What is Compute Sponsoring?
Proposal in brief
Compute Sponsoring couples a clearly labeled ad exposure to a unit of AI compute and defines the constraints for consent, separation, no influence, and proof. Showing an ad during the wait window is one manifestation, not the entire definition.
Why Compute Sponsoring exists
GenAI has a unit-cost problem: every prompt consumes compute during inference. Costs scale with usage.
If a product runs a meaningful free tier, it needs a variable revenue stream. Advertising is one option, but only works if influence and privacy are bounded.
Compute Sponsoring is the proposed constraint set for coupling advertising to compute: keep ads out of model I/O, require consent for semantic relevance, and make every billed impression auditable.
The first visible reference implementation is often a placement shown during inference wait time, but the category is broader than that surface alone.
Classic ad standards like OpenRTB cover auction mechanics, not AI-specific separation, consent, or proof.
Compute Sponsoring v1.0 proposes these requirements so ad-funded inference can be transparent and verifiable.
The Principles
These principles are architecture requirements, not policy guidelines. For separation, the ad delivery path is isolated from model I/O at the infrastructure level so they cannot interact.
Separation
ad delivery is isolated from model I/O at the infrastructure level
No input influence
sponsor information never enters the prompt or context window
No output influence
ads never cause prioritizations, omissions, or rephrasings
Data sovereignty
sponsors never receive user-level semantic data
Advertising transparency
ads are always clearly and fully labeled
Consent-based targeting
semantic relevance targeting requires explicit user consent
Classification
Compute Sponsoring classifies systems along two dimensions. Semantic Source, S0 to S3, describes what data is used to target ads. Semantic Persistence, P0 to P2, describes how long that data is stored. Together, they define three profiles:
CS-N
Compute Sponsoring Neutral
No semantic data is used for ad targeting. Ads are matched based on system metadata only, like language or time of day. This is the least invasive profile. Example: a coding assistant that shows untargeted banner ads during inference.
CS-S
Compute Sponsoring Situational
Semantic data from the current prompt is used to match ads, but nothing is stored beyond the current request or session. This profile allows relevant ads without building user profiles. Example: a travel chatbot that shows flight deals when the user asks about weekend trips, but forgets the topic on the next request.
CS-P
Compute Sponsoring Profiling
Semantic data is accumulated across sessions. This enables profile-based advertising and requires the strongest consent and transparency mechanisms. Example: an AI assistant that learns user preferences over time and targets ads based on accumulated interests.
wavebird operates as CS-S (S1/P0): situational relevance based on prompt topic, with no persistence of semantic prompt-derived targeting profiles across requests. Semantic targeting requires explicit user consent.
How wavebird implements it
wavebird is infrastructure and a reference implementation built on top of Compute Sponsoring v1.0. It currently operates as CS-S (S1/P0) for semantic prompt-derived targeting profiles: situational relevance from the current request, with no persistence of those profiles across requests.
The delivery signal is produced by the data firewall (topic category + language). Raw prompts are not shared by default; if prompt sharing is explicitly enabled, prompts are sent only to wavebird's firewall for ephemeral classification and are not sent to SSPs, DSPs, or advertisers.
Consent gating is explicit; without consent, delivery falls back to CS-N. Failure behavior is fail-closed (see Safety).
Within that stack, CSI is the proof and audit subsystem. Compute Sponsoring is the category; CSI is one product subsystem used to make it verifiable.
Read the full proposal
Read the Compute Sponsoring v1.0 white paper for the full matrix, notation, and disclosure rules.
Feedback and critique are collected on GitHub.
Related
Last updated: 2026-04-20