Ad-Tech’s LLM Dilemma: You Don’t Need a Poet to Run a Bidder

I recently read a headline claiming that major DSPs are “empowering their demand-side platforms with an LLM data backbone for the agentic world.”

Let’s be honest: That is a load of nonsense.

The industry is currently obsessed with “AI-washing” every legacy system, but in the world of Real-Time Bidding (RTB), using a general-purpose Large Language Model (LLM) to handle core logic isn’t just overkill—it’s bad engineering.

The Tool vs. The Task

RTB is a high-performance, high-precision environment. It requires extreme speed and mathematical efficiency.

• The LLM is an interface for humans. It is designed for nuance, conversation, and creative synthesis.
• The Bidder is a high-speed machine. It needs to process millions of requests per second with microsecond latency.

If you need to stamp 10,000 metal parts an hour, you buy a pneumatic press. You don’t hire a master craftsman—or an army of interns—to hand-hammer each one. Using a clumsy, resource-heavy LLM to do a straightforward, high-speed job is a fundamental misunderstanding of how our tools are built.

Moving Beyond the Hype: RAG and the Bidder

The fact that these conversations are happening reveals a massive gap in how people understand ad-tech infrastructure.

At Teads, we aren’t chasing buzzwords; we are building for performance. We are currently developing a system that performs RAG-like (Retrieval-Augmented Generation) requests within the bidder itself.

This is a massive leap beyond the simplistic “AI backbone” descriptions we see in press releases. By integrating the logicof retrieval without the bloat of a general LLM, we can move away from static bidding and toward a more dynamic, data-aware system that still maintains the speed required by the open web.

Where LLMs Actually Fit: The Niche Use Case

Does this mean LLMs have no place in the bidder? Not necessarily. But their role will be surgical, not structural.

The most intriguing opportunity lies in Contextual Targeting. Historically, contextual has always struggled with the “Goldilocks” problem: it’s either too broad (useless) or too precise (impossible to scale).

The future isn’t a giant LLM running the whole DSP. Instead, I expect to see purpose-built mini-models running inside the bidder on behalf of specific, high-value campaigns. Imagine a lightweight, specialized model performing real-time semantic analysis to find that “just right” context at scale.

That is a vision worth talking about. That is a PoC I expect to see very soon.

The Bottom Line

We need to stop pretending that every problem is an LLM-shaped nail. In ad-tech, performance is king. Let’s leave the poetry to the interfaces and keep the bidders focused on what they do best: high-speed precision.