跳到正文
实时
Tomasz Tunguz·· 20 小时前

Lance 数据湖:AI 基础设施中的核心组件

The Lakehouse for AI

要点速览

Lance 数据湖让企业高效处理 AI 训练与推理所需的大规模文件,已在 Exa、Runway、Microsoft AI、Adobe 和 NVIDIA 等公司实现生产部署,支持 35PB 数据和千万级节点推理。LanceDB 企业版整合了非结构化数据处理、列式追加、分布式读取和 SQL 支持,年收入年增 10 倍,获 5700 万美元 B 轮融资。

正文

Read on tomtunguz.com →

Across the largest AI labs & AI infrastructure companies, there is a common component : a lakehouse that stores data for training & inference called Lance.

Lance allows companies to manipulate massive collections of files in the most efficient way possible for AI.

The LanceDB Multimodal Lakehouse unifying curation, feature engineering, training & search

Five production deployments show how that architecture runs at frontier scale :

Exa stores 35pb in Lance tables across 100b semantic web pages with crawl-to-search latency under five minutes.1 Index lead Jan van der Vegt explained :

“These column backfills allow us to keep the old data in place & only write what is new.”

Runway runs Lance & LanceDB Enterprise across its video-generation stack. CTO Kamil Sindi noted :2

“We first started using the Lance format in 2024 because it was the only option that handled unstructured data, column appends, distributed reads & ad-hoc SQL all at once.”

Microsoft AI runs LanceDB inside its own tenant & security boundary. Head of Data Platform Suqiang “Jack” Song shared :3

“We rely on LanceDB as the data layer for training across billions of rows of image & code data, with distributed inference reading the format directly across thousands of nodes.”

Adobe manages 40pb to 50pb across 800k tables & 220b daily events, contributing 20+ pull requests upstream.4 As Adobe’s platform team put it :

“Where should the AI-native data live? Our answer is in the same lakehouse, under the same contract as everything else.”

NVIDIA closed the loop on physical AI. Cosmos tech lead Max Li explained :5

“First ever, we train on thousands & tens of thousands of GPUs at scale with Lance tables directly.”

Since its Series A last June, LanceDB has grown revenue more than 10x year over year across Netflix, Microsoft AI, Runway, Luma AI & Midjourney, while open-source downloads grew 6x past 120m.6

Theory partnered with Chang She & Lei Xu at Seed & led their $30m Series A.7 Today we are thrilled to join Bessemer Venture Partners & Innovation Endeavors, alongside Dimension, Swift Ventures, CRV & Y Combinator, in LanceDB’s $57.5m Series B.

Learn more about how to use Lance here.


  1. Jan van der Vegt (Exa), Managing an Index of 100s of Billions of Pages with Exa Skeleton, Reverie Summit, November 5, 2026. ↩︎

  2. Ethan Rosenthal (Runway), Scalable, Multimodal Data Curation, Reverie Summit, November 5, 2026, & Kamil Sindi (CTO, Runway) in LanceDB’s Series B announcement. ↩︎

  3. Suqiang “Jack” Song (Head of Data Platform, Microsoft AI), Building the Data Flywheel for Frontier AI Model, Reverie Summit, November 5, 2026, & LanceDB Series B announcement. ↩︎

  4. Jing Chen He & Miao Wang (Adobe), How Adobe Built a Multimodal Lakehouse for AI-Native Experiences, Reverie Summit, November 5, 2026. ↩︎

  5. Max Li (Technical Lead, NVIDIA Cosmos), Cosmos : Omnimodal World Models for Physical AI, Reverie Summit, November 5, 2026, & LanceDB Series B announcement. ↩︎

  6. Chang She, “LanceDB Raises $57.5m as AI Research Outgrows the Modern Data Stack,” November 2026, & Yichen Wang (Luma AI), From Raw Delivery to the Training Loop : Luma’s Data Platform on Lance, Reverie Summit, November 5, 2026. ↩︎

  7. Tomasz Tunguz, A Data Foundation for Multimodal AI : Our Series A Investment in LanceDB, June 24, 2025. ↩︎

来源: Tomasz Tunguz · tomtunguz.com