LanceDB, the startup building a data layer designed specifically for AI training and inference, announced on Monday a $57.5 million Series B co-led by Bessemer Venture Partners and Innovation Endeavors, with Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator participating. The round brings total funding to $98.5 million, up from a $30 million Series A led by Theory Ventures last June.
The company's growth numbers, which are self-reported, have scaled quickly alongside the multimodal AI boom: revenue has grown more than 10x since the Series A across customers including Netflix, Microsoft AI, Runway, Luma AI and Midjourney, and open-source downloads have grown 6x to past 120 million. Co-founders Chang She and Lei Xu spent nearly two decades building core data infrastructure before founding LanceDB, and She says research teams routinely lose weeks or months assembling training data before a run — the problem the company exists to remove.
The technical foundation is Lance, an open-source format that stores structured records, images, video and other large binary objects together in tables designed for AI workloads. Unlike analysis formats optimized for whole-table scans, Lance supports record-level retrieval and modification, direct SQL queries and vector search on the same data, and column additions — a team can append a column of embeddings to a petabyte-scale table without rewriting it. LanceDB Enterprise adds distributed indexing, high-throughput ingestion, training-data loaders and feature engineering, running inside a customer's cloud or on-premises while the underlying data stays in place.
The customer deployments the company and its investors describe, all company-reported, show the scale it is being asked to handle. Adobe says it manages 40-50 petabytes across roughly 800,000 tables and 220 billion daily events on the platform, and has contributed more than 20 pull requests upstream. Search startup Exa stores 35 petabytes of Lance tables covering 100 billion web pages with crawl-to-search latency under five minutes. Microsoft AI runs LanceDB inside its own security boundary as the training data layer across tens of billions of rows of image and code data, Netflix built a media data lake on the format for content creation, and Runway uses it across its video-generation stack — its CTO Kamil Sindi cites the ability to add embeddings, resolution and duration to a training example as experiments evolve.
Bessemer partner Mike Droesch framed the round as backing "one of the most important infrastructure shifts in AI: the move to training models on multimodal data at massive scale." The new funding goes toward large-scale data preparation and loading — with the stated goal of keeping thousands of GPUs fed during training — and deeper support for video generation, world models and robotics, workloads that pile sensor data, synthetic data and annotations on top of ordinary media.
The company is riding a real architectural argument: AI teams are tired of maintaining parallel copies of data across blob storage, Parquet files and vector databases, and a single lakehouse contract that serves both training and inference is an appealing consolidation. The open question is whether the enterprise platform's performance claims hold at production scale — but with the frontier labs themselves among the reference customers, LanceDB has at least convinced the most demanding buyers to test it.
Comments (0)
Log in to join the discussion
Log InNo comments yet