Pinterest and Nvidia Build Multimodal AI Search Infrastructure to Cut Inference Costs and Latency
Pinterest has unveiled a standardized multimodal AI infrastructure built with Nvidia's Blackwell B200 GPUs and open-source Dynamo framework, aiming to make visual search faster and more cost-efficient across its fleet of roughly 14,000 Nvidia GPUs. Company benchmarks claim precomputing visual representations reduced overall latency by 7.3 times and cut inference costs to less than 8% of closed proprietary models.
Pinterest has developed a standardized multimodal AI infrastructure in partnership with Nvidia, combining Blackwell B200 GPUs, Nvidia's open-source Dynamo framework, and Pinterest's own visual embeddings to improve the speed and cost-efficiency of its visual search capabilities, the company announced on September 21, 2026.
The system is designed to run vision-language workloads across Pinterest's fleet of roughly 14,000 Nvidia GPUs. According to company benchmarks, precomputing visual representations reduced overall latency by 7.3 times and accelerated initial responses by roughly 85 times. Pinterest's AI assistant, Pinterest Assistant, can now process up to 25 times more visual context per request under the new architecture.
"Building the next generation of AI-powered discovery means investing in infrastructure that can keep up with the scale and complexity of Pinterest," said Kartik Paramasivam, chief architect at Pinterest, as quoted by TechRepublic.
Pinterest processes over 80 billion monthly searches across 640 million active users, with more than 96% of text queries described as unbranded, according to TechRepublic. The publication reported that post-training open-source models on Pinterest's first-party Taste Graph using Nvidia hardware has enabled the company to run inference at less than 8% of the transaction cost of closed proprietary models.
The infrastructure is intended to address a persistent gap in digital shopping discovery. TechRepublic noted that shoppers often lack the precise terminology to describe visual concepts — such as a specific mid-century furniture silhouette or a woven textile pattern — making text-based search insufficient on its own. By precomputing image representations and separating different stages of inference, Pinterest aims to avoid repeatedly processing the same raw visual data during a conversation.
Early advertiser tests of Smart Assembly, an automated creative tool under Pinterest Performance+, yielded a 6% bump in average click-through rates, according to TechRepublic.
The publication also noted that operating large vision models introduces real infrastructure challenges, including strain on memory bandwidth and key-value cache capacity from high-dimensional image embeddings and multi-turn conversations. Maintaining responsive query speeds during real-world peak traffic requires continuous dynamic autoscaling, creating ongoing trade-offs between server capacity and user responsiveness, TechRepublic reported.
Prepared with AI assistance and reviewed by the editorial team.
