long-running agents are changing the economics of compute i believe they will also have an outsized impact on AI infrastructure financing: a long-running agent is one that keeps inference demand running independently of human prompts. long-running agents turn spiky user-prompt inference workloads into baseload. current workloads vary enormously in utilisation: a chat app spikes and goes quiet, whereas an agent grinding through every contract, every vuln, every customer record just runs flat out around the clock until its task is complete. baseload is the profile debt investors and banks actually know how to underwrite. predictable utilisation is a financeable asset in a way spiky consumer traffic never was. the more long-running agents are able to operate autonomously, the more AI compute starts to look less like speculative capacity built ahead of demand and more like infrastructure with contracted, persistent utilisation. that changes the financing equation. higher and more predictable utilisation means more predictable cash flows, greater debt capacity, lower cost of capital, and ultimately more infrastructure that can be built against a given amount of equity. the biggest impact of agents may therefore not just be how much compute they consume but more how financeable they make that compute.