Ad tech hopes open-source AI like DeepSeek will level the playing field with walled garden rivals
By Marty Swant • February 4, 2025 •
Ivy Liu
The open web has spent years competing for crumbs against companies like Google and Meta. But now, some ad tech players say the Chinese startup DeepSeek creates a new chance to build their own AI offerings beyond the walled gardens.
Chris Vanderhook, COO and co-founder of Viant, still remembers when Google’s AdX crushed the ad network business model right out of the gate. “Once you were in it and were bidding into them, they had such scale, and at any moment, they could tilt the rules in their favor and have all the value,” he said.
The jury’s still out on whether the Chinese startup’s open-source R1 model is safe for U.S. companies to use — or if it’s as cheap or as accurate as initially positioned. However, Vanderhook and others don’t think they’ll be able to benefit from its innovations if they’re able to soon replicate their own AI models thanks to open-source technology without relying on models like Google’s Gemini, Meta’s Llama, or OpenAI’s GPT models.
“In the back of your mind, you know Google’s in the advertising business, so is Meta,” Vanderbook said. “It’s still their LLMs, sure there are some white papers, but you don’t know the weightings, don’t know the data. Also, these are all one-size-fits-all models.”
Viant, which has its own AI-powered adtech offering called Viant AI, is already thinking about various ways to not just innovate using open-source models like R1. However, the democratization of LLMs also has him rethinking his pitch. If everyone can make their own AI model, it’ll put data quality and transparency more front and center.
Open-weight models still require users to trust that their goals are aligned with the AI model’s data, said Jaysen Gillespie, head of analytics & data science at RTB House. However, they noted open-weight models still offer data privacy gains for users by running on local hardware to reduce data leakage. Although open-weight models don’t remove walled garden advantages like protected inventory and audiences, they could help ad tech players in other ways.
“Open internet ad tech serves as a price cap for walled gardens, which can only charge marketers as much as the next best alternative,” Gillespie said. “Tech has always sworn by the ‘Good. Fast. Cheap. Choose any two’ mantra. Did the Chinese just blow that to bits?”
The challenge with ad tech is often doing a lot in very little time and for very little money, said Index Exchange CTO Ray Ghanbari. One way he sees DeepSeek’s innovation helping is through its model distillation, which makes LLMs cheaper to make, faster to run, and easier to fine-tune for specific knowledge areas like ad tech.
Because programmatic ads require extreme speed to process massive transaction volume, distilled models might be easier to deploy within ad tech platforms. That could help improve content categorization and ad-targeting beyond what’s possible today. Ghanbari likened R1 to an iPhone moment by helping people rethink how the industry considers training AI models for ad tech.
“A lot of the advantages that an LLM can provide have been just over the horizon,” Ghanbari said. “And I’ve been very keenly waiting for a breakthrough like DeepSeek because it’s telling us that the horizon is getting much closer than it appears.”
Ghanbari is just one of many ad tech execs excited about how open-source AI models could improve contextual analysis. Another company, Chalice AI, uses both off-the-shelf and open-source LLMs to develop its own version of existing models.
Tylynn Pettrey, vp of data science at Chalice AI, said the company is exploring ways to speed up how to optimize LLMs to improve page-level analysis with more granular and fresher data. She mentioned one partner Chalice is working with is Sincera, which recently was acquired by The Trade Desk. Chalice received funding from the Trade Desk’s venture arm in 2021.
“A lot of this information needs to be updated quickly because articles change, page content changes,” Pettrey said. “And if you’re doing this once per week or once per month, a lot of that information is stale. So what may have been a safe site last week suddenly has a politically charged news article that a brand would not consider safe anymore.”
Other ad tech startups like OpenAds.AI have also spent the past week testing the R1 model. Steven Liss, OpenAds.AI’s co-founder, said he’s been testing ways to use DeepSeek to generate and curate training data to improve speeds for ad-targeting and ad creation. Liss also thinks AI will soon become cheap enough to enable URL-level analysis — even if reasoning models like R1 are still too slow at the moment. (The startup’s running R1 on private endpoints using U.S. cloud providers.)
“Our first product — real-time generative ads on AI chat and search — was only possible once inference costs fell below CPMs,” Liss said. “This next jump in model efficiency lets us scale generative ads to the whole programmatic ecosystem. LLMs in [real-time bidding] sounds crazy now, but we’ll get there.”
https://digiday.com/?p=567556
