Compute is about to become the next big commodity in advertising.
The economics demand it. AI has gotten expensive enough, fast enough, that holdco CEOs are now being asked outright how funding it won’t come at the expense of staying profitable. The IAB Europe’s chief economist Daniel Knapp put it plainly: “No one really knows quite yet how to price that in, so you’re seeing the business sort of try and come up with solutions to try and figure that out.”
One answer being floated: turn the holdco itself into a futures market for tokens, whereby they buy them in bulk ahead of demand, price the risk, and sell it on with a margin attached. Some are already doing it, baking token costs into principal media deals.
It’s the same model holdcos have run for years across media and data. Compute is just the latest thing running through that same machine. And just like those other times, the approach has divided opinion. In a LinkedIn poll conducted by Digiday last week, 42% of 58 respondents said agencies shouldn’t become futures markets for tokens at all. Fewer than 10% called it a legitimate new revenue opportunity. The largest swing vote — 36% — landed on maybe, conditional on transparency.
Here’s the case for and against.
Case for
Whether you buy this depends on how much sympathy you have for agencies right now. They’ve spent close to two years building AI infrastructure at group level, largely carrying the running costs on their own balance sheets rather than billing for them. If they hadn’t, they’d have been stuck justifying a new line item to clients who already assume AI means the bill goes down, not up.
That arrangement had a shelf life, and it ran out this year, as agencies reached the point where they could no longer afford to keep absorbing those costs unrecovered. Not as the push for more AI adoption and increasingly autonomous agents drove token consumption higher and higher. That capex needs to show up somewhere on the other side of the ledger eventually, or shareholders start asking why it hasn’t.
So the holdcos went looking for a way to offset it all. Turns out they didn’t have to look too far. Principal media, as already mentioned, was already how agencies put a number on uncertainty and got paid to carry it. Token costs are being handed the same treatment, because nobody can say with confidence where they’ll land next quarter, let alone next year, and somebody has to hold that exposure so the client doesn’t have to.
It also isn’t pure speculation on the agencies’ part — token prices are widely believed to be subsidized right now, the same way early Uber and Amazon pricing was, and expected to rise once that correction comes. Buying ahead of that curve is a hedge as much as a sales tactic.
“I think it’s a big bargaining chip to walk in and say, ‘Hey, I can give you a better deal on both compute and media, and that’s why you should work with me, and that’s why you should pay my fees,” said Ana Milicevic, co-founder of consultancy Sparrow Advisers.
In that regard, agencies will need all the help they can get. When CMOs and their procurement teams think about AI, they think more for less. Not less relative to the growth it drives. Just less. And it will stay that way until agencies find a clear, transparent way to show clients how they deliver value from their AI spend. Folding AI investment into principal media margin, rather than itemizing it, sidesteps that fight until they can win it, and lets agencies fund the infrastructure they’ve already built in the meantime. And it’s not just talk — holdcos are already getting real contracts signed on this basis, not just floating it in pitches.
This is what happens when an industry gets forced to adapt to something big on the fly. It’s not always the right choice that gets made, just the practical one. The average client has no mental model of how many tokens map to what outcome. Outsourcing that risk to an intermediary who can price it and carry it makes sense on those terms. The client gets certainty instead of exposure to a number they can’t forecast, and the holdco gets a shot at the pricing power its time-and-materials fee model has never given it.
“What we saw in the last two years is that most agencies were carrying those costs entirely on their balance sheet, and they’ve been using it as a differentiator to win clients,” said Ruben Schreurs, CEO of media management firm Ebiquity. “But now, starting this year, agencies can’t afford to continue to subsidize all those costs.”
The AI companies have their own reasons to prefer it this way, too — servicing thousands of individual advertisers directly isn’t a business any of them particularly want, any more than Google or Meta ever wanted to replace agencies rather than resell through them.
The case against
Strip away the tokenomics language and you’re often looking at a fee discount tied to a guaranteed slice of an agency’s most profitable inventory. One executive gave a number that said more than any analysis could: a holdco offered a total fee of zero in exchange for a large chunk of spend going through principal media.
Nobody outside the deal knows what the agency paid for those tokens or what margin sits on top. Clients buy an outcome and take the AI cost on trust. Procurement teams have no benchmark for what a unit of AI work should cost. Media at least has a rate card people can argue about. This has nothing. One person close to these negotiations put it simply: “you just don’t know quite how much the agency pays for it versus you.”
Then there’s the expectation problem. Most clients still assume AI should shrink the bill, since it’s supposed to replace headcount rather than add a new cost. An agency that tries to charge for AI usage runs into that assumption from both directions. Fees should fall because AI is doing the work. Don’t charge extra for the AI doing it. Burying the AI cost inside principal media doesn’t fix that. It just moves the argument somewhere clients can’t easily see it happening.
“Principal trading is disclosed in holdco earnings as a growth line,” said Robert Webster, former WPP exec and founder of AI marketing consultancy TAU. “The AI layer makes thousands of micro-decisions inside that permission, and nobody reviews them. Disclosed doesn’t mean aligned.”
The model has a failure mode too, and almost nobody’s talking about it. Lock in three years of token volume at a fixed rate, then get good at prompting or orchestration, and suddenly you’re holding far more capacity than you need. Sell it at a loss or write it off. Schreurs, who isn’t hostile to any of this, still called it “obviously the danger” sitting underneath the whole arrangement.
“That’s obviously the danger in all of this,” he continued. “If agencies become a futures market and commit to token volumes for three years at fixed rates, and their team engineers [get] much better prompting or a better agentic orchestration system, meaning they need a fraction of the tokens they thought they would, they’re stuck with this massive amount of tokens, which they need to resell in order to not incur a write-off.“
Whether any of this is dishonest depends entirely on what happens next: whether the guardrails, the auditing, the contractual language actually show up, or whether this becomes one more version of the old story everyone in this industry has heard before.