployed as an empirical claim about any of them. It is deployed as a procurement argument: a reason to skip the evaluation, extend the contract, adopt the framework agreement. It forecloses the comparison rather than winning it.

Events have been unkind to the headline version. Within the same month, Germany’s domestic intelligence agency, the BfV, selected ArgonOS, an AI analysis platform from the French company ChapsVision, over Palantir, after its president signaled a European-first course. In June, the French government went further: Prime Minister Sébastien Lecornu announced that the DGSI, France’s domestic intelligence service and a Palantir customer since the aftermath of the November 2015 attacks, will drop Gotham for ChapsVision in the name of strategic autonomy. Germany’s military cyber chief had already said in April that granting an outside vendor’s staff access to national databases was not something he could see happening. Whether ArgonOS matches Gotham operationally is a fair empirical fight that its own parliamentary supporters concede must be settled in use. But two domestic intelligence services choosing the alternative within weeks of a NATO commander’s “no real competitor” is a reminder of what the phrase does: it describes the procurement default, not the market. Vandier, to his credit, was more careful than the headline, distinguishing single-supplier dependence from infrastructure dependence from data control. The headline version is the one that travels through procurement committees.

One more sequencing detail from the German case deserves notice: the BfV’s ArgonOS deployment cannot be completed until Germany passes its planned intelligence-law reforms. Law first, then software. It is the inverse of the Baden-Württemberg order, and the fact that it reads as remarkable is a measure of how normalized the other sequence has become.

Brussels already runs on this

The EU institutions are not observers of the pattern. As Euractiv reported in June, EPSO, the office that runs recruitment for the EU institutions, is deploying a Job Matching Application built by Accenture on Anthropic’s large language model and hosted on AWS, used to identify, score, and rank candidates feeding into a Unified Talent Pool. Model, integrator, infrastructure: the full stack is American, assembled through a contractor chain in which no single EU decision-maker ever faced the question “should a US foundation model help sort applicants to the European civil service?” as a policy matter. The Commission’s position is that the tool remains a support tool with humans retaining final say, and the European Data Protection Supervisor’s assessment is ongoing; an EDPS official argued publicly this week that scoring candidates with a language model demands a legally and scientifically rigorous assessment it has not yet received. The scale is not hypothetical. EPSO’s last generalist competition drew 174,922 applications for a single reserve list. Under the AI Act the EU itself wrote, recruitment scoring is a high-risk application. The administration that authored the rulebook acquired the system through the layer the rulebook is least equipped to watch.

The layer question

The European sovereignty debate is conducted at the level of models, data residency, and regulatory text. The operative decisions are moving one layer down, through vendor relationships, framework contracts, validated architectures, and hardware refreshes, where the political content of a dependency is procedurally invisible. This is not an argument that procurement officers are doing their jobs badly. It is an argument that their jobs were scoped for a different question, and that the vendors bundling contested software into infrastructure understand the scoping better than the governments buying it.

The practical implication is a reporting one as much as a policy one. If the venue has moved, attention has to move with it. Framework agreements, call-off contracts, break clauses, reference architectures, and integrator relationships are now primary documents of technology politics. Watching procurement channels is a way of watching politics. Increasingly, it is the only way to watch it in time.