Startups & Funding

LanceDB Raises $57.5M Series B to Expand Its Multimodal AI Lakehouse

LanceDB raised $57.5 million on November 17, 2026, reporting more than tenfold revenue growth, over 120 million downloads and a platform for petabyte-scale training data.

LanceDB Raises $57.5M Series B to Expand Its Multimodal AI Lakehouse
LanceDB Raises $57.5M Series B to Expand Its Multimodal AI Lakehouseblickpixel / pixabay
By Marcus Bennett6 min read

Why it matters

  • $57.5 million Series B announced November 17, 2026; total funding reached $98.5 million.
  • Revenue grew more than tenfold year over year since June 2025; downloads surpassed 120 million after sixfold growth.
  • LanceDB reports 130,000+ requests per second, latency below 210 milliseconds for 95% of requests, and training across 8,000 GPUs.
  • Tables can hold hundreds of petabytes or more than 40 billion vectors; managed compaction supports up to 100 billion multimodal rows.
  • NVIDIA Cosmos curated roughly one million hours of video with Lance for Cosmos 3 and is moving toward direct training.

LanceDB raised $57.5 million in a Series B announced on November 17, 2026, bringing the company’s total funding to $98.5 million.

Bessemer Venture Partners and Innovation Endeavors led the round; Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator also participated.

LanceDB plans to put the capital toward a broader role across model development, including training-data management, feature engineering, experiment execution and data loading.

The investment targets a bottleneck LanceDB identifies in frontier AI: researchers can lose weeks or months obtaining and curating training data even though frontier AI advances are measured in days and weeks.

LanceDB says manual pipeline oversight and data-loading problems can slow research. The Series B gives the company capital to address both.

LanceDB wants one platform for the work surrounding model training, rather than a tool limited to retrieval. Its plan spans datasets that change, experiments that run and training jobs that begin.

What does LanceDB’s multimodal lakehouse do?

LanceDB describes a multimodal lakehouse—a system for organizing and delivering training material—as the core of its platform. LanceDB says it handles videos, images and sensor data. Retrieval, in this use, means finding and returning relevant stored records.

  • Model-development scope: LanceDB says the platform supports frontier models, world models (models of simulated environments), physical AI (AI systems operating in the physical world) and enterprise applications.
  • Data preparation: LanceDB says curation, feature engineering (preparing model inputs) and experiment execution sit alongside training and data loading.
  • Search and retrieval: LanceDB says large-scale retrieval serves changing datasets used in training and other model-development work.

LanceDB says the goal is to unify training data, experiments and rollouts (deployments of trained models), reducing the need for teams to hand-manage separate pipeline stages.

In LanceDB’s design, one changing dataset can support curation, transformation, search and training without teams handing work between separate pipeline stages.

How large can LanceDB’s platform operate?

LanceDB reports the following operating and capacity figures:

  • Query throughput: LanceDB says the platform supports more than 130,000 requests per second.
  • Latency: LanceDB says 95% of requests complete in less than 210 milliseconds.
  • Table capacity: LanceDB says an individual table can hold hundreds of petabytes of data or more than 40 billion vectors (numerical representations indexed for search).
  • Training scale: LanceDB says the platform supports training across 8,000 GPUs (graphics processors) in parallel.

LanceDB’s figures address three constraints: response time for stored-data queries, individual table capacity and parallel model training. These are LanceDB’s own figures.

The company also reports more than 120 million downloads of its open-source library, a sixfold increase in total downloads from the same point one year earlier.

Since its Series A in June 2025, LanceDB reports more than tenfold year-over-year revenue growth.

LanceDB names Netflix, Microsoft AI, Runway, Luma AI and Midjourney as customers. It also identifies NVIDIA Cosmos, an open-source project, as both a user and partner.

The Series B gives LanceDB capital to expand a platform with reported download and revenue momentum. Its customer roster covers enterprise AI, generative-media companies and an open-source model project.

How can shared tables change during use?

LanceDB’s Enterprise additions focus on a specific problem: training tables can change while teams read from, write to and query them.

  • Live ingestion: LanceDB says a durable, memory-accelerated write path (the route for live data into a table) handles real-time ingest.
  • Distributed indexing: LanceDB says its distributed indexing scales across large datasets.
  • Managed compaction: LanceDB says managed compaction (continuous background table optimization) supports up to 100 billion multimodal rows while multiple researchers concurrently read, write and query the same table.

LanceDB positions the combined features as a way to keep shared tables usable during live ingestion, indexing and continuous optimization. That matters because the customer workflows behind the round repeatedly add or revise data while research continues.

How are customers changing training datasets?

LanceDB’s customer examples focus on data that changes as model development proceeds.

Runway CTO Kamil Sindi said: “A training example at Runway doesn’t stay still. It starts as a video and a prompt, then we add embeddings, resolution and duration metadata, whatever the next experiment needs.”

An embedding is a numeric representation of content. Sindi’s account describes how each experiment can add another layer of metadata to the same underlying training example.

LanceDB says Runway began using the Lance data format in 2024 and later became a LanceDB Enterprise customer. The company now uses Lance and LanceDB Enterprise together so researchers can move through large, changing datasets more quickly.

LanceDB identifies NVIDIA Cosmos, an open-source project, as both a user and partner. Technical lead Max Li said the team curated roughly one million hours of video with Lance for Cosmos 3.

Li said the team is moving toward training directly from the Lance format. He also said its object-storage-native architecture, built around object storage, helps the team stream, filter and inspect large datasets as those datasets are revised while supplying data to large GPU clusters.

Microsoft AI data-platform head Suqiang (Jack) Song said his organization uses LanceDB as its training-data layer for billions of rows spanning images and code.

Song said distributed inference jobs (model-prediction workloads) read Lance directly on thousands of nodes (compute machines). He said operating within Microsoft’s own tenant (an isolated cloud environment) and security policy was non-negotiable for production.

What will the Series B fund next?

LanceDB intends to bring training data, context (information supplied to a model), logs (records of model runs) and rollouts into one system at petabyte scale. The objective covers the full model-improvement cycle rather than a single stage before training.

Research teams would also be able to run large numbers of AI-agent experiments in parallel without contaminating shared datasets. Isolation would keep one experiment from changing data used by another.

LanceDB also plans to deepen its feature-engineering tools. A team would define a transformation once; the platform would then apply it throughout the dataset, backfill existing rows (apply the transformation to stored data) and process new data as it arrives.

Further work targets more preparation before training, plus better data loading and caching (reusing stored data to speed access) during training runs.

The Series B pushes LanceDB beyond retrieval into the continuing work of changing datasets, conducting experiments and supplying training jobs. Its stated goal is to reduce manual pipeline oversight and data-loading problems across the full model-improvement cycle.

LanceDB’s next phase will test whether one platform can accommodate both frontier-scale data volumes and constant, experiment-by-experiment changes described by Runway. That test matters because LanceDB’s Series B thesis depends on one system handling high-volume data and frequent revisions, rather than merely answering searches quickly.

Source: unite.ai

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

718 articles

Related articles

  1. Humyn Labs raises $47M for real-world robot training data
  2. Modal Labs Nears $750M Round at $15.75 Billion Valuation
  3. Sequel.io Raises $15 Million Series B to Fund AI Agents and Real-Time Website Personalization
  4. Dreamteam Raises $53 Million for AI-Native CRM With Four AI Coworkers
  5. Trajectory raises $45M to automate AI model post-training

« Previous articleNext article »

LanceDB Raises $57.5M to Expand Multimodal AI Lakehouse — AI In Context