Startups
LanceDB raises $57.5M for multimodal AI training data platform
Series B funding will expand infrastructure that manages petabyte-scale datasets across text, video, images and sensor data without moving files.
Key takeaways
- LanceDB raised $57.5 million Series B led by Bessemer Venture Partners and Innovation Endeavors, bringing total funding to $98.5 million.
- The platform manages multimodal AI training datasets reaching 10-100 petabytes without copying or moving data from existing cloud storage.
- Runway AI, Netflix and Microsoft use LanceDB to consolidate data management for video generation, media workflows and large-scale model training.
- The Lance open-source format supports rapid retrieval of individual records rather than processing entire tables, optimizing for AI search and preparation tasks.
- New funding will expand data curation and loading systems to keep GPU clusters efficiently supplied during training runs for video, robotics and world models.
LanceDB has closed a $57.5 million Series B round to expand its multimodal data platform designed to accelerate AI model training workflows, the company announced today. Bessemer Venture Partners and Innovation Endeavors led the round, with participation from Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator. The funding brings LanceDB's total capital raised to $98.5 million.
The company addresses a persistent infrastructure bottleneck: research teams routinely spend weeks or months assembling training datasets scattered across incompatible storage systems. LanceDB's platform consolidates access to text, images, video, 3D sensor data and other modalities without requiring organizations to copy or relocate their data.
"There's a missing data layer for AI today, and that's what we're trying to build," co-founder and CEO Chang She told SiliconANGLE.
How the platform works
LanceDB built its technology on Lance, an open-source data format optimized for AI workloads. Unlike traditional analytics formats that process entire tables, Lance supports rapid retrieval and modification of individual records—critical when searching and preparing massive training datasets.
The company packages this format into a lakehouse architecture that combines the raw storage capacity of data lakes with database-style query and management capabilities. The platform connects directly to customers' existing cloud storage and distributes computational workloads across the necessary infrastructure, leaving the underlying data in place.
This approach matters because modern AI training datasets regularly reach 10 to 20 petabytes per table, with some workflows approaching 80 to 100 petabytes. Autonomous vehicle models, for example, must process 3D lidar point clouds alongside images, video, geospatial coordinates, time-series telemetry and semantic annotations—all of which need to remain synchronized during training.
Production deployments
Runway AI adopted LanceDB's enterprise platform as its video generation training requirements scaled. "A training example at Runway doesn't stay still," said CTO Kamil Sindi. "It starts as a video and a prompt, then we add embeddings, resolution and duration metadata, whatever the next experiment needs."
Netflix built a media data lake on the technology for content creation workflows. Microsoft uses the platform as a data layer for training models on tens of billions of rows combining image and code data. These implementations consolidate data management, search, curation and feature engineering into unified workflows that operate at petabyte scale.
Why it matters
The speed of the AI development cycle increasingly depends on how quickly teams can curate and load training data. She describes this as the "AI flywheel"—the faster research teams can prepare quality datasets, the faster they can iterate toward higher-performing models. Infrastructure that eliminates weeks of data movement and format conversion directly accelerates that cycle.
LanceDB plans to use the new capital to improve large-scale data curation, feature engineering and data loading systems that keep thousands of GPUs efficiently supplied during training runs. The company will also deepen support for emerging workloads including video generation, world models and robotics, where researchers must manage growing volumes of first-person sensor data and synthetic training examples.
Details of the funding were first reported by SiliconANGLE.