LanceDB's Multimodal Lakehouse Lands Midjourney and $98.5 Million

The open-source vector database, built by a pandas co-creator, is expanding into a unified system for managing the massive datasets that train AI models.

About LanceDB

Published

For an AI company, the most expensive part of building a model isn't the GPU time. It's the data. Managing petabytes of images, video, and text for training, then billions of vector embeddings for retrieval, creates a sprawling infrastructure problem that most teams solve with a patchwork of specialized databases and object storage. LanceDB, an open-source project founded in 2022, started by offering a simpler answer for one piece: a serverless vector database. Its real bet, now backed by $98.5 million in venture funding, is that teams need a single system for the entire data lifecycle, from curation to training.

Chang She, LanceDB's CEO and a co-creator of the ubiquitous pandas Python library, frames the problem in practical terms. "You have this massive corpus of multimodal data. You need to search it, you need to version it, you need to extract features, and you need to feed it efficiently into training pipelines," he told TechCrunch in 2024. "Splitting that across five different systems creates massive engineering overhead and data silos." LanceDB's answer is what it calls a multimodal lakehouse, built on its own Lance columnar file format, designed to handle both the search and the SQL-based exploration required for model development [TechCrunch, May 2024].

From Vector Search to Training Data

The company's wedge was its open-source, Apache 2.0-licensed vector database, which gained traction for its serverless architecture and integration with cloud object storage like S3. By June 2025, the company reported more than 20 million downloads of its open-source packages over the preceding year [AI Engineer, June 2025]. This community adoption provided the top-of-funnel lead generation for its commercial product, LanceDB Enterprise.

The enterprise layer is where the lakehouse concept takes shape. It adds capabilities for managing the scale of AI training data: versioning datasets, tracking data lineage, executing feature engineering jobs, and providing a unified SQL interface over both structured metadata and unstructured content. The company claims its system supports workloads involving billions of vectors and petabytes of training data, a scale evidenced by its flagship customers [LanceDB, June 2025].

The Team Behind the Format

The technical credibility of LanceDB is anchored in its founders' deep data infrastructure backgrounds. Chang She is not just a pandas co-creator; he previously served as VP of Engineering at TubiTV, where he built its machine learning stack, and was a co-founder of DataPad, which was acquired by Cloudera. Co-founder and CTO Lei Xu is described as an Apache Hadoop committer who led machine learning infrastructure at the autonomous vehicle company Cruise [TechCrunch, May 2024]. This combination of open-source library creation and large-scale, production ML platform experience is a precise fit for the problem LanceDB is solving.

Seed (May 2024) | 8 | M USD
Series A (Jun 2025) | 30 | M USD
Series B (Nov 2026) | 57.5 | M USD

Traction at the Frontier of AI

LanceDB's early customer list reads like a who's who of generative AI pioneers. The company has named Midjourney, Character.ai, and Runway as enterprise users, with Airtable and WeRide also cited as customers [TechCrunch, May 2024]. These are organizations dealing with some of the largest, most complex multimodal datasets in the world. For a model like Midjourney, searching across billions of image embeddings to inform generation is a core production workload. LanceDB's positioning is that it can handle that retrieval latency while also being the system that stores and manages the original training images.

This traction has attracted a tier of investors focused on foundational infrastructure. The company's $8 million seed round in May 2024 was led by CRV, Essence VC, and Swift Ventures. A $30 million Series A followed in June 2025, led by Theory Ventures with participation from Databricks Ventures and RunwayML. Most recently, in November 2026, Bessemer Venture Partners and Innovation Endeavors co-led a $57.5 million Series B, bringing total disclosed funding to $98.5 million [LanceDB, November 2026].

The Competitive Landscape

LanceDB does not operate in a vacuum. Its evolution from vector database to lakehouse places it in a competitive matrix with several established players.

Competitor Primary Focus LanceDB's Positioning
Turbopuffer High-performance vector search Broader data management for training, not just search.
Delta Lake / Iceberg Traditional data lakehouse formats Native multimodal support, vector indexing, and AI-centric tooling.
Proprietary Vector DBs (e.g., Pinecone) Managed vector search service Open-core model, deeper integration with training pipelines, and object storage flexibility.

The company's argument is that existing data lakehouse formats like Delta and Iceberg were built for tabular analytics, not for the unique demands of AI data. Meanwhile, pure vector databases lack the tools for the full data curation and feature engineering lifecycle. LanceDB's bet is that the convergence of these workflows creates a new category.

Technical Breakdown and Scale Risks

The core of LanceDB's architecture is the Lance columnar format. It's designed for high-performance random access, which is critical for both vector search and sampling batches for training. The format allows for efficient versioning and delta updates, which is a non-trivial problem when dealing with petabyte-scale datasets that evolve daily. The enterprise layer then adds the necessary controls for production: SOC 2 Type II, HIPAA, and GDPR compliance, along with encryption for data at rest [LanceDB].

The sober assessment for any system claiming petabyte-scale management lies in operational complexity at the limits. While the technology can theoretically handle the load, the practical challenges emerge in areas like:

  • Global data synchronization: Keeping vector indexes instantly consistent across regions for a globally distributed application like Character.ai is a distributed systems challenge of the highest order.
  • Query performance isolation: Ensuring a long-running training data export job doesn't degrade latency for a real-time inference search requires sophisticated resource governance.
  • Vendor integration fatigue: As the lakehouse expands, it must integrate with an ever-growing stack of cloud services, model training frameworks, and monitoring tools. Maintaining these integrations becomes a tax on the engineering roadmap.

LanceDB's next twelve months will be about proving it can manage these scale risks not just for a handful of elite AI labs, but for the broader enterprise market. The recent Series B capital is likely earmarked for scaling its sales and customer engineering teams to tackle that expansion. The open-source community, now reportedly downloading millions of packages per month, provides a strong foundation, but the leap from embedded library to enterprise data platform is a formidable one. If the team can navigate that transition, they won't just be another vector database. They'll have defined the data layer for the next generation of AI models.

Sources

  1. [TechCrunch, May 2024] LanceDB, which counts Midjourney as a customer, is building databases for multimodal AI | https://techcrunch.com/2024/05/15/lancedb-which-counts-midjourney-as-a-customer-is-building-databases-for-multimodal-ai
  2. [AI Engineer, June 2025] LanceDB Raises $30M Series A to Build the Multimodal Lakehouse | https://www.lancedb.com/blog/lancedb-raises-30m-series-a-to-build-the-multimodal-lakehouse
  3. [LanceDB, November 2026] LanceDB Series B Funding | https://www.lancedb.com/blog/lancedb-series-b-funding
  4. [Y Combinator] LanceDB: Open-source, serverless vectordb for production-scale generative AI | https://www.ycombinator.com/companies/lancedb
  5. [Theory Ventures, June 2025] A Data Foundation for Multimodal AI: Our Series A Investment in LanceDB | https://theoryvc.com/blog-posts/a-data-foundation-for-multimodal-ai-our-series-a-investment-in-lancedb

Read on Startuply.vc