Big-platform ad prediction, sold to everyone else.

Two thirds of screen time, outside the big platforms

Moloco starts from something its chief executive saw while building ad systems inside Google. Ikkjin Ahn joined YouTube around 2008, when he says the team had about 80 people, built its video profitability prediction system and later led the Android data team as its founding tech lead manager. He says that mixing machine learning with the economics he had studied lifted ad returns there by more than 50%. When he looked at where people actually spent their digital time, he concluded that platforms such as Google held only about a third of it. The other two thirds sat in independent apps, shops and services that had plenty of data but could not turn it into revenue. In 2013 he co-founded Moloco, short for Machine Learning Company, with engineers who had built similar systems inside large technology companies.

It looks as if big tech like Google has taken over the whole digital world, but when we did the numbers it was only a third.

Ikkjin Ahn · Co-founder and CEO, Moloco · Interview with The Korea Economic Daily, published in Korean; translated by Signet · 2026-01-04

The belief that sets Moloco apart is that one shared prediction engine, trained across many advertisers and millions of apps, can do for everyone else what in-house systems do for the largest platforms. The company calls that engine CARA, a compound ad recommendation architecture with six technical domains: campaign automation, supply, ad recommendations, bidding, creative and signals. By Moloco's own estimates it weighs more than a trillion ad opportunities a day, learns from more than 6 billion events a day, and redeploys updated models as often as every two hours. The Seoul machine learning posting puts the load at over 6 million bid requests a second, with predictions in under 7 milliseconds. The claim is that gains compound: better creative choices produce better training data, and richer supply data sharpens bidding.

Why now: phones, privacy and first-price auctions

Three changes made the bet timely. Smartphones multiplied the data and the ad slots; Ahn says watching people carry phones convinced him that the volume and variety of new data would open a new market. Privacy rules then pushed businesses toward their own customer data; in the 2021 funding announcement Ahn said it was increasingly critical for businesses to activate their own data as consumer privacy restrictions grew stricter. And most auctions in the app ecosystem now run as first-price auctions, where the winner pays exactly what it bid, so pricing each bid precisely is worth real money. Moloco also took its time. It says it spent six years building its machine learning and data infrastructure, and Korean coverage of its tenth anniversary dated serious commercial selling to 2018.

Setting the price of an ad is the hardest problem.

Donghwan Jeon · Co-founder and CTO, Moloco · Media briefing in Seoul before MOLOCON 25, reported in Korean; translated by Signet · 2025-06-12

One campaign through the engine

Take an illustrative app marketer at a Korean game publisher launching in the United States. The marketer picks the outcome that matters, such as a purchase, a deposit or a subscription, connects a mobile measurement partner such as AppsFlyer or Adjust, uploads creatives and sets a budget; the minimum for a new campaign is $300 a day. From then on CARA works impression by impression. For each ad request that arrives through more than 35 exchanges or Moloco's own SDK, it predicts how likely this user is to install, convert and keep spending, prices a bid within tens of milliseconds, and picks or assembles the creative. Results flow back through the measurement partner, and the models retrain. The marketer watches spend, installs and return in Moloco Ads Manager, down to publisher, placement and creative.

People still make the decisions the engine cannot. The marketer chooses the goal and the creative, and decides whether to trust the numbers. That last step is where Moloco's own staff come in. For Nexon in January 2026, Moloco's data scientists designed a randomized test with ghost bids so the publisher could see what its connected TV ads added beyond what attribution showed. On the retail side, retailers keep governance and creative review while Moloco handles targeting, product selection and bidding underneath.

Most figures in this chapter are Moloco's own internal estimates, each footnoted on its website. The practical test of the engine belongs to the customer: the strongest cases in this file are ones where a customer ran its own comparison, such as Nexon's incrementality test or Musinsa's comparison with other channels during its summer promotion.

In Seoul most of the open roles sit inside this engine: ranking and bidding models, the Ads Manager that marketers use, test automation and the CTV program. Two sit on the commerce media side, helping retailers trust the same engine with their own shoppers. Either way, the job is to make a prediction system's results believable to the people paying for them.

One engine, two businesses and a supply door

Moloco describes itself as one company with two businesses on the same AI system. The first and larger is Moloco Ads, a demand-side platform: app marketers use it to buy ad slots inside other apps, automatically, through real-time auctions. User acquisition and re-engagement campaigns run in one architecture, so signals from both feed the same models. The company says Moloco Ads can reach more than 2 billion people a day across more than 2.9 million independent apps, that 89% of its spend in the fourth quarter of 2025 went to video, interstitial and native formats, and that in 2025, 45 of the 50 highest-grossing mobile app publishers advertised with it. Its FAQ positions it beside Google and Meta rather than against them: those platforms sell mostly their own properties, while Moloco reaches the independent apps around them.

In April 2026 Moloco took the same engine into the living room with Performance CTV for app marketers. Each connected TV impression is optimized in real time toward an app outcome, and attribution runs through the advertiser's chosen measurement partner. In its launch announcement the company reported early results of up to 1.5 times higher return on investment on CTV than on mobile for campaigns running on both, and said about two thirds of users who installed after seeing a Moloco CTV ad did so within six hours. These are the company's own early figures.

The second business, Moloco Commerce Media, sells the engine to retailers and marketplaces so they can run their own ad business. Brands and sellers on the platform buy sponsored products, search ads, sponsored display and sponsored brands. The retailer can use Moloco's ready-made campaign manager, embed a widget in its seller portal, or build its own tool on Moloco's APIs. Each retailer's models are trained only on its own first-party data, and the documentation says data, training and inference stay in a silo per customer. Moloco says a retailer can go live in as little as six weeks, and that its retailers have reached ad revenue of 1% of gross merchandise value within the first year and 2% within two, with the strongest passing 3% within three. It competes with retail media technology that matches ads by keywords and fixed rules.

The third piece works on the supply side. Publishers install the Moloco SDK so Moloco's advertisers can bid directly on their ad slots, and so CARA can render richer formats and collect better signals. Moloco said in July 2025 that nearly 500 publishers used it, that it had become available through Google AdMob and Ad Manager, and that it was already a certified bidding partner in AppLovin's MAX and Unity's LevelPlay. Its SDK page says the number of publishers more than doubled during 2025.

Named customers and what each one measured

The clearest Korean case is Musinsa, the fashion platform. With the agency Madup it used Moloco Ads to bring back active, semi-active and inactive users, sending each group to its own event page through deep links, and then showed those returning users Product Ads powered by Moloco Commerce Media inside its app. During its largest summer promotion, cost per active session was 10% and 22% better than two other DSPs, video creatives were 2.5 times more efficient than images, and users attributed to both products had 66% higher average revenue per paying user than users attributed to neither. Moloco published the case in September 2025.

Nexon tested Performance CTV in January 2026 to see whether TV ads could bring new players to a live-service game. Moloco's data scientists split the audience into test and control groups and placed ghost bids for the control group, so both faced the same auctions. By Nexon's analysis, the incremental lift in new sign-ups was up to three times what attribution alone showed, and 1.6 times for returning users. In commerce media, W Concept, which launched its ad business with Moloco in 2023, saw ad spend as a share of gross merchandise value rise 2.6 times over 18 months to June 2026, a Moloco figure the retailer confirmed. Outside Korea, Scopely's Monopoly GO! drew more than 1 million installs through Moloco in its first 30 days and beat its return-on-ad-spend target by 2.6 times.

How to read these cases

Every case here is published by Moloco, and each compares against its own baseline: other channels during one promotion, attribution-based results, or a retailer's own starting point. They show what happened for one customer under stated conditions. A useful interview question is which baseline a team uses when it reports a win.