What moved power in late 2026 was not a model release. It was deployment topology: who physically holds the servers. Six events between October 5 and November 18 rank almost entirely by that criterion.
Rank them by one question: how much does each event change the set of vendors a European defence buyer can realistically sign with in 2027? On that test Helsing’s Clarity comes first, because it runs on customer-owned infrastructure. The Pentagon’s removal of Anthropic comes second. Britain’s £90M naval package, real money with a named mission, comes last.
Three things happened in that window that together changed what European defence buyers are choosing between. The Pentagon stopped using Anthropic products on October 5 after designating the company a national-security supply-chain risk. Palantir executive Louis Mosley told Reuters on October 21 that neither Palantir nor the US government could remotely disable the Maven Smart System. And Helsing unveiled Clarity on November 11 in Berlin, running on customer-owned infrastructure.
The criterion is not headline size and not contract value. Each item is ranked by how much it constrains or expands the buyer’s options at signature time: who can host, who can revoke, and who can be removed from the chain by a government that is not yours. A $1.275B ceiling matters less here than a sentence in a usage policy, if that sentence removes a supplier from your shortlist.
Why Clarity’s hosting model beats Maven’s contract value in this ranking
Helsing ships Clarity on customer-owned infrastructure
The first credible European answer to Maven that does not require trusting a vendor’s promise, because the hosting question is answered by the architecture instead.
The Pentagon drops Anthropic as a supply-chain risk
A frontier lab was removed from a defence stack by its own government in a single announcement. That is the scenario European buyers spent the year worrying about, demonstrated in reverse.
Palantir publicly denies a NATO kill switch
Reassurance offered at the level of corporate statement rather than architecture, which tells you the question had already become load-bearing in European capitals.
Anthropic’s usage policy extends the weapons ban to control software
A commercial lab drawing a line through drone arming and autonomous vehicle control narrows what any integrator can legally build on top of that model family.
China’s Announcement No. 49 licences high-power drone motors
Software sovereignty is worth less when the airframe stops at customs. This is the hardware floor under every drone AI programme.
Britain’s £90M naval drone and AI package
Real money with a named mission, but it buys capability from the existing supplier base rather than changing who can supply it.
Clarity tops the list because of what it removes from the negotiation. According to Reuters coverage of the Berlin launch, the platform runs hundreds of AI agents across drone, satellite, sensor, report and open-source data on a shared military ontology, identifying assets and preparing target handoff to a weapon, with human approval required. Co-CEO Gundbert Scherf put it plainly: “a capability this critical must be sovereign.” The deployment model is the product claim. If the software sits on your hardware, the question of whether a foreign vendor can switch it off becomes a question about your own data centre.
What Helsing did not say matters too. No customer named, no pricing disclosed, no operational accuracy results. Der Spiegel reported that the Bundeswehr is testing Clarity, but that is a journalist’s sourcing rather than a vendor’s announcement, and BattlePolicy’s write-up makes the same point. Against Maven, which reached full technical operational capability at NATO’s SHAPE headquarters in June 2026 and now has Marine Corps organisations issuing individual user accounts under an administrative message effective October 1 (Marine Forces Reserve Command included), Clarity is a launch event with an architecture behind it. That is not the same as a fielded targeting system with accounts provisioned down to reserve commands.
Why the Anthropic removal ranks above the kill switch denial
The Anthropic case is the only item here where a government actually cut a model supplier out of a defence stack, and it did so in one step. A Department of Defense official told the BBC on October 5 that the department “has ceased the use of Anthropic products” after designating the company a national-security supply-chain risk.
Read that next to what Anthropic published three days later. The 2026 Usage Policy update, effective November 12, states the company had “seen people attempt to use our models to build guidance and control software for weapons,” and extends the weapons prohibition to software and components that make weapons operate, explicitly including “arming drones and other autonomous vehicles.”
So a government removed the supplier, and the supplier simultaneously tightened the policy that constrains the use case. Both directions of severance happened inside one week. For a European buyer, that is the concrete version of the abstract dependency argument: a model provider in your stack can be removed by a state you do not vote in, or can remove itself by changing a document you do not negotiate.
I rank the Anthropic events second because they are evidence rather than assurance. Palantir’s denial asks you to believe a statement about capability. The Pentagon’s action shows you what happens when a government decides a supplier is a risk. Advising anyone on vendor risk in Europe, I would weight demonstrated severance far above denied severance, regardless of which company it happened to.
Can a US vendor switch off military AI it sold to NATO?
On the sourced record, Palantir says no, and nobody has published an architectural demonstration either way. Mosley told Reuters on October 21 that neither Palantir nor the US could remotely disable the Maven Smart System used by NATO HQ and several European militaries.
I take that at face value as a description of the current build. My concern is different, and it is a reasoning step rather than a reported fact: a kill switch is the least interesting failure mode. Software that reached full operational capability in June 2026 and runs to a contract completion estimate of May 28, 2029 needs updates, support engineers and export authorisation for all of it. Denying a remote off switch says nothing about what happens when a licence renewal is withheld, or when a support contract is not extended. Those are slower levers, and they are harder to deny publicly because nobody frames them as a switch.
The procurement record already reflects this. Germany and France declined Maven for their own national forces, citing dependence on US technology, while NATO bought it for SHAPE in 2025. A May 20, 2025 Pentagon modification raised the Maven ceiling by $795M, from roughly $480M to $1.275B. Two European states looked at that programme and said no on sovereignty grounds, before Clarity existed as an alternative. That is the demand signal Helsing launched into.
I have written before about how concentrated military AI contracting actually is, and the concentration is the point here: European buyers are choosing from a supplier base shaped by one government’s budget.
The drone motor licence is the dependency nobody ranks highly enough
China’s Announcement No. 49, issued October 19, revises dual-use export controls across 13 categories. From November 18, qualifying high-power drone motors require an export licence.
I rank this fifth, and fifth is probably too low. It sits below the software items because the licensing regime is new and its effects on specific European programmes are not yet documented in anything I can cite. But the logic is uncomfortable. A sovereign targeting platform running on your own servers, handing off to a drone whose motor needs a licence from Beijing, has moved the dependency rather than removed it.
Britain’s package is the clearest case of money that buys capability without shifting power. On November 10, Britain announced £90M ($121M) for drones, sensors and AI to help the Royal Navy track Russian submarines in the North Atlantic. Specific mission, specific domain, no change to who can supply it.
The three questions I would put in a 2027 defence AI questionnaire
Across all six items, every form of control a European buyer wants turns out to be a question about physical and legal custody rather than about vendor intent. Hosting location, model licence terms, component export licences, support contract jurisdiction. None of those are answered by a reassuring interview.
So: where do the weights and the inference run, and can you operate for twelve months with no vendor network connectivity? Under which jurisdiction’s export regime do model updates fall, and what is the documented process if an authorisation is refused? Which physical components in the kill chain, motors included, are subject to a third country’s licensing, and what is the substitution plan? The Palantir and Anthropic events of October answer none of these for you. They do demonstrate why asking them is now reasonable rather than paranoid.
The uncertainty I cannot resolve: Clarity has no published accuracy results and no named customer, while Maven has operational capability at SHAPE and provisioned accounts down to reserve commands. **A platform that runs on your servers but performs worse at identifying assets is a worse targeting system with better press.** The Bundeswehr test Der Spiegel reported is the thing I would want results from, and until those exist, this ranking is about who can supply you rather than who performs best. If you are sitting on a 2027 procurement decision and want a second opinion on the dependency questions rather than the feature matrix, that is a conversation worth having before the RFP text is locked, because hosting and licence clauses are much harder to renegotiate afterwards.
