There's a new currency in advertising, and it isn't media impressions or billable hours. It's tokens. AI tokens, specifically — the computational units that power every ChatGPT prompt, every Midjourney image, every Claude conversation. And agencies are having a hell of a time figuring out who should pay for them.

If you've used any AI tool, you've consumed tokens. Type a prompt, that's tokens. Get a response, more tokens. Generate a Coca-Cola holiday ad that requires 70,000 prompts? That's millions of tokens, and at even fractions of a cent each, it adds up fast.

The question facing every agency right now: do you eat those costs, or pass them through to clients?

The Eat-It Camp

Full-service agency RPA absorbs token costs entirely. Lisa Herdman, their SVP of video investments, put it bluntly: "We're still in the testing phases. We can't expect to charge our client for something we don't know is actually going to work for them."

There's logic here. AI is still experimental for most agencies. Charging clients for compute costs on tools that might not deliver feels like a money grab — which is exactly what Chris Neff at Anomaly called it. "It feels like a money grab," he said, declining to pass through AI overheads.

The problem? As AI becomes more central to agency work, eating those costs becomes increasingly expensive. And agencies already operate on thin margins.

The Pass-Through Camp

At Merge, token costs go to clients on a metered basis. Big Spaceship treats AI compute like any other production cost — catering, equipment rental, token usage. Line items on an invoice.

Silverside AI, the production studio backed by Pereira O'Dell's founders, uses a subscription model with seat-based pricing. "There's no one model that's the silver bullet," explained co-founder Johnny Rohrbach. "You have to use a lot of different models and all of those have different token economics."

The sophisticated players are negotiating bulk deals with AI providers. Brandtech's Pencil platform charges clients for "generation credits" — one credit equals one chat response, one image, or one second of video. Volume commitments let them negotiate better rates with Anthropic and OpenAI, then pass the savings on.

The Principal Media Problem

Here's where it gets politically uncomfortable. If agencies can buy tokens in bulk at a discount, what stops them from marking those costs up when passing them to clients? It's the principal media buying debate all over again — the practice that's been one of advertising's most contentious issues for years.

Industry observers aren't too worried though. James Londal, founder of consultancy Avelin, argues the bigger economic shift isn't token arbitrage: "It's labor compression. If AI runtime meaningfully replaces reporting teams, duplicated dashboards and middleware connectors, the financial upside dwarfs any marginal bulk discount dynamic."

Translation: Don't worry about agencies making money on tokens. Worry about them using AI to do the work of your entire media planning team.

The Coming Audit

Ebiquity CEO Ruben Schreurs already sees the future: token audits. "If it's part of the contract, then yes, we'd audit that," he told Digiday. Just as Ebiquity audits media spending, they'll eventually audit AI compute costs.

This is probably healthy. Transparency has always been a problem in agency-client relationships, and AI adds a new layer of opacity. Clients deserve to know what they're paying for.

But Schreurs also offered a useful reframe: "It's less about the input and more about how incremental business returns are being delivered." In other words, maybe we should stop obsessing over the cost of tokens and start asking whether the AI work is actually driving results.

That's the real question nobody wants to answer yet.