I couldn’t agree more with Richard Whittle here. The dependencies that can be established here are deeply concerning because if AI integration can’t be rolled back, spending will have to be cut elsewhere when the pricing shock hits:
For Manchester’s public institutions including Manchester City Council, GMCA, NHS trusts, housing associations, Further Education colleges, and the city’s universities, the AI pricing problem has direct and urgent implications. These institutions are, for reasons of scale, civic visibility, and political ambition, highly attractive customers for AI vendors. Greater Manchester has been nationally prominent in its commitment to digital and AI modernisation. The region’s public bodies have embraced AI as a component of service delivery, efficiency improvement, and regional economic leadership.
Public institutions that embed AI services at today’s subsidised prices and then face cost-recovery pricing during the transition period may experience significant operational disruption. Budget commitments made on the basis of current AI pricing may prove undeliverable under true-cost conditions. And the services that may be cut to accommodate rising AI costs are, in many cases, precisely the services on which Manchester’s most disadvantaged residents depend.
https://markcarrigan.net/wp-content/uploads/2026/09/d9851-11.05manchesterresilientaireportv11.pdf
He argues that resilient adoption is possible. This isn’t an argument against integration but rather to predicate that integration on building capabilities to negotiate the risks inherent in the pricing shock:
… including procurement that is evidence-based, contracts that include robust exit provisions, spending decisions that are stress-tested against true-cost pricing scenarios, governance structures that can assess value for money with substantive independence from the political incentives that drive procurement in the first place, and investment in the technical capability and process management required to switch providers or models over time where quality or cost considerations demand it.
This is a robust economic analysis of what Ed Zitron referred to (in a metaphor that didn’t quite work if you dwelled on it too long) as ‘subprime AI‘. There is a disaster waiting to happen here and we urgently need to think about what inference rationing will mean for organisations which are seizing on LLMs, at the high point of investor subsidy, as a means to square the fiscal circle they find themselves confronting.
