LanceDB, the company behind the open-source Lance data format, announced a $57.5 million Series B on Tuesday, November 17, co-led by Bessemer Venture Partners and Innovation Endeavors. The round brings its total funding to $98.5 million, according to SiliconANGLE, and will pay for tools that help AI labs curate, transform and load the huge image, video and sensor datasets used to train models. No valuation was disclosed.
Who invested and what the money is for
Besides the two lead investors, the round includes Theory Ventures, which led LanceDB’s $30 million Series A in June 2025, along with Dimension, Swift Ventures, CRV and Y Combinator.
LanceDB says the money will help it bring its “multimodal lakehouse” to researchers, model builders and platform teams. SiliconANGLE reports the focus is on large-scale data curation, feature engineering and data loading fast enough to keep thousands of GPUs busy, with video generation, world models and robotics named as target workloads. Chief executive Chang She, a co-author of pandas, told the outlet that data stays in the customer’s own storage instead of being copied.
Unite.AI reports company figures we could not verify, including revenue up more than tenfold since the Series A and over 120 million library downloads. Named users include Netflix, Microsoft, Runway, Luma AI, Midjourney, ByteDance, Applied Intuition and Nvidia’s Cosmos team.
What Lance and LanceDB actually do
Lance is three things in one specification: a file format, a table format and a catalog spec, released under the Apache 2.0 licence. A single Lance table can hold raw images or video next to their metadata, embeddings and other features, with SQL, full-text (BM25) and vector search on that data.
First, Lance is built for fast random access, so a training job or a retrieval query can fetch individual rows without scanning large chunks of a file. Second, it can add a new column, such as a fresh set of embeddings or captions, without rewriting the whole table. Lance also keeps automatic, zero-copy versions of each table, with time travel, tags and branches.
The open-source LanceDB library sits on top, while LanceDB Enterprise adds distributed indexing, high-throughput ingestion and training data loaders, running in the customer’s cloud or on premises.
How Lance compares with Parquet and other vector databases
Against Parquet. Apache Parquet is the standard columnar file format for analytics. It stores data by column and uses heavy compression and encoding, which makes it excellent for scanning whole columns in data warehouses. Lance keeps the columnar idea but optimises for the access patterns of AI work: picking random samples for training, fetching a handful of matching rows during retrieval, and appending new columns often. The Lance README claims up to 100 times faster random access than Parquet or Iceberg, based on a preliminary benchmark on the Oxford Pet dataset. Treat that as a vendor figure. Parquet still has far wider engine support, though Lance integrates with Arrow, pandas, Polars, DuckDB, Spark and Ray.
Against dedicated vector databases. Services such as Pinecone, Qdrant, Weaviate and Milvus, covered in our guide on how vector databases work and how to choose one, keep embeddings in their own storage and serve low-latency similarity search. LanceDB takes a different approach: vectors live as one column in an open-format table on object storage, next to the source files and metadata. That avoids copying and suits teams whose retrieval data is also training data; a pure, latency-critical retrieval service may still fit a dedicated database better.
For a retrieval-augmented generation (RAG) app, our explainer on how retrieval-augmented generation works covers the pipeline, and our piece on hybrid search and reranking for RAG explains why combining BM25 and vectors, which Lance supports in a single table, often improves results.
What to watch next
- Whether LanceDB publishes independent or reproducible benchmarks against Parquet, Iceberg and dedicated vector databases.
- How Nvidia’s Cosmos team progresses on training directly from Lance tables, as Theory Ventures describes.
Key takeaways
- LanceDB raised a $57.5 million Series B on November 17, 2026, co-led by Bessemer Venture Partners and Innovation Endeavors.
- Total funding reaches $98.5 million; no valuation was disclosed.
- Lance is an open-source format that stores media, metadata and embeddings in one table with fast random access and cheap column additions.
- Bessemer likens Lance’s role for multimodal AI data to Parquet’s role for analytics tables.
- Performance and growth figures in the announcement are company claims, not independent results.
Frequently asked questions
What is LanceDB?
LanceDB is a company and an open-source database built on the Lance data format. It stores images, video, text, metadata and vector embeddings together and supports vector search, full-text search and SQL on the same tables.
Is Lance a replacement for Parquet?
Lance is designed for AI workloads that need fast random row access and frequent new columns, where Parquet is optimised for scanning whole columns in analytics. Many teams will use both, keeping Parquet for warehouse analytics and Lance for training and retrieval data.
Is LanceDB free to use?
The Lance format is open source under the Apache 2.0 licence, and LanceDB also offers a free open-source library. LanceDB Enterprise, which adds distributed indexing and data loaders, is a commercial product.




