Achieving Near-Linear Training Scalability for Pinterest’s Foundation Models

**Pinterest Engineering** team members Sheng Huang, Pong Eksombatchai, Saurabh Vishwas Joshi, Gaurav Arora, and Karthik Anantha Padmanabhan detailed infrastructure advancements in their published work, "Achieving Near-Linear Training Scalability for Pinterest’s Foundation Models." The covered foundation models directly power recommendation systems serving over **600 million monthly active users**.

Technically, the authors presented their latest **Foundation Model** at **ACM RecSys 2025**, built to deliver near-linear training scalability. The system pre-trains on **two years of user activity data** to learn long-term user representations and is deployed directly into **Hom**.

For machine learning and platform engineers, establishing near-linear training scalability over **two years of user activity data** addresses primary compute throughput bottlenecks. This scalability enables platform teams to pre-train foundation models on massive multi-year dataset windows without encountering non-linear training overhead.

In the competitive landscape of large-scale recommendation platforms serving **600 million monthly active users**, efficient dataset ingestion is a primary differentiator. **Pinterest** demonstrates at **ACM RecSys 2025** how foundation model pre-training can scale linearly when processing multi-year user activity histories.

Engineers implementing large-scale recommendation infrastructure should evaluate near-linear pre-training architectures for multi-year behavioral logs. Key developments to monitor include performance metrics from the deployment into **Hom** and the operational efficiency of pre-training on **two years of user activity data**.