Use Case Intelligence RSS Search use cases Search
NVIDIA Media and Entertainment Global

Pinterest

Marketing Personalization

Pinterest needed to serve vision-language AI for shopping and discovery at internet scale, but multimodal requests involving thousands of images created prefill-heavy inference, memory pressure, and routing challenges that conventional LLM infrastructure could not handle efficiently.

Read the original story Explore all use cases
14,000 NVIDIA GPUs
25x more visual context per request
369x faster time to first token
44x faster end-to-end latency

Reported impact

Business outcome

Pinterest scaled multimodal AI across a 14,000-GPU fleet, cut inference transaction costs to less than 8 percent of prior closed-model costs, improved ad click-through rates by 6 percent, processed 25x more visual context per request, and accelerated end-to-end latency by up to 44x for Pinterest Assistant.

  • Cost Reduction
  • Customer Experience
  • Revenue Growth
  • Scalability

Solution

Pinterest rebuilt its multimodal serving layer on NVIDIA Blackwell and NVIDIA Dynamo so visual discovery workloads could run as a shared production platform instead of isolated model stacks. NVIDIA Blackwell B200 GPUs provide the compute layer for Pinterest's vision-language models, while NVIDIA Dynamo orchestrates routing, disaggregated inference phases, and service integration through a common Chat Completions API. Pinterest pairs Dynamo with vLLM and a proprietary image encoder that serves precomputed visual embeddings rather than reprocessing raw pixels on every request. That design allows the Pinterest Assistant and other product surfaces such as rerankers, content safety, OCR, and signal generation to share the same serving foundation. The platform also uses AI Configurator, DynoSim, and Dynamo Planner to simulate deployment options, optimize parallelism, and autoscale compute against latency SLAs as traffic changes.

Data flow

User search, shopping, and conversational requests enter Pinterest Assistant or other product surfaces, proprietary image embeddings and visual context are routed through Dynamo and vLLM on Blackwell GPUs, the serving layer combines text and image understanding in real time, and the resulting multimodal outputs flow back into discovery, ads, and safety experiences.

Solution architecture

6 components · 4 layers
  1. Compute
    • NVIDIA Blackwell B200 GPUs Supply the inference compute needed for prefill-heavy vision-language workloads at Pinterest scale.
  2. Serving
    • Pinterest Assistant Provides the user-facing visual discovery experience that consumes the multimodal serving platform.
    • vLLM Serves as the inference engine under Dynamo for Pinterest's production multimodal workloads.
  3. Orchestration
    • NVIDIA Dynamo Acts as the inference orchestration layer that routes requests and exposes a common Chat Completions API.
    • Dynamo Planner Autos-scales VLM and LLM compute in response to real-time traffic patterns and latency targets.
  4. Governance
    • AI Configurator Simulates deployment configurations so Pinterest can tune hardware and parallelism with less manual effort.

Architecture clues

  • Pinterest uses precomputed visual embeddings through Dynamo and vLLM instead of re-encoding raw pixels on every request.
  • The same multimodal infrastructure is shared across assistant, ads, safety, OCR, and ranking use cases.
  • The serving stack disaggregates different phases of inference rather than treating VLM traffic like standard LLM inference.

Evidence from the source

Pinterest achieves inference transaction costs at less than 8% of comparable closed proprietary models.
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 built its vision-language model (VLM) serving stack on NVIDIA Blackwell.
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.
With NVIDIA Dynamo, Pinterest Assistant can process 25x more visual context per request ... with projection embeddings delivering up to 369x faster time to first token and 44x faster end-to-end latency.
Transparency Evidence & methodology 100/100 reference value · 89% confidence · 1 source

Review how this reference was scored and trace each normalized claim back to its documented source.

01

Scoring methodology

How scores are calculated

Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.

Customer reference value100/100
  • Problem clarity20/20
    Measures whether the customer's operational problem is specific enough to compare.
  • Implementation depth25/25
    Rewards named components, data flow, integrations, and architecture detail.
  • Outcome specificity20/20
    Measures whether the reported benefit is concrete rather than a generic claim.
  • Quantified impact25/25
    5 metrics describe measurable impact.
  • Deployment scale10/10
    Rewards evidence of a live or scaled deployment and meaningful integration scope.
Evidence confidence89%
  • Source authority20/25
    The main evidence is an official vendor-published customer story.
  • Claim traceability21/25
    10 of 12 key fields link to a named source.
  • Direct evidence20/20
    5 direct source quotes retained.
  • Date provenance10/10
    The publication date comes from the source page or publisher.
  • Source transparency8/10
    1 source reference document publisher, URL, authority, retrieval, and supported fields.
  • Implementation specificity10/10
    Rewards source-backed components, data flow, quantified impact, and a specific outcome.
02

Traceability

Sources & provenance

Field coverage shows which facts each source supports. Dates are shown only when documented by the source.

  1. NVIDIA

    Pinterest customer story

    First-party
    https://www.nvidia.com/en-us/case-studies/pinterest/
    Source type
    Vendor-published customer story
    Published
    Sep 14, 2026
    Supports
    • Customer
    • Business problem
    • Solution
    • Products
    • Solution components
    • Data flow
    • Integrations
    • technical area
    • use case category
    • Business outcome
    • outcome category
    • Metrics
    • Architecture
    • evidence quotes
    • Publication date

Date evidence

  • Published on source page Sep 14, 2026 · NVIDIA · First-party
    Open date source
Not separately sourced Industry, Region, Company size. Treat these normalized attributes as lower-confidence until a dedicated source is added.

Related

Similar use cases

See all like this
Google Cloud Modified · Oct 9, 2026 HOLYWATER TECH Media and Entertainment Ukraine HOLYWATER TECH wanted to produce long-form fantasy video content faster and more consistently, but traditional filming and fragmented AI tools made it hard to keep characters, settings, and sound synchronized across… 10x ARR growth for My Muse since implementing Marketing Personalization BigQuery · Gemini · Gemini platform Google Cloud Modified · Oct 8, 2026 Meitu Media and Entertainment China Meitu needed one cloud platform that could supply advanced image-generation models, developer productivity tooling, performance telemetry, and large-scale analytics so it could improve AI imaging quality for a global… As of December 2025, Meitu had 276 million monthly active users around the world. Marketing Personalization BigQuery · Firebase · Gemini Code Assist Databricks Sep 30, 2026 Showtime Media and Entertainment United States Showtime wanted to leverage data to drive content strategy, but they struggled with scaling limitations of legacy systems and inefficient data pipelines. Unfortunately, legacy technology architectures were pulling teams… 6x Faster data pipelines enables faster decision making Data Lake / Data Warehouse ModernizationMarketing Personalization Databricks Data + AI Platform