LanceDB Raises $57.5M Series B for Multimodal AI Data
A dataset can outlast the model trained on it. LanceDB is building a business around keeping that dataset useful as researchers change what they want to learn from it, carrying images, video and the information derived from them through successive experiments.
The San Francisco company announced a $57.5M Series B on November 17, 2026, co-led by Bessemer Venture Partners and Innovation Endeavors. Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator participated. LanceDB says the financing brings total funding to $98.5M and will support the next phase of its multimodal lakehouse, including more of the preparation and access work surrounding model training.
CEO and cofounder Chang She and CTO and cofounder Lei Xu are taking that expansion into a demanding market. A research team needs to preserve what it already knows about its data while continually adding something new: another embedding, a different caption, a feature that makes an overlooked training example searchable. The commercial opportunity sits inside that recurring work.
The Open-Source Format Beneath the Enterprise Sale
Lance and LanceDB occupy different layers of the proposition. As Bessemer explains in its investment account, Lance provides the open-source storage format, table format and catalog specification; LanceDB supplies the database and a commercial enterprise platform around those foundations. A table can keep raw media alongside metadata, vectors and computed features, with SQL and vector search available over the data.
That arrangement matters when an experiment changes the shape of a dataset. Adding embeddings to an existing table can mean writing the new column rather than rewriting the whole collection. For the researcher, the practical attraction is continuity: the images remain reachable while a new representation makes a different question possible. For the platform team, the same choice affects how much infrastructure must be rebuilt each time research changes direction.
The enterprise offering adds the capabilities needed to operate across larger workloads, including distributed indexing, ingestion, feature engineering and training data loaders. Bessemer describes deployment in a customer's cloud or on premises. Open-source adoption therefore gives LanceDB a route into the environment where a future commercial decision may occur, close to the data and the people already using it. Supporting that environment becomes an ongoing operating responsibility as more experiments depend on it.
Founders Who Have Worked on Both Sides of the Data
Chang She helped build pandas and ran ML engineering at Tubi, according to Bessemer and the official Y Combinator company profile. Lei Xu's background includes HDFS at Cloudera and ML infrastructure at Cruise, Bessemer says. Their experience connects familiar analytical tooling with the video, images and sensor streams that occupy more of the training workflow LanceDB now serves.
The financing history shows how long the multimodal platform has been under construction. LanceDB's June 24, 2025 announcement disclosed a $30M Series A led by Theory Ventures, with plans to expand feature engineering, model training and enterprise scale. The new Series B brings Bessemer and Innovation Endeavors into the lead while Theory continues investing. The $98.5M cumulative figure comes from the current company announcement; it should be read separately from the $57.5M new round.
LanceDB has not disclosed a valuation in the checked announcements. Its reported revenue growth of more than 10x year over year and more than 120M cumulative open-source library downloads offer evidence of commercial and developer activity, with different meanings. The revenue claim lacks a disclosed base, and a package download is a distribution event rather than a count of paying customers. Those distinctions keep the financing story attached to what the company actually reported.
The Research Question Keeps Changing the Dataset
Theory Ventures' account of its returning investment describes researchers at Luma launching new captioning columns themselves, sometimes through an agent. The example makes the architecture easier to understand than another list of database features. A change in a model can create a need for new information across existing media, and the people testing the idea need a practical way to produce and inspect it.
Lei Xu's technical explanation of physical-AI data mining follows that problem into sensor logs and video. A timestamp is already structured information; a maneuver or pedestrian interaction may require additional computation before a researcher can search for it. Feature engineering helps turn recorded experience into attributes that can be filtered, compared and assembled into training or evaluation datasets. It gives the next question somewhere concrete to run.
The operating implication is that data preparation continues after a collection has first been stored. A new hypothesis can send a team back through its history, with a new feature to compute and a new subset to inspect. LanceDB's planned work on preparation, loading and caching reaches further into that cycle. The work also puts the company close to the decisions that determine how a team spends its research time, where an infrastructure product earns a place in the daily routine.
Where the New Capital Meets Daily Operations
LanceDB names Netflix, Microsoft AI, Runway, Luma AI and Midjourney among its customers. The breadth of those company-reported relationships gives the platform several kinds of production work to accommodate, while the public announcement leaves individual contract economics undisclosed. NVIDIA is separately described as an open-source user and partner, a relationship that should remain distinct from the customer list.
The broader funding conversation includes robotics data infrastructure and systems for turning AI failures into training data. These are neighboring parts of the data problem, with their own buyers and product boundaries. For LanceDB, the opportunity is to make the same underlying data useful across storage, retrieval, feature development and training access, reducing the need to maintain disconnected representations of every experiment.
The company says it is hiring in engineering, field engineering and go-to-market as that scope expands. Each additional deployment brings another research team with existing data, established infrastructure and its own way of deciding whether an experiment was worth running. The Series B finances more of the work that follows those teams into production, where yesterday's dataset will acquire tomorrow's questions.
AI Infrastructure funding, last 30 days
DevCuration's funding database tracked 34 AI Infrastructure rounds totaling $30.2B in disclosed capital over the past 30 days. Recent deals we covered:
- Sqwish Labs Raises $3M Pre-Seed for AI Decision ModelsPre-Seed · $3M · Nov 19
- Linq Raises $90M Series B, Adds Ben Horowitz to BoardSeries B · $90M · Nov 18
- Compute Desk Raises $45M Seed for GPU Compute MarketsSeed · $45M · Nov 18
- BXDC Agrees to $5B AWS Data Center Portfolio Acquisition$5B · Nov 17
- Daisytuner Raises €6M Seed to Expand Daisyflow AI InferenceSeed · €6M · Nov 16
Frequently Asked Questions
Why does adding a feature matter to a multimodal training dataset?
A new caption, embedding or derived signal can make existing media useful for another research question. LanceDB allows new columns to be added to Lance tables without rewriting the entire table, connecting feature development with data retrieval and training access.
How do Lance and LanceDB Enterprise differ?
Lance provides open-source data-format foundations. LanceDB builds the database and commercial platform around them, with Enterprise capabilities including distributed indexing, ingestion, feature engineering and training data loaders.
What development work does the Series B support?
LanceDB says it will expand support for curation, enrichment and experimentation around multimodal training data. Its stated plans include extending preparation capabilities and improving loading and caching for training runs.
How should readers interpret LanceDB’s reported growth figures?
The company reports more than 10x year-over-year revenue growth and over 120M cumulative open-source library downloads. The revenue base is undisclosed, and downloads should not be interpreted as unique users or paying customers.
Where the Money Moved
The intelligence briefing of the innovation economy. Funding, M&A, debt and fund closes, read as market signal rather than deal announcements.
Subscribe to Where the Money Moved