The paper envisions an agent-native CDN for prefill, while acknowledging unresolved issues such as lossless KV compression and a cross-party payment/royalty mechanism.

Loadable KV caches face practical hurdles because KV data is nearly incompressible, making per-load egress costs potentially higher than prefill savings; hosting caches provider-side can mitigate this by removing egress.

The work lays groundwork for an agent-native CDN architecture that reduces redundant computation and scales, yet faces challenges including compression and cross-party monetization.

Proposals include shipping KV caches via provider-side hosting (like prompt caching) to avoid egress costs and enable a vendor-friendly model through API tariffs for cache reads.

Publishers would precompute a document’s KV cache and allow agents to load it on demand, achieving exact equivalence to on-the-fly prefill in tokens.

Experiments with Qwen3-4B show reuse is 9-50x cheaper in compute than redoing prefill, with bigger savings for longer documents due to prefill’s O(L^2) attention cost.

The Compute It Once paradigm proposes precomputing a document’s key-value caches once and licensing their use to multiple AI agents, eliminating repeated prefill steps and enabling token-exact results.

This approach aims to optimize AI agent performance by monetizing precomputed KV caches, reducing redundant computation across agents.

Reusing precomputed KV caches can deliver token-exact results with no accuracy loss, enabling substantial compute savings and better scalability as document length increases.

A concrete example shows serving a 3,774-token document to 80 million agents costs about $1.5 million with re-prefill versus only about $0.03 million with reuse, illustrating nearly 50x savings.

On models like Qwen3-4B, reuse yields substantial compute savings, with the efficiency gap widening as document length grows because of quadratic attention scaling in prefill.

ArXiv publication details the concept and notes related startup activity and potential implications for AI agent deployment and cost efficiency.