Caches are only so large. Expanding them doesn't buy you much, and increases costs greatly.
The key benefit to a cache is that a small set of content accounts for a large set of traffic. This can be staggeringly effective with even a very limited amount of caching.
Your options are:
1. Maintain the same cache size. This means your origin servers get far more requests, and that you perform far more cache evictions. Both run "hotter" and are less efficient.
2. Increase the cache size. Problem here is that you're moving a lot of low-yield data to the cache. On average it's ... only requested once, so you're paying for far more storage, you're not reducing traffic by much (everything still has to be served from origin), and your costs just went up a lot.
3. Throttle traffic. The sensible place to do this IMO would be for traffic from the caching layer to the origin servers, and preferably for requesting clients which are making an abnormally large set of non-cached object requests. Serve the legitimate traffic reasonably quickly, but trickle out cold results to high-demand clients slowly. I don't know to what extent caching systems already incorporate this, though I suspect at least some of this is implemented.
4. Provide an alternate archival interface. This is its own separately maintained and networked store, might have regulated or metered access (perhaps through an API), might also serve out specific content on a schedule (e.g., X blocks or Y timespan of data are available at specific times, perhaps over multipath protocols), to help manage caching. Alternatively, partner with a specific datacentre provider to serve the data within given facilities, reducing backbone-transit costs and limitations.
5. Drop-ship data on request. The "stationwagon full of data tapes" solution.
6. Provide access to representative samples of data. LLM AI apparently likes to eat everything it can get its hands on, but for many purposes, selectively-sampled data may be sufficient for statistical analysis, trendspotting, and even much security analysis. Random sampling is, through another lens, an unbiased method for discarding data to avoid information overload.
