Ad Platforms Are Like LLMs. Which Frontier Model Are You Running On?

When GPT-4o replaced GPT-3, the performance gap wasn't because OpenAI wrote better code. The model trained on more data types simultaneously: text, images, audio, structured data. The relationships between those modalities generated intelligence that a single-input model could not replicate.

Large language models (LLMs) became “large multimodal models” (LMMs) the moment researchers realized that intelligence didn't live in text alone. It lived in the relationships between text and images, between images and behavior, between behavior and outcome.

Ad platforms are making the same transition. The companies that recognized it early are now running on a different architecture than the ones that didn't. The gap is already compounding.

The modalities are different. The logic is identical.

For an ad platform, there are multiple data modalities: creative performance (which formats, messages, and visuals drive response), attribution (which impressions converted and through what path), bidding dynamics (how auction pressure and floor prices affect margin and volume), audience behavior (how users move across apps, content, and surfaces), and supply signals (what inventory is available, at what quality, with what context).

A platform trained on all of these simultaneously can see relationships that a platform trained on only one or two cannot. The performance gap between a GPT-3 and a GPT-4o is not a mystery. Neither is the gap between a platform running on fragmented signals and one running on a closed loop. The advantage compounds: Each closed loop, from creative to auction to attribution, generates more training signals which improve the next cycle.

Also Read: Best AI Infrastructure Providers for Enterprise AI in 2026

What the deals are actually buying

Fox acquiring Roku, years after acquiring Tubi, reads on the surface as a media consolidation story. The more precise reason is that Fox is buying modalities.

It already knows what content audiences want through its legacy linear broadcast and sports rights portfolio, which gives Fox first-party viewership data across live events and primetime programming. Tubi’s acquisition added behavioral data from free ad-supported viewing. Roku adds device-level footprint and platform engagement across roughly 100 million households. Stack those together, and the data graph that none of the three could build independently now exists inside one system.

Publicis acquiring LiveRamp follows the same logic: an agency holding company acquiring the connective tissue of the data ecosystem. The attribution loop closes within a single system rather than across multiple vendors. Walmart's expansion into CTV advertising is the same thesis.

Each acquirer is buying a data modality it didn't own, to feed a model it's building. These are not scale plays. They are training set acquisitions.

The integration tax

Platforms that don't own their full signal chain pay an integration tax on every campaign. Someone manually reconciles the measurement. Something gets lost between the buying platform and the attribution vendor.

Creative performance data sits in a report that no bidding algorithm ever reads. Each handoff is a place where the training signal degrades.

Best-in-class stacks, by contrast, pay this tax on every campaign. Someone has to manually reconcile the signals. Over a single campaign, the cost is tolerable. Over a year of campaigns, across thousands of auctions, the degradation accumulates into a compounding disadvantage.

This shows up clearly in mobile. Platforms have evolved to optimize toward CPE, CPA, and ROAS, giving advertisers a direct link between spend and outcomes. That only works when the signal chain is intact from impression to outcome. Break the chain by splitting measurement across vendors and the model is training on incomplete data.

What advertisers should actually evaluate

The evaluation criteria for ad platforms need to change. Auditing on point-solution quality asks the wrong question. The right question is how many data modalities flow into a single model, and how fast does that model improve.

Where does creative performance data go after a campaign ends? Does it inform the next bid, or does it sit in a dashboard? Who owns the signal between impression and outcome, and are they the same party running the next auction? A platform that can answer those questions with "all of it stays inside the same system" is compounding. One that can't is paying the integration tax and passing some of it to you.

The real divide in this market is between platforms with a unified, cross-modal data graph and those still operating on single-channel signals. That divide is widening. A platform that fuses more data types into its models this quarter will have a larger advantage next quarter, not a similar one. The modalities a platform doesn't own today are not a temporary gap. They are an absence in its training set that compounds against it.

The compounding rate is the product

The M&A wave in media and ad tech is not about scale. It is about training data. The broadcasters, agency holding companies, and retailers buying into adjacent data layers are not building bigger organizations. They are building better models. Just as the leap from LLM to LMM wasn't about adding parameters, it was about adding modalities and the relationships between them, the leap from ad network to integrated platform isn't about adding inventory. It is about closing the loop between every signal in the stack.

Advertisers who align with platforms that have already closed those loops get the benefit of faster-compounding intelligence. Those running fragmented stacks keep paying the integration tax, and that tax rises every quarter the gap grows.