Pinterest Deploys 14,000 Nvidia GPUs To Drive Multimodal AI Across Its Platform

Pinterest has built a new AI infrastructure layer using Nvidia’s Blackwell GPUs and Dynamo serving framework, giving the visual discovery platform a unified foundation for deploying multimodal AI across search, content safety and its conversational Pinterest Assistant product.

The work represents a deepening of a partnership between the two companies that has spanned close to five years, beginning with recommendation systems and expanding into generative AI and vision-language models (VLMs).

Pinterest now operates a fleet of 14,000 Nvidia GPUs spanning Blackwell, Hopper and previous-generation architectures.

Handling images at scale

Pinterest serves 640 million monthly active users, processing more than 80 billion searches per month across more than 16 billion user-created boards. More than 96 per cent of those text-based searches are unbranded, meaning users are typically looking for ideas or products without a specific brand in mind.

That dynamic makes the platform attractive to advertisers, but it also places considerable demands on the AI systems underpinning the experience. Every interaction on Pinterest is grounded in images, and the company’s AI models must interpret visual content alongside text in real time.

Unlike text-only large language model workloads, VLM inference is what engineers describe as “prefill-heavy.” Encoding image content is computationally expensive, request payloads grow when they carry multiple images, and multi-turn conversations accumulate visual context across exchanges, compounding memory pressure with every turn.

A single Pinterest request may include thousands of images, each requiring preprocessing, tokenisation and encoding before the language model can act on it.

Blackwell and Dynamo as the compute foundation

Pinterest standardised on Nvidia’s Blackwell B200 GPUs as its primary compute layer for inference workloads. In preliminary benchmarking for Pinterest Assistant, B200 instances delivered more than 2x latency improvement compared with Hopper-generation hardware.

For orchestration, Pinterest selected Nvidia Dynamo, a serving framework that coordinates how inference workloads are distributed and managed across GPU infrastructure. Dynamo was chosen for its compatibility with Pinterest’s service discovery systems and its inference-engine-agnostic design.

The company uses vLLM as the inference engine underneath Dynamo, and teams across Pinterest can now build on the stack through a common Chat Completions API rather than creating custom infrastructure for each product.

Pinterest’s Chief Architect, Kartik Paramasivam, described the rationale behind the investment.

“Building the next generation of AI-powered discovery means investing in infrastructure that can keep up with the scale and complexity of Pinterest,” Paramasivam explained. “Our collaboration with Nvidia helps us deliver faster, smarter and more personalised experiences for the hundreds of millions of people who use Pinterest.”

Cost and performance gains

One of the more tangible outcomes of the new infrastructure is a reduction in inference costs. Pinterest post-trains open models on Nvidia GPUs using its own proprietary data, achieving inference transaction costs at less than 8 per cent of what it previously paid for comparable closed proprietary models.

On the advertising side, Pinterest’s AI creative optimisation tools are already producing measurable results. Smart Assembly, a capability within Pinterest Performance+ that automatically builds and serves ad combinations from advertiser-uploaded images, delivered a 6 per cent average click-through-rate improvement in early alpha testing.

Pinterest Assistant gets a visual upgrade

Pinterest Assistant is the company’s AI-powered conversational tool designed to help users find products and ideas through a combination of image and text understanding.

A key technical contribution is Pinterest’s internal image encoder, which allows the platform to serve precomputed visual embeddings through Dynamo and vLLM rather than processing raw pixels for every request.

With Dynamo, Pinterest Assistant can now process 25x more visual context per request. When using projection embeddings, the system delivers up to 369x faster time to first token and 44x faster end-to-end latency compared with raw image processing.

The Vice President of AI and HPC Frameworks and Libraries at Nvidia, Ujval Kapasi, pointed to the scope of the collaboration.

“By building its multimodal AI infrastructure on Nvidia accelerated computing and software, Pinterest can bring the next generation of visual and conversational AI experiences to its users at scale,” Kapasi remarked.

Broader platform ambitions

The infrastructure is not limited to a single product. Pinterest is using the Nvidia Dynamo stack across ranking systems, content safety, optical character recognition pipelines, signal generation and conversational agents.

Pinterest is also adopting Dynamo’s AI Configurator, a tool that can simulate more than 10,000 deployment configurations in seconds, alongside Dynamo Planner, an autoscaler designed for VLM and LLM workloads that dynamically adjusts compute allocation based on real-time traffic patterns.

The company confirmed it is tracking Nvidia’s Vera Rubin platform roadmap for future inference capability.

Gen Z is now Pinterest’s largest and fastest-growing user cohort, and the platform has sustained a multi-quarter streak of record user numbers and double-digit user growth. The company appears to be betting that a shared multimodal AI platform will allow it to ship new experiences faster without each team needing to build bespoke serving infrastructure from scratch.

Last Updated on September 15, 2026 by Nick Ross