Core concepts
Compute Sponsoring
Quick answer
Compute Sponsoring funds consumer AI experiences through separate sponsored placements alongside the application flow.
The publisher keeps ownership of model output while configuration controls sponsorship defaults, pacing, consent, and safety.
Definition
The September 2026 paper by Christian Leinberger proposes linking an advertising exposure to an identifiable compute unit. A job can provide that reference: user intention, the generation process and its result. The paper develops a category definition, not a technical specification or certification.
Why it matters
Funding is one part of the argument. Because AI content is generated in the moment, an output alone cannot reliably show whether advertising influenced it. Coupling advertising to the compute process makes the separation addressable for publishers, users and advertisers.
Source and persistence
Semantic Source distinguishes no semantics (S0), user intention (S1), model-derived semantics (S2), and both sources (S3). Semantic Persistence distinguishes transient signals (P0), retention within a session (P1), and retention across sessions (P2).
Neutral is S0/P0. Situational uses S1, S2 or S3 with P0 or P1. Profiling uses semantic signals with P2. These are descriptive categories, not a list of wavebird features.
How to read the notation
CS-S (S1/P0) describes prompt-derived situational relevance without semantic persistence. The notation applies only when advertising is coupled to a compute unit and the proposed principles hold; a label or API flag does not establish that by itself.
Configuration
Configure sponsorship through project defaults, formats, pacing, consent, and safety policies. Job requests can still carry documented per-request overrides.
Publication
Current paper: Christian Leinberger, September 2026. The earlier February 2026 white paper was co-authored with Mario von Bassen.
Read the current paper →Publication record (DOI) →Earlier version (February 2026) →
Need help with your integration?
Share the affected endpoint, request ID, and the behavior you expected. Leave out keys and user content.