Social Media

How Pinterest Delivers 10M AI Recommendations per Second on AWS

PPinterestSource: AWS ↗March 2025

Pinterest built an AI-powered discovery engine on AWS processing 18TB daily, delivering 10 million AI recommendations per second across 10,000+ GPU instances, driving 17% revenue growth and 70% AI-driven discovery.

Impact

17% YoY

Revenue Growth

70%

AI-Driven Discovery

11%

MAU Growth

10M per second

AI Recommendations

Challenge

Needed to process billions of images and deliver personalized recommendations from 500+ petabytes while maintaining user trust at massive scale.

Solution

Built microservices architecture on AWS with 10,000+ GPU instances, Pinterest Canvas diffusion model, visual search recognizing 2.5B objects, and voice-enabled AI assistant.

Tools & Technologies

ASAmazon SageMaker AIML PlatformABAmazon BedrockML PlatformAEAmazon EKSCloud Infrastructure

What Leaders Say

“For over a decade, we have been leveraging AI to craft a uniquely positive online experience, striving to make every moment on Pinterest additive, not addictive.”

Kartik Paramasivam, Chief Architect, Pinterest
Get the full story.

Sign up to read complete case studies, access detailed metrics, and unlock all use cases.

Full Story

Pinterest needed to evolve its AI-powered discovery system to process billions of images and deliver personalized recommendations from 500+ petabytes of data without compromising user trust or prioritizing addictive engagement.

Pinterest implemented a microservices architecture on AWS with Amazon EKS for rapid AI deployment, 10,000+ EC2 G5 instances for inference, and 600+ P4/P4de instances for training. Key innovations include Pinterest Canvas (latent diffusion for image generation), visual search recognizing 2.5 billion objects via SageMaker, and Pinterest Assistant (voice-enabled conversational AI). The infrastructure processes 18 terabytes of data daily.

Results: 17% year-over-year revenue growth, 11% increase in monthly active users, 230 basis point improvement in search fulfillment, and 70% of user discovery now AI-driven.

Similar Cases

BO
Blue Origin
2,700+
ai agents deployed

Blue Origin deployed 2,700+ AI agents with 70% company-wide adoption, achieving a 90% reduction in hardware development time using Amazon Bedrock.

ManufacturingAerospace & DefenseABAmazon BedrockAEAmazon EKS
T
TaskUs
20%
average handle time reduction

TaskUs is a leading outsourced digital services company providing next-generation customer experience (CX) for innovative global brands. To move beyond flat-file embedding storage and scaling limitations, TaskUs built TaskGPT—a proprietary GenAI platform—with Pinecone as the core vector database for semantic search, RAG-based knowledge retrieval, and client-specific recommendations. The result: a 20% reduction in average handle time and a 5% increase in customer satisfaction across client deployments.

Business Process OutsourcingPPineconeABAmazon Bedrock
TR
Thomson Reuters
3,000+
subject matter experts' knowledge delivered via ai

Thomson Reuters integrated Claude via Amazon Bedrock into its AI platform, CoCounsel, to make the expertise of 3,000+ subject matter experts and 150 years of authoritative content accessible to legal and tax professionals. The solution combines Retrieval-Augmented Generation (RAG) architecture with multi-model deployment to deliver comprehensive, accurate professional analysis. Early adopters report dramatic efficiency gains, with some estimating task time cut in half or more.

Professional ServicesC3Claude 3 HaikuC3Claude 3.5 Sonnet
P
Phagos
10+ years to 2 months
treatment development time

French biotech Phagos uses Amazon SageMaker AI to match phages to bacteria 99.5% faster, cutting treatment development from 10+ years to 2 months and reducing wet lab testing by 50%.

BiotechnologyASAmazon SageMaker AI
CA
Cox Automotive
17 (from 57 evaluated)
production ai solutions

Cox Automotive deployed 17 production AI agent solutions using Amazon Bedrock AgentCore, reducing estimate completion from 48 hours to 30 minutes, achieving 3x consumer response rates, and projecting 17,000 hours saved.

AutomotiveABAmazon Bedrock AgentCoreABAmazon Bedrock
P
Postman
Up to 1,150/year
developer hours saved

Postman selected Claude Opus 4.6 as the default model for Agent Mode, saving developers up to 1,150 hours per year and nearly $1M annually for a 10-person team in API development automation.

TechnologyCAClaude APIABAmazon Bedrock