LanceDB raises $57.5M to accelerate AI model training with multimodal data platform

LanceDB Inc., a company that makes database technology to help artificial intelligence models work faster, announced today that it has raised $57.5 million in early-stage funding. The money will help it bring a multimodal lakehouse to researchers, AI model developers and platform teams.
Bessemer Venture Partners and Innovation Endeavors led the Series B funding round. Other investors included Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator. This brings the company's total funding to $98.5 million.
AI is making it easier to create and understand more complex types of data. But creating and using all this data also creates a challenge: companies need the right computer systems to manage it. LanceDB is working to solve this problem with one platform that brings different types of data together. Researchers, businesses and AI agents can use it to find, access, prepare and train on data without spending weeks managing the systems behind it.
"Research teams usually have to store different parts of their datasets in different places. They can then lose weeks or months getting all the data ready for one training run," co-founder and Chief Executive Chang She told SiliconANGLE in an interview. "AI is missing an important data layer today, and that's what we're trying to build."
LanceDB's technology is based on Lance, an open-source data format built for AI tasks. It brings together organized information, images, videos and other large files. Traditional data formats for analytics are designed to process entire tables. Lance can quickly find and change individual records, which makes it useful for searching and preparing very large AI training datasets.
The company uses this format to build a large-scale lakehouse for finding and accessing data. This is important because modern AI models need huge amounts of information for training. A lakehouse combines the ability of a data lake to store large amounts of raw data with database features that help organize, search and manage that data.
"One thing that makes multimodal data different and harder to work with is how large the data files are," She explained.
By adding support for different types of data to the lakehouse, LanceDB gives developers one place to search, organize and prepare AI training data. They do not need separate systems for each type of data. The platform can connect directly to customers' existing cloud storage, so the data stays where it is. It then spreads the computing work across the systems needed to process the data.
Why different types of data are difficult for AI
Multimodal AI is about more than putting text, audio, images and videos together. Self-driving cars, robots and world models also use 3D sensor data, location information, data collected over time and notes that explain what the sensors are detecting.
For example, a self-driving car does more than take a picture. Its sensors may create a 3D model of the area around the car. That model must then be connected to information that identifies roads, people walking, plants and other objects.
The huge amounts of data from 3D lidar scans, images, videos, audio and other sources can take up an average of 10 to 20 petabytes in a single table for many large AI training projects. Some projects can be even bigger, reaching close to 80 to 100 petabytes. Researchers and AI model developers may have to keep large amounts of data in many different systems because traditional platforms were not built for the needs of multimodal AI.
Researchers then have to bring the different parts of a training dataset together, make sure they match and store them in a way that makes them useful for training. Preparing the data can take weeks or months because it must be moved, prepared and processed. LanceDB aims to manage the systems needed for this work without copying or moving data out of the storage systems used by businesses and research teams. This gives users one place to access it all.
"The data stays where it is," She explained. "It still belongs to you. We don't need to copy it."
In other words, LanceDB is not just offering a new way to store data. It aims to make data easier to access without having to copy it, change its format or move it across networks. This can greatly speed up what She calls the "AI flywheel." This is the process of giving AI models good data as quickly as possible so they can be improved and released.
"The faster we can make that process run, the faster research teams can build better models and help create even more interesting AI applications," She said.
Some of the company's customers include Runway AI Inc., which started using LanceDB's business platform as its AI training needs grew. Netflix Inc. built a media data lake using the technology for content creation. Microsoft Corp. also uses the platform as a data layer to train models on tens of billions of rows of image and code data. The platform lets these companies manage data, search, prepare it and create useful features in one connected process. This helps researchers prepare training datasets at petabyte scale without having to manage many separate systems.
"A training example at Runway keeps changing," said Kamil Sindi, Runway's chief technology officer. "It starts as a video and a prompt. Then we add embeddings, information about its resolution and length, and whatever other details the next experiment needs. We first started using the Lance format in 2024 because it was the only option that could handle unstructured data, adding columns, reading data across distributed systems and running SQL queries when needed, all at the same time. Since then, we have become a LanceDB Enterprise customer."
She said LanceDB plans to use the new funding to grow its multimodal data platform and help speed up AI model development. The company especially wants to improve how large datasets are prepared, how useful features are created from data and how data is loaded. These improvements will help keep thousands of graphics processing units supplied with data while they train AI models. The company also plans to improve support for newer AI tasks, including video generation, world models and robotics. These fields require researchers to manage growing amounts of data from first-person sensors, computer-generated data and other types of information.
