Inference as Infrastructure

Utilities share characteristics: the output is fungible, suppliers are interchangeable, consumption is metered, and buyers optimize for price, reliability and jurisdiction rather than for brand. Inference increasingly fits. A summarization request produces comparable output across providers; the meaningful variables become cost per token, latency, where the computation happens and what the supplier retains. Organizations that recognize the shift early build for substitution; those that treat model choice as strategic keep re-litigating a decision that matters less each quarter.

Centralpoint procures inference this way. The model is selected per request, consumption is metered per execution and per skill, cost can be consolidated onto a single invoice below published rates, and locality is a deployment decision — including embedded local models where the workload cannot leave the environment at all. The governed cache means the most repeated work stops consuming the utility entirely.