Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarLanceDB Blog→ original

LanceDB Raises $57.5M for a Database That Trains Models on Video and Images

Training models on video, images and robot logs is awkward: the data sits in different places, and researchers spend weeks collecting and cleaning it. The startup LanceDB is building a single "multimodal" database where everything is stored, searched and prepared for training. The Series B round totaled $57.5M, bringing the total raised to $98.5M. If you are building a training pipeline for unstructured data, it is worth taking a look at the open Lance format.

AI-processed from LanceDB Blog; edited by Hamidun News
LanceDB Raises $57.5M for a Database That Trains Models on Video and Images
Source: LanceDB Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceLanceDB Blog ↗
Editor-in-Chief
Key metric #1The round was led by Bessemer Venture Partners and Innovation Endeavors
Key metric #2Revenue grew more than 10x in a year, and the library has over 120 million downloads
Key metric #3A single table is claimed to hold more than 40 billion vectors and to support up to 8,000 GPUs in parallel

When progress in AI is measured in weeks, researchers cannot afford to spend months searching for and preparing training data. This is especially painful with video, images, sensor recordings and logs, which are scattered across different storage systems and formats. The startup LanceDB raised $57.5M in a Series B round to solve this problem with a single platform.

Who Provided the Money

The round was led by Bessemer Venture Partners and Innovation Endeavors. Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator also participated. In total, the company has raised $98.5M. Its customers reportedly include Netflix, Microsoft AI, Runway, Luma AI and Midjourney.

The company says revenue has grown more than 10x since the Series A round, roughly a year ago. The open-source library has been downloaded over 120 million times, six times more than a year ago. These figures come from the company itself and have not been independently verified.

How It Works in Practice

Take a training example at Runway. It starts as a video and a text prompt. Then embeddings, resolution, duration and everything the next experiment will need get added to it. In a typical data lake, such additions are inconvenient: you have to rewrite huge files or set up separate tables. According to Runway CTO Camill Sindi, the company chose the Lance format in 2024 because it could handle unstructured data, add columns, read data in a distributed way and run SQL queries all at once.

LanceDB stores, searches, enriches and serves training data within a single system. NVIDIA used Lance to select about a million hours of video for Cosmos 3 and is now moving to training directly from that format. Microsoft AI uses LanceDB as a data layer for training on billions of rows of images and code, in its own cloud environment and under its own security rules.

The platform's claimed capabilities:

  • more than 130 thousand queries per second with 95th-percentile latency below 210 ms;
  • hundreds of petabytes in a single table;
  • over 40 billion vectors in a single table;
  • up to 8 thousand GPUs training in parallel.

What's New in the Enterprise Version

The company has added three things. The first is a reliable, fast write path for data arriving in real time. The second is distributed index building over very large datasets. The third is managed table compaction, where the system tidies up storage in the background. This makes it possible to work with tables of up to 100 billion multimodal rows while several researchers read, write and query the same table at once.

The next step is feature engineering. The team writes a transformation once, and LanceDB applies it to the entire dataset, backfills older rows and processes new ones as they arrive. Going forward, the company wants to take on more of the data preparation before training and improve loading and caching during the training runs themselves. It also describes a scenario where researchers and their agents run many parallel experiments without corrupting shared datasets.

What This Means for Practitioners

The market is shifting from text models toward video, world models and physical AI, and the bottleneck is no longer training code but data. If your team works with petabytes of video or robot logs, it is worth looking at the open Lance format and comparing it with conventional data lakes built on Parquet and Iceberg.

There are caveats. Most of the numbers come from the company's own blog, and the customer quotes were chosen for the press release. You should verify the claimed scale on your own data. The platform is also designed for very large workloads, and a small team with a couple of terabytes will likely be better served by simpler tools.

Story

Key Questions Answered

Who Provided the Money?

The round was led by Bessemer Venture Partners and Innovation Endeavors. Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator also participated. In total, the company has raised $98.5M. Its customers reportedly include Netflix, Microsoft AI, Runway, Luma AI and Midjourney. The company says revenue has grown more than 10x since the Series A round, roughly a year ago. The open-source library has been downloaded over 120 million times, six times more than a year ago.

How It Works in Practice?

Take a training example at Runway. It starts as a video and a text prompt. Then embeddings, resolution, duration and everything the next experiment will need get added to it. In a typical data lake, such additions are inconvenient: you have to rewrite huge files or set up separate tables. According to Runway CTO Camill Sindi, the company chose the Lance format in 2024 because it could handle unstructured data, add columns, read data in a distributed way and run SQL queries all at once.

What's New in the Enterprise Version?

The company has added three things. The first is a reliable, fast write path for data arriving in real time. The second is distributed index building over very large datasets. The third is managed table compaction, where the system tidies up storage in the background. This makes it possible to work with tables of up to 100 billion multimodal rows while several researchers read, write and query the same table at once. The next step is feature engineering.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…