The cheapest computation is the one that never runs. Use AI to strip bloat out of software and stop the world from computing the same answers over and over.
For decades, software has expanded to fill whatever hardware exists, so apps that once ran in megabytes now need gigabytes. AI could reverse that by systematically rewriting bloated code to run on a tenth of the resources: heavy desktop apps built on web frameworks, legacy enterprise stacks, and unoptimized database queries. That frees up modern chips without salvaging a single device. The second half is a planetary memo cache. The same image gets resized, the same page rendered, and the same query answered millions of times a day. A global “has anyone already computed this?” layer would turn repeated computation into a lookup. Old devices are a natural fit for this layer, since caching needs plenty of storage and very little processing power.