Today, we’re pleased to announce that we’ve raised $57.5M in our Series B funding to bring the LanceDB multimodal lakehouse to the researchers, model builders, and platform teams working at the frontier.
Our Series B is led by Bessemer Venture Partners and Innovation Endeavors with participation from Theory Ventures, Dimension, Swift Ventures, CRV, and Y Combinator, and brings our total funding to date to $98.5M. Our customers include Netflix, Microsoft AI, Runway, Luma AI, and Midjourney.
"Chang and Lei have spent nearly two decades building core infrastructure that powers many of the world's data platforms, and with LanceDB they're leading one of the most important infrastructure shifts in AI: the move to training models on multimodal data at massive scale," said Mike Droesch, Partner at Bessemer Venture Partners. "The company's growth speaks for itself, from a widely-adopted open source technology to a fast-growing enterprise platform trusted by the industry's most sophisticated AI labs."
AI research needs a new data stack
When advances on the frontier of AI are measured in days and weeks, researchers are losing weeks and months obtaining and curating the best data for their training runs. LanceDB gives model builders, whether they’re working at the frontier, on world models, physical AI, or within an enterprise, a multimodal data platform for curation, search and retrieval, and feature engineering for the most demanding scale.
Today, LanceDB can support the following:
- 130K+ requests per second, under 210ms P95 latency
- Hundreds of petabytes of data in a single table
- Over 40 billion vectors in a single table
- 8K GPUs training in parallel
This lets researchers focus on improving the next model faster, rather than preparing and finessing data infrastructure.
LanceDB becomes the new standard
Since our Series A last June, revenue has grown by more than 10x year over year. LanceDB has also continued to grow open source traction, with the open source library having been downloaded over 120 million times, a 6x growth in total downloads compared to this time last year.
We’ve also rolled out some notable new features: LanceDB Enterprise now provides:
- A durable, memory-accelerated write path for real-time ingest,
- Distributed indexing that scales across massive datasets, and
- Managed compaction that continuously optimizes tables in the background, enabling teams to work with up to 100 billion multimodal rows while multiple researchers concurrently read, write, and query the same table.
All while removing operational complexity for our customers.
What customers and users say about LanceDB
“A training example at Runway doesn't stay still,” said Kamil Sindi, CTO, Runway. “It starts as a video and a prompt, then we add embeddings, resolution and duration metadata, whatever the next experiment needs. We first started using the Lance format in 2024 because it was the only option that handled unstructured data, column appends, distributed reads, and ad-hoc SQL all at once. We’ve since become a LanceDB Enterprise customer. We use both Lance and LanceDB Enterprise to provide our researchers a faster way to explore that data at scale.”
“Physical AI is entering a new scaling regime,” said Max Li, Technical Lead for NVIDIA Cosmos, a LanceDB open source user and partner. "With Cosmos 3, we curated roughly a million hours of video using Lance; now we’re taking the next step and training directly from Lance. Its object-storage-native architecture makes it far easier to stream, filter, inspect, and continuously improve massive multimodal datasets while keeping large GPU clusters fed. Lance is becoming a critical part of the infrastructure that allows us to scale Cosmos.”
"We rely on LanceDB as the data layer for training across billions of rows of image and code data, with distributed inference reading the format directly across thousands of nodes,” said Suqiang (Jack) Song, Member of Technical Staff, Head of Data Platform, Microsoft AI. “Doing that in our own tenant, under our own security policy, was non-negotiable, and LanceDB gave us what we needed for full confidence in production."
The next phase of development for the multimodal lakehouse
When we raised our Series A last year, customers already relied on LanceDB to do retrieval over billions of rows of multimodal data including images, videos, and sensor data to create AI models and applications.
Since then, we've continued to build out workflows around multimodal data, especially for pre- and post-training AI models such as frontier models, world models, and robotics action models. LanceDB lets model builders store, curate, enrich, and train on their multimodal data in one system, rather than being forced to go on a scavenger hunt to find the most impactful training examples.
With the Series B, we’re focused on supporting AI research teams across the full end-to-end flywheel for model improvement. Researchers and their agents can manage petabytes of training data, context, logs, and rollouts in one place, and run massively parallel agentic experiments without polluting shared datasets. And they can do so without babysitting pipelines, struggling to feed GPUs, or losing days and weeks fighting their data.
We’re also investing in our feature engineering capabilities: teams write a transformation once, and LanceDB runs it across the dataset, backfills existing rows, and processes new data as it arrives. Over time, we’ll extend the platform into more of the data preparation that happens before training, and improve data loading and caching for the training runs themselves.
We're working with customers at the frontier of AI building models and systems that weren't possible a year ago. Cutting-edge teams should be able to increase research velocity and have the only limit be their ideas, not their infrastructure.
How you can get involved
Replacing the data layer beneath AI research is a long project and we're still in the early innings. If you're wrangling petabytes of video, sensor logs, or images into shape for a training run, LanceDB can help. Watch the recorded sessions from Reverie, the Summit for AI Builders hosted by LanceDB, to hear how teams at NVIDIA, Runway, Netflix, and Luma AI are building on multimodal data at scale.
If you want to get involved with LanceDB directly, consider contributing to the Lance project or connecting on Discord. We're also hiring across engineering, field engineering, and go-to-market.





