How Scribd Cut GenAI Costs 90% and Boosted Sign-Ups with Databricks

Scribd, Inc. connects millions of users to knowledge across three distinct platforms: Scribd, a user-powered library; SlideShare, a repository of presentations; and Everand, a subscription service for audiobooks and eBooks. The company’s content library has grown to more than 250 million pieces — multilingual, media-rich, and highly varied in format and quality. At that scale, ensuring content is discoverable, properly tagged, and free of low-quality material is not a manual task.

The company had identified generative AI as the right approach to enrich metadata, improve search, and moderate content at scale. But its existing infrastructure made it difficult to act on that opportunity. Data pipelines were fragmented across separate tools for transformation, LLM experimentation, and model serving. The feedback loop between data scientists and production teams was slow. There was no shared environment where engineers, data scientists, and product managers could collaborate on AI development without constant context switching between platforms.

Scribd, Inc. built its AI foundation on the Databricks Data Intelligence Platform, which unified ingestion, transformation, model development, and inference into a single governed environment. Data scientists run Spark jobs, build features, experiment with LLMs, and test ideas inside Databricks Notebooks without changing tools. Databricks Model Serving handles both real-time and batch inference — serverless endpoints for low-overhead prototyping, then seamless scale-up when production jobs process tens of millions of prompts against the company’s largest tables. The team selects open-weight models — Llama, Mistral, Claude — based on the specific task at hand, served through Mosaic AI. Delta Lake and Unity Catalog provide the data reliability and governance layer underneath. Lakeflow Jobs orchestrates batch workloads. The Databricks team was embedded with Scribd during critical early deployments, helping tune model performance on AWS and navigate GPU constraints in real time.

The most commercially visible AI application is auto-generated content metadata. Every piece of user-uploaded content now receives AI-generated titles, summaries, descriptions, and category tags, making it more discoverable through search and more visible in SEO. This metadata enrichment directly contributed to a 7% increase in new user sign-ups and a 7% reduction in churn. At the same time, streamlining how large language models are built and run through Databricks’ batch inference capabilities cut the cost of running those models by more than 90%, making it economically viable to expand GenAI across moderation, discovery, and chat-based features like Everand’s Ask AI.

Scribd, Inc. is extending this foundation further. The next wave of AI features — intelligent topic extraction, in-document navigation, and slide-level search on SlideShare — are designed to let users jump directly to the content they care about within a document. These features are only possible because the company now operates AI and data on a single scalable platform rather than a collection of disconnected services.