Can I Buy Your KV Cache?

Researchers propose a marketplace for precomputed key-value (KV) caches in large language models, where publishers can sell the right to reuse cached computations on identical documents rather than having every AI agent recompute them from scratch. The approach achieves 50x cost savings compared to redundant prefill computation and is token-exact with no accuracy loss, though hosting the caches provider-side is necessary since transmitting them over networks is prohibitively expensive. The economic opportunity is substantial—serving a popular document to millions of agents could generate millions in provider margin while offering users a 10x discount on inference costs.