S4 Capital posts improved margins but faces exploding AI token costs

Full analysis

Your draft

Sir Martin Sorrell just admitted his company has been "ill-disciplined" on AI token spend, the per-use fees that AI providers charge every time you run a request. That admission, buried inside a doubling of operating profit to £35.2 million, is the tell every agency and ad-tech operator should read closely. The efficiency story the whole sector is selling investors runs on cheap AI that may not stay cheap.

What's actually being decided: not whether AI helps agencies. It's who eats the cost when AI providers move from flat software subscriptions to charging by consumption, and whether "we automated it" survives that switch. This is Type 1 for anyone who has already repriced client contracts around cheap inference. Reversing a signed rate card is hard. The forcing function is the pricing model shift itself, which Forrester's Jay Pattisall says arrives when equity and private-equity money stops subsidizing AI economics.

The Market Analyst. Two stories are colliding here. S4 sells a margin-recovery narrative: 12.3% margins, headcount down from 8,200 to 6,150. But the token bill is a cost structure nobody has repriced. In plain terms: investors are valuing agencies as if AI makes them cheaper, when it may make them more expensive. S4's pain is at least visible and partly baked into a beaten-down stock. WPP, Publicis, Omnicom carry the same token exposure invisibly, because consensus models still assume flat-fee AI. That's where the earnings surprise lives in 2027. The contrarian read: the lean, already-restructured player absorbs this shock better than the giants with legacy fixed cost.

The Skeptic. The Goldman 24-times-token-growth forecast is carrying this entire story, and it deserves a hard look. Enterprise AI curves usually undershoot the early hype. More to the point, inference costs have fallen by orders of magnitude in two years, and open-weight models from Meta and Mistral keep commoditizing the stack. Pattisall's "devastating" assumes providers hold pricing power. Against open-source pressure, that's not obvious. And S4 doubled operating profit while the token problem was already live. That reads like an operational discipline issue, not an existential one. The narrow, real risk: agencies that promised clients AI savings on flat-fee assumptions now owe a contract renegotiation. That's a lawyer problem, not a sector collapse.

The Operator. Wesley ter Haar's "horses for courses" model, cap tokens on low-value tasks, leave them loose for coding, sounds elegant on a slide. On Tuesday it's brutal. Somebody in finance or procurement now has to adjudicate which task is worth burning tokens on, and that decision tree collapses at scale. Coding agents are already making spend "explode" by Sorrell's own words, and coding is exactly where they left the limits off. Expect Q3 ops teams retrofitting usage governance into Monks.flow before it eats the margin they won by cutting 2,000 people. In plain English: they fired humans to pay for AI, and the AI is now threatening to cost more than the humans did.

The CFO. Here's the math that should keep every agency CFO up. You cut 2,000 heads to fund an AI transition and booked the savings as margin. Those savings are permanent. The token costs are variable and, if Goldman is even directionally right, growing. You've swapped a fixed cost you controlled for a variable cost your vendor controls. Sorrell's own quote is the horror scenario stated plainly: it may end up "more expensive to automate advertising than to hire humans to create it." Monks.flow's ability to switch models by cost is a real hedge, the AI equivalent of routing spend to the cheapest supplier. But building and running that routing layer is itself a cost, and S4 is mid-sized and capital-constrained.

The Customer / End User. The advertiser bought a promise: AI makes this cheaper and faster. If the agency's input cost triples when consumption pricing lands, one of two things happens. The agency eats it and margins crater, or the client gets a repricing conversation nobody wants. Clients who signed AI-efficiency deals on flat-fee assumptions are holding contracts that no longer reflect the agency's economics. Smart clients will start asking who bears token-cost inflation before they renew. That question, not the technology, is what reshapes the next round of agency pitches.

Where the council splits:

The Skeptic versus the Market Analyst on whether this is real. The Skeptic bets falling inference costs and open-source models outrun consumption growth, making the whole scare a repricing footnote. The Market Analyst bets the giants are carrying invisible token exposure that detonates in 2027 earnings. Both can't be right about 2027.

The CFO versus the Operator on the hedge. The CFO sees Monks.flow's model-switching as genuine margin defense. The Operator sees the governance layer as an operational nightmare that breaks at scale and costs real money to run. The routing layer is either the moat or the next line item that eats you.

And underneath both: does owning the orchestration layer, the software that picks the cheapest capable model per task, actually defend margin, or does model commoditization erode any routing edge as fast as you build it?

What this hinges on. Three beliefs. One, whether AI providers can hold pricing power as they shift to consumption billing, or whether open-source keeps inference cheap. Two, whether agencies locked client contracts to flat-fee AI economics they can't sustain. Three, whether a proprietary routing layer is a durable moat or a treadmill.

The council leans toward this being real but narrower than "devastating." Not a sector collapse. A brutal renegotiation cycle that punishes agencies who priced AI savings to clients on assumptions that break the moment billing goes consumption-based. The lean, transparent operator survives the repricing better than the giant whose exposure is still hidden from its own investors.

What to verify before believing either extreme: the actual trajectory of per-token inference prices over the next few quarters, and whether any holdco discloses AI cost of goods as a line item. The one that discloses first is telling you it's a problem.

Prediction: At least one of the big four holding companies (WPP, Publicis, Omnicom, or Dentsu) will name rising AI or token costs as a specific margin pressure in its full-year 2026 results reporting, by the March 2026 earnings season.

Confidence: Medium. Sorrell said it out loud first, and peers with the same exposure tend to follow once one of them puts it on the record.

Why: Sorrell just put "ill-disciplined" token spend on the record while posting good margins, which gives every rival CFO cover to acknowledge the same cost without looking uniquely mismanaged. The whole sector runs the same AI stack on the same consumption-pricing threat, so the exposure is shared, not S4-specific. Once one holdco frames token cost as a known, managed pressure, the others follow rather than let analysts assume they're hiding it. The opposite outcome, total silence, only holds if inference prices fall fast enough that nobody notices, and coding-agent spend already "exploding" at Monks argues against that.

Revisit by 2026-04-15: We're right if a WPP, Publicis, Omnicom, or Dentsu FY2026 report or earnings call explicitly cites AI/token/inference cost as a margin factor. We're wrong if all four get through full-year reporting without naming it.

The tell to watch isn't the headline margin. It's whether anyone starts breaking out AI as its own cost line. The first one to do that is admitting the subsidy party is ending.