Pinterest taps NVIDIA to power next generation visual AI products

Pinterest has built a new multimodal AI infrastructure layer with NVIDIA to support visual search, content understanding, safety systems and conversational features across its platform.

The system runs on NVIDIA Blackwell GPUs and NVIDIA Dynamo while incorporating Pinterest’s own visual embeddings and open source models. The shared infrastructure is designed to let Pinterest teams build and deploy multimodal AI products without creating separate systems for each use case.

Pinterest said the platform helps turn more than 80 billion monthly searches into signals for AI powered discovery and shopping. It also allows Pinterest Assistant to process 25 times more visual context per request.

In benchmark testing, Pinterest said using precomputed visual representations improved response startup times by about 85 times and reduced overall latency by 7.3 times compared with repeatedly processing raw images.

The infrastructure was initially developed to support Pinterest Intelligence, the company’s AI layer used throughout the platform. It is now being used across ranking, safety, optical character recognition, signal generation and other AI applications.

Pinterest said it plans to expand its use of vision language models in products including Pinterest Assistant as it builds more visual and conversational experiences.

The project builds on nearly five years of collaboration between Pinterest and NVIDIA across more than 14,000 NVIDIA GPUs.

Pinterest Chief Architect Kartik Paramasivam said the partnership gives the company infrastructure capable of supporting increasingly complex AI discovery experiences at scale.