LanceDB: The data layer for multimodal AI
Bessemer Venture Partners co-leads LanceDB’s $57.5M Series B to build the new data stack for AI research.
Mike Droesch, Eric Kaplan, Libbie Frost & Jason Scheller
Underneath every AI model lies a pile of multimodal data. Even a text model is trained on scraped web pages full of images, PDFs, and videos, while models for robotics, autonomy, and media are trained entirely on multimodal data. That pile is growing faster than ever because AI itself is producing it. Robots and vehicles log sensor streams around the clock, cameras and microphones capture more of the world than anyone can review, and generative models produce billions of images and videos. At the same time, open source is winning the infrastructure layer, and the smartest companies are beginning to own their intelligence by post-training open models on proprietary data.
All roads lead to the same problem: the systems we use to store and move data were designed for rows of numbers and strings of text that now strain when trying to curate, search, and stream petabytes of multimodal data into training or inference.
LanceDB was built for exactly this moment by a team that lived the problem. CEO Chang She is a co-author of pandas and ran ML engineering at Tubi, where he built training pipelines on Parquet and concluded the format needed to be rebuilt from first principles. CTO Lei Xu has been building file systems his whole career, from a searchable file system he founded during his PhD to HDFS at Cloudera to the ML infrastructure behind Cruise's self-driving fleet. It didn’t take us much time with Chang and Lei to be convinced this is the team to bet on. This is why Bessemer is leading LanceDB's $57.5M Series B alongside Innovation Endeavors, with continued support from CRV, Theory Ventures, Dimension, Swift Ventures, and Y Combinator.
From primitives to platform
Lance is an open-source file format, table format, and catalog spec that defines how multimodal data is laid out in storage so software can read it efficiently—similar to what Parquet does for the analytics tables underneath every data warehouse. A single Lance table holds raw images or video alongside their metadata, features, and vectors. Vector search and SQL work natively on the data, and a team can add a column of embeddings to a petabyte table without rewriting it. LanceDB is the company and the database built on that format. Its open-source library and format is how many developers first use Lance, and LanceDB Enterprise is the commercial platform that adds distributed indexing, high-throughput ingestion, training data loaders, and feature engineering, all running inside the customer's own cloud or on-prem deployment.
Users download LanceDB tens of millions of times each month, up 6x within a year, and many of the most respected AI researchers in the world are adopting it on their own. Some of the most sophisticated frontier labs in the world use it for pre-training at scale, helping them consolidate many separate systems and homegrown pipelines into a single Lance-based platform for petabyte-scale curation and feature generation. AI companies like WorldLabs, Applied Intuition, and ByteDance use it to post-train, and enterprises like Netflix use it for production retrieval and ML ops.
The why now for multimodal builders
Most prior generations of data infrastructure have been won by an open-source project the community trusted, paired with a company that made it work at scale. We believe the AI generation will be no different. Frontier labs curate petabytes of text, images, and video for each training run. Robotics companies post-train on their fleets' sensor logs. Enterprises want to fine-tune on proprietary data to own their own intelligence, and agents retrieve context at inference time. Underneath, they all need the same thing: data primitives that can curate, version, search, and serve multimodal data at scale. LanceDB is the most credible candidate to be that primitive, and the company is hiring across every function to keep up.
We're proud to partner with Chang, Lei, and the LanceDB team. If you're building on video, audio, images, sensor data, or anything else that doesn't fit in a spreadsheet, LanceDB is where it belongs.
Contributors

Mike Droesch
Partner
Mike Droesch is a partner in the Boston office primarily focusing on AI applications, cybersecurity, supply chain software, and B2B marketplaces. He currently serves on the board of directors for Tackle, Curri, Raft, Cypress, Netography, Optimal Dynamics, and VendorPM.
Previously, Mike worked with a few venture-backed startups in Boston, including two Bessemer portfolio companies, Fuze and InsightSquared, where he supported sales operations and analytics efforts. He started his career as a systems engineer and then spent nearly three years in management consulting at Navigant, where he helped companies build new businesses around emerging energy and infrastructure technologies in the U.S. and Middle East.
He holds an MBA from the MIT Sloan School of Management, a Master’s Degree in naval architecture and engineering from the University of Michigan, and a degree in mechanical engineering from Tufts University.

Eric Kaplan
Vice President
Eric is an investor in the New York City office, where he invests across the AI stack from infrastructure, to data infra, to applied AI. Eric is drawn to product-obsessed founders across categories innovating at the frontier.
Prior to joining Bessemer, Eric worked at FirstMark Capital and Inspired Capital investing in early stage businesses. Prior to working as an investor, Eric was an early growth product manager at Via, a global provider of on-demand and transportation software. While at Via, Eric managed user funnels, account creation, and global payments. He started his career in management consulting at Monitor Deloitte.
Eric earned his B.A. with honors in philosophy, politics, and economics from the University of Pennsylvania and his MBA from Harvard Business School. At Harvard, Eric was actively involved the entrepreneurship ecosystem.
In his free time, Eric enjoys drawing, historical fiction, and Seinfeld trivia.

Libbie Frost
Investor
Libbie Frost is an investor in the New York office where she focuses on vertical software, fintech, and healthcare.
Before joining Bessemer, Libbie worked at Accel on their Growth Team in Palo Alto and Insight Partners on their Diligence and Growth Strategy Team in New York City.
Libbie graduated magna cum laude from the University of Notre Dame where she earned her BBA in Finance and BA in Industrial Design. While in school, she helped launch and grow the Notre Dame Venture Capital Club and was an alum of Girls Who Invest. Libbie has studied the Italian Renaissance in Florence, Italy, and in her free time she enjoys painting murals, running with friends, and reading a nice biography.

Jason Scheller
Investor
Jason Scheller is an investor based in Bessemer's New York City office, where he focuses on enterprise software and defense tech.
Prior to joining Bessemer, Jason was one of the first hires at Patch, a Series B startup, where he worked in growth.
Jason holds a BBA from the University of Michigan's Stephen M. Ross School of Business. In his free time, he enjoys running, lifting, cold plunging, and hunting (waterfowl, deer, turkey).



