Solution
Pinterest, serving 600M monthly users, runs its AI-powered visual discovery platform entirely on AWS, processing 500+ petabytes of data and over 10M AI recommendations per second. The recommendation system trains on 600+ Amazon EC2 P4/P4de GPU instances and serves inference on 10,000+ Amazon EC2 G5 GPU instances, all orchestrated by Amazon EKS for rapid microservices deployment. Pinterest Canvas, a homegrown latent diffusion vision foundation model, generates and manipulates high-resolution product images on EC2 GPUs for advertiser creative tooling. Visual search across 2.5B+ recognized objects uses vision language models built on Amazon SageMaker, while Pinterest Assistant adds a voice-enabled conversational discovery experience powered by foundation models from Amazon Bedrock.
Data flow
Pinterest ingests user signals, Pin metadata, and uploaded images at scale; training data (18+ TB/day) flows to EC2 P4/P4de clusters that fit multimodal recommendation and Canvas diffusion models. Trained models deploy via EKS to EC2 G5 inference fleets, which generate 10M+ recommendations per second for the feed, search, and ads. Visual search queries route to SageMaker-hosted vision language models for object recognition, while Pinterest Assistant queries call Bedrock foundation models for conversational responses.
Solution architecture
7 components · 3 layers - Compute
- Amazon EC2 G5 instances Run 10,000+ GPU instances for real-time recommendation and ranking inference at over 10M predictions per second.
- Amazon EC2 P4 instances Provide 600+ high-end GPU instances for training large multimodal recommendation and content-understanding models on 18+ TB of daily data.
- Amazon EC2 P4de instances Extend P4 training capacity for the most demanding vision and foundation-model workloads including Pinterest Canvas.
- Amazon EC2 Provides the underlying compute fabric for microservices, data processing, and supporting infrastructure across 500+ petabytes of data.
- Amazon SageMaker Builds and serves the vision language models powering visual search across 2.5B+ recognized objects for refined in-image search and discovery.
- Serving
- Amazon Bedrock Supplies managed foundation models that power Pinterest Assistant's voice-enabled conversational discovery experience.
- Orchestration
- Amazon EKS Orchestrates microservices-based recommendation, ranking, and safety pipelines, enabling rapid AI deployment across the entire Pinterest platform.
- Amazon EKS for rapid AI deployment
- Microservices-based architecture on AWS since 2010
- Pinterest Assistant: generative AI voice-enabled conversational experience
- Pinterest Canvas: homegrown latent diffusion vision foundation model on EC2 GPU instances
- Recommendation system built entirely on AWS
- Visual language models built using Amazon SageMaker
17% year-over-year revenue growth ... 11% increase in monthly active users ... 230 basis point improvement in search fulfillment ... 70% of user discovery now AI-driven.
Built entirely on AWS, Pinterest uses Amazon EKS for rapid AI deployment across the business, 10,000+ Amazon EC2 G5 instances for inference and 600+ P4/P4de instances for training, analyzing over 18 terabytes of data daily.
Pinterest processes over 10 million AI-powered recommendations per second to support personalized discovery, advanced content moderation, and responsible features.
With the help of AWS, Pinterest scales processing, storage, and analysis of over 500 petabytes of data with ease.
