Training data is the advantage. Do you know where yours is going?

Listen to anyone talk about large language models right now and they land on the same conclusion. The architecture is not the moat. The compute is not the moat. The training data is the asset that compounds, because every new signal makes the next model better, and a better model attracts more signal.

This raises a question every operator should be able to answer and most cannot. Who is learning from the campaigns you run?

When a campaign runs through infrastructure you own, the answer is obvious. You are. When it runs through infrastructure someone else owns, the answer is less clear, and the question is worth sitting with. Where does the signal go? Whose model does it make smarter?

The questions the pixel era taught us to ask too late

We have been here before. A decade ago, publishers and advertisers put Google's and Facebook's pixels on their pages. The pixels were free, they worked, and few asked what those platforms were learning from all that data. By the time the industry did ask, the platforms had already compounded the advantage.

The lesson is not that the same thing is happening again. The lesson is that asking late is expensive. 

The real question is your infrastructure's incentives. Is the party running your rails neutral, or does it have a commercial stake in the outcome of your campaigns?

Publicis's pending acquisition of LiveRamp is the event crystallizing this for the industry right now. But it's part of a pattern, not an isolated deal. Publicis bought Epsilon in 2019, picked up Lotame last year, and is now acquiring LiveRamp. Even WPP's InfoSum, built specifically to keep client data from pooling across a shared model, now sits inside a holding company. The infrastructure doesn't have to be centralized for the ownership question to matter. The concern is structural. It applies anywhere the entity running the infrastructure also holds a position in the market that infrastructure serves.

Once the operator of the rails has a stake in what runs across them, what that operator learns from you stops being theoretical. A campaign generates signal at every step, what creative worked, which audience responded, where the spend was wasted. That signal is the most valuable thing a campaign produces. When it runs through centralized infrastructure, is that learning contained to you? Or does the owner of the infrastructure see it across every client it serves?

Take a beverage brand running a campaign through centralized infrastructure. The campaign teaches the platform what messaging landed, which targeting variables underperformed, where the budget was wasted. Now a competing beverage brand runs on the same infrastructure. Is that second brand starting from scratch, or from lessons the first brand paid to learn? 

The walls were built for people, not models

Holding companies already run competing brands inside the same walls. Two agencies under one parent, each serving a rival in the same category, separated by conflict walls meant to keep one client's strategy away from another's. Those walls were built for people. They govern who sits on which account, who sees which deck, who is allowed in which room. 

But those walls were built to separate humans, not to contain a shared model. Whether they can do the second job is not an obvious yes.

In this case, exposure is larger than a data leak. It would mean funding the model that out-targets you. 

This isn't hard to check. Ask whoever runs your infrastructure whether they take a position in media buying or selling. Ask what happens to your campaign signal after the campaign ends, and whether it's used to train anything beyond your own account. If the answer is vague, that's the answer.

Neutral or biased? The question is yours to answer.

You don't need to be inside a holding company for this to apply to you. You just need your signal running through infrastructure one of them now owns.

For most of the last decade, choosing your infrastructure was an operating decision. Cost, coverage, convenience, handled a few layers down from strategy.

It is a different kind of decision now. The infrastructure you route your signals through determines whether the intelligence generated by your spend stays yours or accrues to a platform that also serves your competitors. So the question is worth asking plainly. Do you want your infrastructure neutral, or are you comfortable not knowing whose model your campaigns are training?

The operators asking that question now are the ones who keep their advantage instead of handing it to whoever else is renting the same rails.

Trip Foster is Co-founder and CRO of GrowthCode. GrowthCode provides identity and data infrastructure as-a-service for the digital advertising ecosystem.