There is a particular kind of confidence that attaches itself to money spent on things that do not yet exist. When Jeff Bezos and the government's sovereign AI fund put capital into CuspAI last week, valuing a two-year-old Cambridge company at $2.6bn on the promise that it might one day discover materials nobody has yet identified, the tone from ministers was unmistakably buoyant. Liz Kendall spoke of paving the way for breakthroughs. The company's founders spoke of a single challenge holding back fifty years of industrial progress. Nobody asked what would happen if the search engine for rare materials simply failed to find anything, because the failure, if it comes, is years away and diffuse enough not to attach to any one patient, any one ward, any one week of winter pressures.

That is worth sitting with, because the government now runs two very different experiments in state-backed artificial intelligence, and it treats them with two very different temperaments. The sovereign AI fund has already put money behind Isomorphic Labs, the drug discovery venture spun out of Google DeepMind, on much the same logic that produced the CuspAI deal: back the science, tolerate the uncertainty, let the returns arrive later. That is a defensible model for research capital. It becomes harder to defend as a template once the AI in question is not designing a hypothetical medicine but sitting across live patient records in two hundred hospital trusts.

The Federated Data Platform, the £330m data infrastructure built with Palantir that NHS England signed in 2023, remains the clearest illustration of how little patience the system extends to AI once it touches an actual ward. The Health and Social Care Committee has this month written to ministers urging them to prepare to drop Palantir when the contract's break clause arrives in February 2027, citing what it called serious mistrust among both the public and the medical profession. Greater Manchester's integrated care board, covering 2.8 million people, has refused to join the platform at all. More than a hundred NHS data and digital staff signed an open letter warning that the current arrangement risks patient trust, data quality and the long-term sovereignty of NHS infrastructure itself. None of this reflects a health service hostile to technology. It reflects one that has learned, through the FDP experience and before it, that AI infrastructure imposed from outside a trust's own governance tends to arrive with the benefits unproven and the risks concentrated on people who never signed up for them.

The asymmetry is not hypocrisy so much as a fairly rational response to where the consequences actually land. A materials-discovery failure at CuspAI costs investors money and delays a chip supply chain that most voters will never think about. A data platform that degrades scheduling accuracy, as internal figures reportedly show has happened in a meaningful share of hospitals using the FDP's tools, costs patients appointments and costs clinicians the working assumption that the systems around them are actually helping. Ministers can afford to be swashbuckling about the first kind of AI precisely because nobody will ever trace a missed diagnosis back to a materials science bet. They cannot be equally swashbuckling about the second, and to their credit, mostly they have not been.

The problem this leaves the NHS with is not too much caution but too little institutional machinery for exercising it well. Scrutiny has arrived mainly through select committee letters, open letters from staff, and freedom of information requests rather than through a procurement and evaluation process built from the outset to interrogate AI claims before they are signed off. The government's appetite for risk on CuspAI shows it can move with genuine speed when the politics are favourable. The unresolved question, with the FDP's break clause now eighteen months out, is whether it can build the same speed into deciding, with evidence rather than momentum, what belongs inside the NHS and what does not.