A Data Foundation for Multimodal AI: Our Series A Investment in LanceDB

LanceDB’s customers, including Runway, Midjourney, and Character.ai, manage petabytes of training data. Before a training run starts, someone has to find the right examples, caption them, embed them, and filter out the bad ones. Then the data has to reach the GPUs fast enough to keep them busy, and the finished model needs much of it again at inference time. Most teams run each step on a different system. LanceDB puts them on one.

We’re excited to lead LanceDB’s $30 million Series A, with participation from CRV, Y Combinator, Databricks Ventures, RunwayML, Zero Prime, Swift, and other investors. Co-founders Chang She and Lei Xu are building a multimodal lakehouse: a shared data foundation for training and serving multimodal models.

AI changes what a data platform needs to do

An image can produce a caption, an embedding, and a set of labels. A video can produce thousands of frames and many more features. Each representation helps answer a different question, and each needs to remain connected to its source.

That creates work across the entire development process. Researchers need to explore datasets and prepare training examples. Engineers need to generate features, build search, and serve results inside a product. As models improve, teams need to revisit the same data and derive something new from it.

We believe the opportunity is to make those workflows part of a coherent data platform. Every handoff that requires another export, transformation, or custom pipeline slows the next experiment.

Keeping the data and its representations together

LanceDB’s multimodal lakehouse builds on the open-source Lance columnar format and LanceDB’s search capabilities. Both are widely used in open source: the LanceDB and Lance repositories have about 19,000 GitHub stars combined, and the LanceDB Python package is downloaded about 7.7 million times a month. Its Enterprise offering brings together search, exploratory data analysis, feature engineering, and training workflows.

The platform supports work with raw content, embeddings, and structured metadata. Developers can define transformations in Python and use distributed execution to process their datasets. This gives teams a common foundation for moving from an initial exploration to the data their applications need in production.

For a builder, the value is practical: more time improving the application, with less work maintaining the infrastructure between each step.

Infrastructure tested by demanding AI workloads

In its funding announcement, LanceDB reported Enterprise deployments at Runway, Midjourney, and Character.ai, supporting workloads involving billions of vectors and petabytes of training data. These are environments where data infrastructure directly affects how quickly a team can build.

We also use Lance internally at Theory, as Tom described when announcing the partnership. That gives us a direct appreciation for the problem Chang and Lei are solving.

We believe multimodal AI will make this foundation valuable to a much broader set of developers. We’re thrilled to partner with Chang, Lei, and the LanceDB team as they make complex data easier to work with, from the first experiment to production.