Pinterest expands Nvidia partnership, achieves 85x faster response startup

Pinterest just gave its AI backbone a serious hardware upgrade. The company unveiled a new AI layer built in partnership with Nvidia, combining Blackwell GPUs and Nvidia’s Dynamo inference framework with Pinterest’s proprietary visual embeddings to create what amounts to a shared foundation for all of its AI-driven features. The headline number: an 85x improvement in response startup times. Overall latency dropped by a factor of 7.3x. And the Pinterest Assistant, the platform’s conversational AI tool, can now process 25 times more visual context per request. For a platform that handles north of 80 billion searches every month, those aren’t incremental gains. They’re architectural. What Pinterest actually built The technical stack pairs Nvidia’s Blackwell GPU architecture with Nvidia Dynamo, a software layer designed to optimize how inference workloads run across GPU clusters. Pinterest layered its own visual embeddings on top. These are the numerical representations the platform uses to understand the content and style of images across its catalog. The result is a shared framework that Pinterest’s engineering teams can use across multiple products without building bespoke infrastru...