
LanceDB
LanceDB, the San Francisco startup behind the open-source Lance data format, announced a $57.5 million Series B on Nov. 17 to extend its AI training-data platform from search into data preparation and training-run data loading. Teams choosing where to store model-training data should weigh one fact before the funding headline: the project's marquee claim of 100x faster random access than Apache Parquet appears in LanceDB's own repository, backed by a test the repository describes as preliminary.
Bessemer Venture Partners and Innovation Endeavors co-led the round, with Theory Ventures, Dimension, Swift Ventures, CRV and Y Combinator participating, and LanceDB puts its total funding at $98.5 million. The raise arrives about 17 months after a $30 million Series A that Theory Ventures led in June 2025. None of the announcement coverage states a valuation. Bessemer describes the five non-lead participants as continuing backers.
LanceDB sells what it calls a multimodal lakehouse, a platform that holds text, images, video, audio and sensor data in one system so AI teams can search it, curate it and train models on it. The foundation is Lance, an open-source columnar format. LanceDB, the database on top, adds vector, full-text and hybrid search, and LanceDB Enterprise, the paid tier, runs in a customer's own cloud or on-premises. The platform connects to cloud storage a customer already uses and leaves the data in place.
The company says the capital will support pre-training and post-training workflows for frontier models, world models and robotics action models. Planned work includes data preparation before training, better data loading and caching during training runs, and feature engineering, where engineers write a transformation once, apply it across a dataset, backfill it onto existing rows and run it on incoming data. LanceDB also describes a single system that consolidates training data, context, logs and rollouts and lets multiple AI agents run experiments without contaminating shared datasets. Those are roadmap items, not shipped products.
New Enterprise features announced with the round include a durable, memory-accelerated write path for real-time ingestion, distributed indexing across very large datasets and managed background compaction. LanceDB says compaction supports up to 100 billion multimodal rows while multiple researchers read, write and query the same table. CEO Chang She frames the opportunity as "a missing data layer for AI".
Columnar formats store each column separately, so pulling one complete row for a training shuffle can require many small disk reads, large over-reads or heavy decoding, according to the Lance paper. Parquet needs roughly one read per row because all of a row's buffers in a column sit in one compressed page, yet expensive compression can keep the processor busy while the disk idles, and Parquet's in-memory page index grows too large for big values such as vectors, at about 20 bytes per page.
Lance picks between two layouts by value size, according to a paper by LanceDB's own engineers. For values of 128 bytes or more, such as embeddings, "full zip" stores each row's data together in one buffer and adds a repetition index that points to each row, so a variable-width lookup takes two reads. For small values, "miniblock" splits data into chunks sized at one to two disk sectors, 4 KiB to 8 KiB of compressed data, and a lookup reads and decodes the whole chunk. The design goal is at most one read for fixed-width values and two for variable-width values.
The design carries costs the authors list themselves. Miniblock lookups decode neighboring values along with the target. Full zip restricts the compression schemes it can use and wastes space on nulls: a null 4 KiB embedding still occupies full size. Packing struct fields together roughly doubles whole-record random-read speed at five fields but gives up cheap reads of a single field. The tests ran on a local NVMe drive; on Amazon S3, reads smaller than about 100 KB do not pay off, so gains depend on where training data lives. The paper tested Lance format versions 2.0 and 2.1, and LanceDB described version 2.2 in a March 9, 2026 post. For engineers, the consequence is concrete: random-read gains depend on value size, storage tier and the Lance version a team runs.
The Lance repository states the 100x figure in three places, including "100x faster than Parquet or Iceberg for random access". The benchmark behind it uses the Oxford Pet dataset, which the page calls preliminary performance testing and describes without hardware or method.
Other published numbers differ. Gradient Flow analyst Ben Lorica cites benchmarks showing 3x to 35x faster random reads, without naming their source, and up to 10x faster vector queries than Parquet. Rival format Vortex publishes a benchmark page listing Lance at 2.13 milliseconds per random access against 68.48 for Parquet, a gap of roughly 32x by TechTimes' arithmetic, and Vortex itself at 0.75; the page states no hardware or dataset.
The Lance authors' own paper supplies a caution. With default settings, the Parquet Rust reader managed 5,500 rows per second on small scalar lookups; tuned, it reached 350,000, a spread of more than 60x from configuration alone. The speed benchmarks traceable to a named source come from LanceDB's engineers or from Vortex's project. Dremio's Alex Merced, who works at a lakehouse vendor, describes Lance's advantage as claimed, says results vary by workload and advises teams to measure their own data. Zilliz, whose Vector Lakebase builds on Vortex, publishes a Lance-versus-Vortex comparison that gives no Lance numbers, though a separate Zilliz test reports Lance cutting per-read S3 traffic about 1,500x versus Parquet, a traffic measure rather than a speed measure.
LanceDB says revenue has grown more than 10x year over year since the Series A and that its open-source library has passed 120 million downloads, about six times the total a year earlier. It says a single table can hold hundreds of petabytes or more than 40 billion vectors, serve more than 130,000 requests per second at a 95th-percentile latency under 210 milliseconds, and support training across 8,000 GPUs. One account, Superpower Daily, frames the thousands-of-GPUs figure as a training target rather than a measured result.
LanceDB has not published absolute revenue, customer counts or headcount, so 10x growth has no base to measure against. Download totals measure adoption of the free library, not paying contracts.
That gap is the standard test for open-core companies, which release a feature-limited core free and sell proprietary tiers. MongoDB, Redis, HashiCorp and Elastic each changed licenses, and community forks followed in cases such as Valkey for Redis and OpenTofu for HashiCorp's Terraform. Lance's repository lists the Apache 2.0 license, and nothing in the announcement coverage indicates a change. Three signals will show how far conversion has gone: whether Enterprise adoption broadens beyond the media and generative-AI labs that anchor the customer list, how quickly the feature-engineering and data-loading work ships, and whether LanceDB discloses harder financial figures.
LanceDB names Netflix, Microsoft AI, Runway, Luma AI and Midjourney as customers, and every endorsement reaches readers through the company's own announcement. Runway began using Lance in 2024 and later became an Enterprise customer. Its CTO, Kamil Sindi, says "A training example at Runway doesn't stay still," describing examples that grow from a video and a prompt into rows carrying embeddings and metadata.
Microsoft AI's head of data platform, Suqiang (Jack) Song, says the team relies on LanceDB as its training data layer across billions of rows of image and code data, with distributed inference running inside Microsoft's own tenant. NVIDIA's Cosmos team, an open-source user and partner, curated roughly one million hours of video with Lance for Cosmos 3. Its technical lead, Max Li, calls Lance a "critical part of the infrastructure" for scaling Cosmos.
Netflix's Media ML Data Engineering team has said, in a statement LanceDB reproduces, that it is building a Media Data Lake for media assets on LanceDB. Lorica goes further, writing that Netflix migrated from Parquet to unify video frames and metadata, and he interviewed Chang She for O'Reilly in May 2026. Midjourney, another named customer, faces copyright suits from Disney, Universal and Warner Bros. over AI training, which Midjourney contests as fair use; the claims concern training and outputs rather than data storage.
Read more: Embodied AI World Models Attracted $6 Billion, But the LLM Parallel May Not Hold
Lance is one of several formats contesting Parquet's place in AI pipelines. Vortex, donated by Spiral and now a Linux Foundation project, targets compression and scan speed under multi-vendor governance. Meta open-sourced Nimble from its ML platform for extremely wide feature tables, and Meta reports decoding two to three times faster than earlier columnar formats. Parquet is also moving: Merced lists a variant type, native geospatial types and format version 2.13 as shipped, with a footer redesign under way and a fixed-size list type for embeddings queued as an AI-era addition.
Governance is the second axis. Zilliz, which sells the Milvus database and builds Vector Lakebase on Vortex, notes that Lance is anchored to a single company and had not committed to a foundation home as of late 2025, which can matter in procurement. Zilliz sells a vector database and Merced works at lakehouse vendor Dremio, so both comparisons come from companies with products in adjacent markets.
The sources agree on where Lance wins: random row lookups, built-in versioning and an integrated vector index. Merced says Lance does not claim Parquet's crown on pure batch scans.
Iceberg, Delta Lake and Hudi sit on top of Parquet files and supply the transactions, schema evolution and time-travel queries a lakehouse needs, while Parquet remains the immutable layer underneath. Lance bundles a file format, a table format and a lightweight catalog spec, which Gradient Flow says covers all three layers of a conventional stack. Merced writes that the overlap creates an either-or decision for enterprises with Iceberg estates.
LanceDB frames the relationship differently, publishing a reference stack that pairs Lance with Iceberg. The choice matters beyond file speed: Gradient Flow says Lance brings its own snapshot versioning and time-travel queries, so a team adopting its table layer takes on a different versioning model than Iceberg tooling provides.
Chang She was an early contributor to the pandas data library, which he describes as co-authoring, before co-founding DataPad, which Cloudera acquired in 2014, and later serving as VP of engineering at streaming company Tubi. Co-founder and CTO Lei Xu is a member of the Apache Hadoop HDFS project management committee, worked on HDFS at Cloudera and led machine-learning infrastructure at Cruise. Bessemer partner Mike Droesch says the pair have spent nearly two decades building core infrastructure behind many of the world's data platforms. Both founders co-authored the Lance paper. Innovation Endeavors, the other lead investor, was co-founded by former Google CEO Eric Schmidt.
The Lance paper's own numbers make the practical test straightforward. A team evaluating Lance should first tune Parquet, because configuration alone moved Parquet by more than 60x in the authors' tests. It should then measure both formats on its own access pattern and storage tier, since gains differ for small scalars, large embeddings, local NVMe and S3.
For engineers, the Series B settles little about speed and a lot about durability. LanceDB now has $98.5 million in total funding to push a format whose headline multiplier is self-reported, whose benchmarks come from its authors and a rival, whose stewardship sits with one company and whose overlap with Iceberg forces a stack decision. The random-lookup design is well documented, and customers such as Runway and Microsoft AI describe production use, but until an independent benchmark exists, treat 100x as a vendor ceiling, not a planning number.
Not necessarily. Lance bundles a file format, table format and catalog spec, so it can overlap Iceberg's table layer, and Dremio's Alex Merced says that overlap creates an either-or choice for enterprises with Iceberg estates. LanceDB itself publishes a stack that pairs Lance with Iceberg.
A lakehouse combines a data lake's raw object storage with a warehouse's transactions and enforced data quality. LanceDB's version stores text, images, video, audio and sensor data in one system so teams can search, curate and train from it.
The Lance repository lists the Apache 2.0 license, and LanceDB's JavaScript library carries the same license in Snyk's package listing. LanceDB Enterprise adds a durable, memory-accelerated write path, distributed indexing and managed background compaction, and it can run in a customer's own cloud or on-premises.
The sources point to testing before moving. The Lance paper reports that tuned Parquet runs more than 60 times faster than default Parquet on random lookups, and Merced's closing advice is to measure on your data. Lance suits random row lookups, versioning and built-in vector indexes, while Merced says it does not claim Parquet's crown on pure batch scans.
