Meta CFO: LLMs Not Yet Driving Ad Targeting
Despite the hype, Meta's CFO admitted the company is not yet using large language models to power its core ad-targeting business. The admission highlights that for key business units, generative AI is still in the pilot phase, underscoring the gap between research and production-scale deployment.
While Meta's core ad ranking and recommendation engine does not yet use LLMs, the company is heavily investing in other AI-powered tools for advertisers. Its "Advantage+" suite, first introduced in 2022, automates tasks like audience selection, budget allocation, and ad placements to improve campaign efficiency. Some businesses using these tools have reported up to a 12% reduction in cost-per-acquisition. Meta's generative AI features for advertisers are focused on the creative side of campaigns. These tools can automatically generate variations of ad copy, create new backgrounds for product images, and even turn still images into video ads. More than 4 million advertisers are already using these generative AI offerings. The long-term goal is to use LLMs to fundamentally change ad targeting by reasoning about content and context in ways current systems cannot. Today's ad ranking relies on past engagement signals like likes and shares, which is inherently backward-looking. LLMs could potentially predict a user's interest in a piece of content in real-time without needing that engagement history. This future vision requires a massive infrastructure investment. Meta plans to spend up to $135 billion on AI infrastructure in 2026, a significant increase from $72 billion in 2025. This spending covers data centers, custom AI chips, and securing hardware from partners like AMD, Google, and Nvidia to avoid over-reliance on a single supplier. A key technical hurdle for using LLMs in the core ad auction is latency. Ad auctions require decisions in milliseconds, a speed at which today's resource-intensive LLMs often cannot operate. The probabilistic nature of LLMs also clashes with the deterministic logic required for real-time bidding systems. Starting in December 2025, Meta plans to use data from user interactions with its AI chatbots to inform ad targeting. This means conversations with Meta AI could trigger ads for related products, adding a new layer of real-time intent signals for advertisers to leverage.
