LanceDB lands $57.5M to build unified data platform for multimodal AI training
The database startup raised its Series B from Bessemer Venture Partners and Innovation Endeavors, bringing total funding to $98.5 million. The company's multimodal lakehouse aims to let researchers train AI models without weeks of data preparation.

LanceDB Inc. announced a $57.5 million Series B funding round to expand its database platform for accelerating artificial intelligence model development. 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 fundamental infrastructure challenge in AI: researchers currently scatter different data types across separate systems and spend weeks or months assembling training datasets. LanceDB's multimodal lakehouse consolidates text, images, video, audio and other data formats into a single platform, allowing teams to access, search, prepare and train on data without extensive infrastructure management.
"Typically, research teams have to put different parts of their dataset in different places, and they lose weeks or months getting a whole training dataset ready for that single training run at the end of the day," co-founder and Chief Executive Chang She said. "There's a missing data layer for AI today, and that's what we're trying to build."
LanceDB built its platform on Lance, an open-source data format engineered for AI workloads that handles structured data alongside images, video and large files. Unlike traditional analytics formats optimized for processing entire tables, Lance retrieves and modifies individual records rapidly, making it suited for searching and preparing massive training datasets at scale.
The company's lakehouse architecture combines the raw storage capacity of a data lake with database-style organization, querying and management capabilities. The platform connects directly to customers' existing cloud storage, keeping data in place while distributing computational workloads across infrastructure needed to process it.
Handling the scale and complexity of multimodal data
Multimodal AI extends beyond combining text, audio, images and video. Autonomous vehicles, robotics and world models require three-dimensional sensor information, geospatial coordinates, time-series data and annotations describing what sensors detect.
A self-driving vehicle, for instance, captures more than a single image. Its sensors generate three-dimensional environmental representations that must be linked with data identifying roads, pedestrians, vegetation and other objects. The volume of data from lidar 3D clouds, images, video and audio typically ranges from 10 to 20 petabytes per training table, with some workflows reaching 80 to 100 petabytes. Researchers often maintain these massive datasets across multiple systems because conventional platforms lack multimodal AI design.
Assembling, reconciling and enriching different training dataset components can consume weeks or months. LanceDB's approach manages infrastructure without copying or moving data from enterprise or research storage, providing a single unified access point. "The data stays in place," She explained. "It remains yours. We don't need to copy it."
Rather than introducing only a new storage method, LanceDB aims to reduce data access frequency and eliminate the need to copy, convert or move data across networks. This acceleration of what She calls the "AI flywheel"—feeding models quality data as rapidly as possible to develop and deploy them—enables faster iteration. "The faster we can make that spin, the faster research teams can get to higher quality models and allow us to make even more interesting AI applications," She said.
Enterprise adoption and future direction
Runway AI Inc. adopted LanceDB's enterprise platform as its training needs expanded. Netflix Inc. constructed a media data lake on the technology for content creation. Microsoft Corp. uses the platform as a data layer for training models on tens of billions of rows of image and code data. These organizations consolidated data management, search, curation and feature engineering into unified workflows, enabling researchers to prepare petabyte-scale training datasets without managing separate infrastructure.
"A training example at Runway doesn't stay still," said Runway Chief Technology Officer Kamil Sindi. "It starts as a video and a prompt, then we add embeddings, resolution and duration metadata, whatever the next experiment needs. We first started using the Lance format in 2024 because it was the only option that handled unstructured data, column appends, distributed reads and ad-hoc SQL all at once. We've since become a LanceDB Enterprise customer."
LanceDB plans to deploy the new funding toward expanding its multimodal data platform and accelerating the AI model development cycle through improvements in large-scale data curation, feature engineering and data loading to maintain efficient graphics processing unit utilization during training. The company will also expand support for emerging AI workloads including video generation, world models and robotics, where researchers manage growing volumes of first-person sensor data, synthetic data and other multimodal information.


