
When we led LanceDB’s Series A last year, we imagined a world where every AI application was multimodal at its core, and LanceDB built the infrastructure that made it possible.
This month LanceDB held its first conference, Reverie. The name means daydream, but the talks were all about production. Teams from NVIDIA, Adobe, Applied Intuition, Luma, Runway, and Exa showed how they train and serve frontier models at petabyte scale, building on Lance and LanceDB.
Our vision is now becoming a reality, and we’re excited to double down and join LanceDB’s $57.5M Series B, led by Bessemer Venture Partners and Innovation Endeavors, with participation from Dimension and existing investors.
Physical AI startups raised $47 billion in the first half of 2026, more than in 2022 through 2024 combined. That buys fleets, sensors, and GPUs, which produce and consume multimodal data at a scale today’s infrastructure wasn’t built for.
Applied Intuition stores hundreds of petabytes of data from cars, machines, and robots, and adds more than a petabyte every day. Shiyan Xu explained why his team chose Lance for its training tables: training randomly samples from hundreds of millions of rows, and Lance makes it dramatically more efficient by fetching only the bytes each sample needs.
At NVIDIA, Max Li’s team trains the Cosmos world model on about a million hours of video. Its SILA platform keeps media, metadata, embeddings, and curation signals together in Lance, and the next generation of Cosmos is moving to train directly from those tables.
Physical AI companies have plenty on their plate building datasets, training models, and building machines. LanceDB lets them do that instead of getting bogged down in data infrastructure.
The leading AI companies now ship new models in weeks instead of quarters. How can research teams iterate on that cadence as data volumes explode?
Exa keeps 35 petabytes in Lance tables behind a search index of roughly 100 billion pages. Every time its researchers ship a better embedding model, the entire index gets re-embedded; one recent backfill ran on 35,000 Kubernetes pods. At Luma, researchers used to think twice before testing a new captioning model. Now it’s a new column they can launch themselves, often through an agent.
Nearly every talk came back to this capability. In conventional table formats, adding a column of new embeddings or captions means rewriting the whole table. On Lance, the new column is written on its own. As research agents start launching a hundred experiments overnight, cheap & fast iteration matters even more.
The best companies will be the ones that iterate fastest at scale, and Lance is building the infrastructure for human and agent researchers to do so.
Lance is becoming the standard for the most ambitious teams in multimodal AI, including NVIDIA, Uber, Netflix, Adobe, Exa, ByteDance, and other major AI labs.
As AI moves off the screen and into the physical world, it will run on Lance. Learn more by checking out the Reverie talks here: www.lancedb.com/reverie.