LanceDB raises $57.5 million to tackle the data work behind AI training
The company plans to improve how researchers organize and feed images, video and other data into training, while leaving that data in customers’ storage.
Loading page…
The company plans to improve how researchers organize and feed images, video and other data into training, while leaving that data in customers’ storage.
Listen to this story
LanceDB is positioning its open-source Lance format and paid Enterprise platform as a way to prepare multimodal training data in customers’ existing cloud or on-premises storage, without copying it into a separate system. Teams can evolve datasets and query images, video and structured records with SQL or vector search, while distributed tooling targets large training pipelines. The company announced a $57.5 million Series B on November 17, 2026, co-led by Bessemer Venture Partners and Innovation Endeavors, bringing reported total funding to $98.5 million; proceeds will support data preparation and loading.
Bessemer says teams can add an embeddings column to a petabyte-scale Lance table without rewriting it; the format also supports SQL and vector search.
Runway adopted Lance in 2024 and later became an Enterprise customer, citing mutable columns, distributed reads and SQL queries for evolving video-training examples.
Microsoft uses LanceDB as a training-data layer across tens of billions of image and code rows; SiliconANGLE reports Netflix built a media data lake on the technology.
LanceDB promises less time assembling AI training data, a task its CEO says can consume weeks or months. On November 17, 2026, the company announced a $57.5 million Series B co-led by Bessemer Venture Partners and Innovation Endeavors to develop a platform that organizes and prepares data where customers already store it.
Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator also participated. The company said the round brings its total funding to $98.5 million. Bessemer described those five investors as continuing backers, joining the two firms leading the financing.
In an interview with SiliconANGLE, co-founder and CEO Chang She described research teams keeping different pieces of a dataset in separate places. Before a training run, they must assemble those pieces into something usable. LanceDB’s answer is a shared system for finding, organizing and preparing different kinds of data, rather than maintaining separate systems for each type.
The foundation is Lance, an open-source format that specifies how data is stored and organized into tables. It is designed for multimodal data: material spanning forms such as images, video and structured records. LanceDB is both the company’s name and the database built on that format; its commercial enterprise platform adds tools for operating at larger scale.
A Lance table can hold raw images or video alongside descriptive information and numerical features used by models. According to Bessemer’s investment announcement, teams can add a column of embeddings—numerical representations used for search—to a petabyte-scale table without rewriting it. The format also supports SQL queries and vector search directly on the data.
Lance also supports retrieving and modifying individual records, rather than only processing whole tables. That design serves work such as searching and preparing large training datasets. LanceDB says its platform connects to existing cloud storage and distributes the computing work while leaving the underlying data in place, without copying it into another store.
The open-source library and format offer an entry point for developers. LanceDB Enterprise is the paid layer, and can run inside a customer’s cloud or on its own premises. Bessemer identifies several additions:
Runway offers a concrete example of why changing datasets matter. Its chief technology officer, Kamil Sindi, said a training example can begin as a video and prompt, then gain embeddings, resolution, duration and other information as experiments evolve. Runway began using Lance in 2024 and later became an Enterprise customer, he said, citing its combination of column additions, distributed reads and SQL queries.
Other uses extend beyond video-model experiments. SiliconANGLE reports that Netflix built a media data lake on the technology for content creation, while Microsoft uses the platform as a training-data layer spanning tens of billions of rows of image and code data.
She said the new funding will support better large-scale data preparation and loading, with the goal of keeping thousands of graphics processing units supplied during training. LanceDB also plans deeper support for video generation, world models and robotics—workloads that combine growing volumes of sensor information, synthetic data and other training material.
Loading discussion...
Join the conversation
Explain what openness would—or would not—change for you.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.