Advertising built for AI.

AI products create value through model output. Advertising should help fund that compute without becoming part of what the model says. wavebird keeps the advertising path separate and builds new formats around the moments AI products already create.

When an AI model generates an answer, the user is already waiting for compute to finish. That creates a natural advertising moment that does not exist in a traditional page view.

An inference-time ad uses that existing generation window. The ad can appear while the model is working, while the AI response remains on its own path.

The wait already exists. The ad does not need to create it.

Inference-time advertising is one format, not the whole category. AI products create several moments where sponsorship can be useful without becoming part of the answer.

AI products earn trust through their output. Advertising should not borrow that trust by looking like another sentence from the model.

A text ad placed too close to an AI answer can be mistaken for a recommendation, conclusion, or factual statement generated by the product.

Visual advertising has its own language: brand, color, imagery, product, motion, and identity. It can be clearly recognizable as advertising without pretending to be part of the answer.

Users should always be able to tell what comes from the AI product and what is sponsored.

Compute Sponsoring defines a way to associate a sponsorship exposure with a specific unit of computation while keeping the advertising channel separate from model I/O.

The published Compute Sponsoring working paper defines a compute unit as an identifiable unit of active compute. Depending on the system, supported examples are:

That creates something traditional display advertising does not naturally provide: a measurable relationship between an advertising exposure and the compute moment it helped fund.

CPM and CPC remain useful advertising metrics. But AI products also create units that publishers already measure in their products: generations, tokens, compute cost, and time spent waiting for model output.

A measurement vocabulary can connect advertising with the units AI products already measure.

Advertisers need buying metrics. Publishers need a different answer: how much of the cost of running the AI product can advertising actually support?

At 70% Compute Coverage, advertising funds 70% of the measured compute cost in this example.

Above 100%, advertising revenue exceeds the measured compute cost for that scope.

Traditional advertising tells a publisher where an ad appeared. AI-native measurement can also show which compute moment the sponsorship was associated with and how that changed the product economics.

Illustrative example only. Values do not represent a guaranteed or current wavebird benchmark.

We believe AI advertising should be designed around the economics and product experience of AI itself, not copied from pages and feeds.

That means formats built around generation moments, clear separation between advertising and model output, and measurement that connects sponsorship with the compute it helps fund.

Inference-time ads are one part of that model. Rewarded compute, sponsored generations, and AI-native economic metrics extend the same idea: advertising can help fund AI without becoming the AI.