Pinterest Brings Conversational AI to Visual Discovery With NVIDIA Blackwell and Dynamo

For years, Pinterest ran this kind of intelligence through a mix of recommendation systems and large language models. But as the company expanded into vision-language models (VLMs), a new class of serving challenges emerged. Unlike text-only LLM workloads, VLM inference is prefill-heavy: Encoding image content is computationally expensive, request payloads grow dramatically when they carry multiple images, and multi-turn conversations accumulate visual context across turns, compounding KV cache pressure with every exchange. In production, a single Pinterest request may include thousands of images, each requiring preprocessing, tokenization, and encoding before the language model can act on it.

Serving this class of workload required more than adding GPU capacity. It required disaggregating the distinct computational phases of inference, routing requests intelligently based on visual context, and building custom support for Pinterest’s proprietary image embeddings. And it required a serving AI infrastructure that was flexible enough to expand across not one product surface, but many—recommendation rerankers, content safety systems, OCR pipelines, and conversational agents—without each team rebuilding the same foundation from scratch.

Pinterest built its platform  around two foundational choices: NVIDIA Blackwell B200 GPUs as the compute layer, and NVIDIA Dynamo as the serving orchestration framework.

Pinterest standardized on B200 GPUs for their industry-leading inference economics. Blackwell’s architectural advances in low-precision data formats, memory bandwidth, and transformer engine made it the right foundation for workloads where prefill is the dominant cost. In preliminary benchmarking for Pinterest Assistant, B200 instances delivered more than 2x latency improvement over Hopper-generation hardware.

For orchestration, Pinterest selected NVIDIA Dynamo for its inference-engine-agnostic design, native compatibility with Pinterest’s service discovery infrastructure, and a high-performance router. Equally important was the close working relationship Pinterest built with the NVIDIA Dynamo team, who provided hands-on support throughout development. The result is an end-to-end generative AI serving platform centered on Dynamo, with vLLM as the inference engine that Pinterest’s teams across the company can now build on through a common Chat Completions API.

That platform approach supported Pinterest’s cost and monetization agenda. Pinterest achieves inference transaction costs at less than 8% of comparable closed proprietary models by post-training open models on NVIDIA GPUs using its unique data, giving the company headroom to expand AI-powered experiences. On the advertising side, Pinterest’s AI creative optimization tools are already translating into performance gains: Smart Assembly, a Pinterest Performance+ capability that automatically builds and serves the best-performing ad from advertiser-uploaded images, delivered a 6% average click-through-rate improvement in early alpha testing.