Tailscale CEO Avery Pennarun argues that cheaper AI compute can come with a catch: if an AI system locks users into one provider’s models, storage, networking, and other services, moving to another LLM might mean rebuilding half their infrastructure. He writes that the computing industry has created an AI infrastructure market where the underlying technology keeps getting cheaper but buyers never see the benefit, because suppliers know customers cannot easily leave. Pennarun cites Character.ai’s bankruptcy last month as a painful example: according to its bankruptcy filings, Character.ai was spending roughly $10 million USD ($14 million CAD) a month on compute in 2025. He says the team stopped developing its own foundation models, adopted open-weight alternatives, moved some workloads to cheaper infrastructure, and in one deployment doubled inference throughput, substantially reducing costs but failing to make the overall economics work. Pennarun notes that Character.ai managed to move some workloads to cheaper infrastructure, but whether a customer can take advantage of a better price depends on how much of an application has been built around its existing provider. He compares this to the mainframe era and to cloud computing, which he says offered a different approach to the same lock-in.