Where does all that Meta capex go? Better ads is one prominent area.
On Monday, Meta engineers — 13 engineers to be precise — explained in a blog post how Meta is turning spare AI computing capacity into better ad selection.
The team describes a system already operating that pools temporarily idle capacity and puts it to work in two ways:
-
- Rank More Ads: Evaluate more candidate ads, improving the chance of finding a relevant match.
- Serve Better Models: Send more ad requests through more powerful prediction models.
When that spare capacity is needed elsewhere, the extra work scales back while protecting normal ad delivery. Meta says this improves advertiser outcomes without requiring additional capacity orders, but provides no quantified performance lift.
The most interesting detail is how Meta decides which requests deserve the stronger model. As the engineers explain, “the incremental benefit of the more powerful model does not always align with the absolute value of a request.” The system tests where extra computing improves results most. Meta is also developing a system to automatically allocate spare capacity among its different uses.
Read: “Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations” (Oct. 12) – Meta
From tipsheet: At the end of March, the engineering team explored the company’s “adaptive ranking model,” which helps deliver the right ad at the right time for Meta’s Advantage+ ad platform. That post described how Meta makes more sophisticated predictions fast and cost-efficient enough for ad delivery. October’s post shows how spare computing capacity can improve those decisions further.
The common thread is the economics of inference. Better advertising depends on both stronger models and how efficiently Meta deploys them. Inference efficiency may become a competitive advantage in advertising and not just a way to reduce AI infrastructure costs.
AGENCIES
Q3: Publicis ‘not exposed’ to Apple Safari changes
Welcome to another earnings season.
Yesterday, agency holding company Publicis Groupe was among the first to report its third-quarter results, which were slightly ahead of Wall Street expectations.
The Wall Street Journal reports:
“The Paris-based advertising group said Tuesday that organic net revenue grew 5% in the third quarter, an acceleration from the growth rate of 4.8% it reported for the prior quarter. Overall, Publicis reported third-quarter revenue of 4.61 billion euros ($5.14 billion), up 12% on a reported basis.”
Publicis also said it completed a €500 million bond issuance in September to help finance its planned acquisition of data collaboration platform LiveRamp. The deal is expected to close before the end of 2026.
On the earnings call:
-
- Publicis CEO Arthur Sadoun said that the planned LiveRamp acquisition builds “on our investment in Sapient and Epsilon, and the position we have established in commerce, influencer, and sports.”
- An analyst asked whether Apple’s Safari changes could affect Publicis given the deal. Sadoun said he couldn’t comment on LiveRamp specifically, but said Publicis was “not exposed” to the Apple changes and saw no risk.
Analysts also pressed Sadoun on Publicis Sapient’s AI transformation opportunity. He said the unit’s improving performance is being driven by work with Publicis’ media and creative businesses to transform clients’ marketing. Publicis expects Sapient to return to growth in the fourth quarter, although Sadoun cautioned that it is too early to expect a pickup in large enterprise AI transformation projects.
Read: “Publicis Account Wins Fuel Top-Line Growth” (Oct. 13) – The Wall Street Journal (subscription)
More:
-
- “Publicis Groupe: Third Quarter 2026 Revenue” (Oct. 13) – Publicis Groupe
- “Earnings call transcript: Publicis Groupe posts Q3 2026 revenue beat, stock slips” (Oct. 13) – Investing[dot]com
- “Publicis is selling directly to AI-anxious CEOs as it steps away from traditional pitches” (Oct. 13) – Digiday (subscription)
LLMs & CHATBOTS
GEO strategy includes 24-hour turnaround
Marketers are turning GEO strategy up a notch, according to Adweek’s Trishla Ostwal who attended Adobe’s Semrush Spotlight event in London yesterday.
Brands are moving from asking whether AI mentions them to assigning people and processes to influence those mentions.
Motorway marketer Cassie Hall said at the event, “We can publish new content and see it being cited within 24 hours.” Ostwal reports that Motorway is also preparing to add ChatGPT Ads as a paid channel within its GEO strategy.
Read: “Flo and Motorway’s Playbook for Getting Cited by AI” (Oct. 13) – Adweek (subscription)
Related: “11 of the Best-Funded GEO Startups Racing to Win AI Search” (Oct. 12) – Adweek (subscription)
PLATFORMS
Morgan Stanley on Unity, Applovin and AI
Analysts at Morgan Stanley published new research yesterday previewing upcoming earnings reports (expected in early November) for ad tech companies Unity and AppLovin.
On Unity:
-
- “We believe Unity remains in the early stages of a multiyear innovation cycle, with elevated expectations the clearest near-term risk. Progress on Unity AI and the Google Playground partnership could also reinforce confidence in Create’s positioning for AI-native game development.”
On AppLovin:
-
- “For Applovin, we view 3Q as a pivotal quarter as investor concerns around competition and the sustainability of core gaming growth intensify. While we continue to believe competition is not zero-sum, Applovin needs to demonstrate consistent model improvements and above-market growth to drive multiple upside.”
Morgan Stanley sees Unity’s AI engine Vector as being early in a multiyear growth cycle, with six consecutive quarters of more than 15% sequential growth, compared with AppLovin’s more volatile performance following the launch of AI engine Axon 2.0. Despite Unity’s stronger recent growth momentum, Morgan Stanley sees more valuation upside in AppLovin (which fell precipitously in Q3), with price targets implying gains of 18% and 48%, respectively.
Separately, Unity CEO Matt Bromberg explained the thinking behind Unity Spark, part of the company’s broader push into AI-powered game creation, in a blog post last week.
Read: Why we built Unity Spark (Oct. 7) – Unity blog
From tipsheet: The competition between Unity and AppLovin extends beyond advertising performance. AppLovin took Unity to court earlier this month, alleging theft of trade secrets. The technology and talent behind AI models that determine which ads win and how effectively they perform have become strategic assets worth fighting over.
While general-purpose AI models may become commoditized, specialized models trained on proprietary data can provide a lasting competitive advantage.
LLMs & CHATBOTS
Developments
PEOPLE MOVES
Now hiring
-
- Pebble Post: “Back at it with Adam Soroca, this time at PebblePost where I’ve joined as SVP of Business Development & Partnerships. Adam and I built together at nToggle and then helped transform Rubicon Project into Magnite. His vision for PebblePost is bold, and I’m in to help execute it.” (Oct. 13) – Mark Balabanian on LinkedIn
MORE
- When AI paywall? “Making AI Pay Its Way, With Cloudflare” (Oct. 13) – AdExchanger Talks podcast
- More is better: “Google Ads API Release and Sunset Schedules for 2026-2027” (Oct. 12) – Google Ads Developer blog
- Uber graph: “TripleLift Named First Partner for Uber Advertising’s U.S. Offsite Audience Expansion” (Oct. 13) – TripleLift
- Investor ad tech: “Aperiam Annual Summit, November 10” – Eric Franchi, Aperiam Ventures on LinkedIn
- Courting crawlers: “Reddit Accuses Former Law Firm of Switching Sides to Defend Perplexity AI” (Oct. 12) – Law Commentary