Could the company running a cache quietly tune how often prices refresh to boost its own profit, not your search?

This explores whether whoever controls a cache can set how fresh or stale it is to suit their own interests. The corpus has no study that tests this directly. It does have several pieces that point toward yes, and they come from different directions. The clearest case is travel. AI agents now run close to 200,000 searches for every ticket sold, so Amadeus has stopped pricing every query live. It now precomputes and caches fares for common routes and dates Does caching fares give infrastructure providers pricing power?. The detail that matters is that whoever decides what gets cached, and how often it refreshes, also decides which offers an agent can see. An agent that only sees the cache can't tell a stale price from a real one. The refresh schedule quietly becomes a way to shape the market.

A formal economics model in the corpus shows why a platform might use that power. When an AI platform charges for 'inspection tokens', meaning the cost of an agent looking closely at a seller, the profit-maximizing price pushes buyers to inspect more than they need to Does platform pricing of inspection tokens encourage excess search?. Consumers end up no better or worse off. Sellers lose, and the surplus moves to the platform. Apply that logic to caching: if fresh lookups are billed and cached answers are cheap, a platform has a reason to set refresh rates and pricing so that agents keep paying for 'fresh' checks. The opposite also works. A platform can serve stale cached data to cut its own compute costs while still charging as if it were live.

The same tension shows up inside AI infrastructure. Prompt caching cuts agentic API costs by 41–80% across OpenAI, Anthropic, and Google Does prompt caching actually reduce agentic LLM costs?. But the provider sets the rules: how long a cache lasts, what counts as a cache hit, and how cache reads are priced compared with fresh input. Those rules matter a lot. In one 115-day case study of a persistent agent, 82.9% of all tokens were cache reads Do persistent agents really cost less per token?. When most of your bill is cache reads, a small change in cache lifetime or cache-read price changes your costs more than the headline per-token price does. Meanwhile, remote storage of cached model state now loads faster than recomputing it Can remote storage beat recomputing cached model states?. That means providers can keep cached state around for longer and more cheaply, so the cache lifetime they give customers is increasingly a business decision rather than a technical limit.

There's a limit on how far a platform can push this, and Netflix's recommendation work shows it. Netflix found that adapting recommendations mid-session improves ranking by 6%. But precomputing those recommendations is impossible because the signals arrive in real time, so the system has to recompute them, which costs more and makes bugs harder to reproduce How can real-time recommendations stay responsive and reproducible?. Freshness has real value. A platform that lets its cache go stale to save money will eventually give worse answers, and users or agents will notice. So the freedom to set refresh rates in your own favor is greatest where users can't easily tell whether data is fresh, and agents reading cached fares fit that description well.

The takeaway you might not have expected: in agent-driven markets, cache refresh logic is starting to behave like pricing power. It is set by infrastructure providers rather than sellers, and the agents relying on it mostly can't see it. The corpus identifies this lever but doesn't yet have empirical work on whether platforms are actually using it strategically. That gap is worth watching.