The NHS Foundry Lock-In Problem: Why Palantir's GP Records Lobby Exposes a Deep Architecture Bet
WhatsApp messages obtained under FOI show Palantir privately pushing to extend its £330m NHS data platform into GP records it was explicitly contracted to exclude. The engineering story underneath is about ontology lock-in, tenant architecture, and what it costs to leave.

Something quietly extraordinary happened this week in British healthcare technology. Documents released under a freedom of information request showed that Palantir's UK head, Louis Mosley, messaged a health minister over WhatsApp in February 2026, urging the government to bring GP records into the NHS Federated Data Platform, a data category the firm's own contract with NHS England explicitly excludes. The lobbying, reported by The Guardian on 9 November and confirmed by multiple outlets by 11 November, landed while the contract's future is under active ministerial review and the government simultaneously prepares to put the platform out to competitive tender. It is a remarkable situation. But the more important story sits in the architecture.
What the FDP Actually Is
The NHS Federated Data Platform is not a single national patient database. That framing, common in public debate, misreads the design. As Palantir's own published documentation explains, the FDP is built on its Foundry software and operates with a federated architecture: each participating NHS trust and integrated care board runs its own separate Foundry tenant, acting as its own data controller, while a national tenant provides coordination. Think of it less like a central reservoir and more like a network of local water tanks, each with its own valve, connected by a shared pipe standard. Data stays where it lives; the platform provides a consistent query and integration layer on top.
Foundry's architecture rests on what Palantir calls an ontology: a single, standardised representation of entities, their relationships, and the logic and security policies that govern how applications built on the platform interoperate. This is what Computer Weekly's in-depth technical series described as the platform's genuinely differentiating feature. Unlike a conventional data warehouse, Foundry's ontology handles defined actions that create new data through transactions, with validation rules, access controls, and audit trails managed by the platform itself. The Canonical Data Model in the NHS deployment is the NHS-specific instantiation of this, a unified schema that makes locally built applications portable across trusts.
That elegance is also the trap.
The Lock-In Geometry
Medact's 2026 technical report on what an exit from the FDP would actually involve put it plainly: Palantir owns the "Specially Written Software" within the platform under the contract terms, and applications built on FDP are built on Foundry. Exit means rebuild. The House of Commons Science, Technology and Innovation Committee, in its June 2026 report, recommended the government exercise the February 2027 break clause and publish a fully costed exit plan by end of 2026. That plan, as of this writing, has not appeared.
Parliamentary debate in April 2026 raised the specific technical mechanism. Because each trust's internal systems are connected to Foundry via custom connector software written trust-by-trust, Palantir-owned code now sits inside every participating NHS organisation. One MP described this as Palantir embedding its architecture as a deliberate business tactic. The more neutral engineering reading is that any platform requiring custom connectors to heterogeneous legacy systems will produce exactly this dependency profile. SAP has done it for decades. The question isn't whether lock-in exists; it's whether it was adequately priced in.
The GP Records Question, Technically
So why GP records? The answer is in the data flow logic. The FDP's current scope covers secondary care: hospital waiting lists, theatre scheduling, discharge management. GP records, held in a separate set of systems, most notably EMIS and SystmOne, represent primary care, the entry point for most patient journeys. Including them would complete the longitudinal patient view that the platform's ontology model is structurally designed to handle. From a pure architecture standpoint, adding GP data to the Canonical Data Model is the natural next step; the pipes are already in place, it's the valve that is locked.
The political dimension is that the previous government explicitly ruled GP records out when it announced the FDP contract in 2023. The FOI messages show Palantir acknowledging this directly: internal notes said ministers would need to "get ahead of" critics if the scope were to expand. The National Audit Office is currently reviewing the rollout. NHS England claimed, as of June 2026, that 139 trusts were "live" on the platform; The Week's reporting on 10 November found that 29 trusts listed as benefiting from Palantir's software say it is not in use. The gap between reported adoption and actual utilisation is itself an engineering signal worth watching.
The Sovereign AI Complication
Health Secretary Yvette Cooper told Parliament last week that she saw "a huge long-term benefit to using the NHS to develop UK sovereign AI," a framing that recast the FDP debate inside the government's broader Sovereign AI initiative. DSIT has committed £500 million to that programme, with a separate £100 million R&D procurement scheme opened in September 2026, explicitly targeting NHS productivity as one of its challenge areas. The contract's next endpoint is 2029.
The tension here is structural. The Sovereign AI argument cuts both ways. On one reading, extending and deepening the Palantir relationship is the fastest route to working AI on NHS data at scale: the infrastructure is already deployed across 139 trusts, the ontology model exists, and replacing it would take years and significant capital. On the other reading, building sovereign AI on a proprietary US-vendor ontology, with IP owned by that vendor and custom connectors embedded in every trust, reproduces the exact dependency the sovereignty argument is meant to avoid. You can't be sovereign on someone else's Foundry.
The government says it will put the FDP contract out to tender, forcing Palantir to rebid against UK-based competitors. What those competitors would actually be competing on is worth spelling out: not just the application layer, but the right to define the Canonical Data Model, the ontology, that all NHS AI products will be built on top of for the next decade.
The Trademark Signal
There is a parallel data point from the intellectual property market. AIBD analysis of UK Intellectual Property Office trademark filings (TMD data) shows just 620 Class 38 filings in Q4 2026, down 50.3% on the prior period. Class 38 covers telecommunications services, the connectivity layer where data-platform brands typically plant their flags. That contraction suggests some chilling of brand registration activity in precisely the infrastructure category the FDP debate sits inside. Whether founders are waiting for the tender outcome, or the market is consolidating around a smaller number of serious platform plays, is an open question. It is a number that people building UK-sovereign data infrastructure should be watching.
The Six-to-Twelve Month View
If the February 2027 break clause is exercised, the government faces an immediate architectural decision: extend on modified terms, run a parallel migration, or accept years of disruption during transition. The tender process, if it proceeds, will effectively be a public competition to define the NHS data ontology. That is not a software procurement. It is a standards-setting exercise with clinical consequence.
The GP records question will not go away. The longitudinal patient record is the genuine prize for anyone building AI-driven diagnostics or population health management. Whoever controls the schema that connects hospital discharge data to GP records to prescribing history controls the training data for the next generation of NHS AI. The WhatsApp lobby was inelegant. The underlying engineering logic was not.
