TwelveLabs raises $30M to make video as searchable as text

Finding a five-second clip in thousands of hours of uncataloged footage usually means someone has to sit there and scrub through timelines manually. TwelveLabs wants AI to do the watching instead.

On Thursday, the video understanding startup announced $30 million in fresh strategic funding, lifting its total raised to $107.1 million. The round drew heavyweight backing from Databricks, Snowflake, SK Telecom, HubSpot Ventures, and In-Q-Tel, the strategic venture arm tied to the U.S. intelligence community.

Alongside the cash injection, TwelveLabs appointed former Apple Siri lead architect and ex-SK Telecom CTO Yoon Kim as president and chief strategy officer to guide global expansion.

Video-first AI over general chat models

While frontier players like OpenAI and Google build massive multimodal models where video feels like a secondary add-on, TwelveLabs built its engine strictly for video from day one.

Its models translate natural language into visual actions, objects, and background sounds. Users can ask questions like "When did the person in the red shirt enter the restaurant?" and jump directly to that exact frame.

Beyond simple search, TwelveLabs offers Marengo, an any-to-any search model that lets users search across audio, images, and video, as well as an Embed API designed to generate multimodal embeddings for enterprise data.

That enterprise focus is already showing up in distribution. Both Databricks and Snowflake are weaving TwelveLabs directly into their ecosystems, letting corporate customers query video archives inside standard data pipelines.

According to co-founder Jae Lee, over 30,000 developers are building on the platform for use cases spanning highlight reels, ad placements, and municipal emergency response.

Lingering questions remain around algorithmic bias, however. TwelveLabs previously promised to release formal model-ethics benchmarks but has yet to deliver, though Lee maintains internal testing is active.

Why it matters

We think TwelveLabs played this round brilliantly by securing distribution through Databricks and Snowflake rather than trying to build a standalone enterprise silo. Video is by far the heaviest, most expensive enterprise data format, and by positioning itself as the native video search layer inside modern data warehouses, TwelveLabs makes itself very hard to rip out.

For media companies, retailers, and developers managing vast video repositories, TwelveLabs is worth testing now. Tailoring a video-native embedding model to specific archives generally yields cleaner pinpoint search than burning expensive tokens on broad multimodal models.

Still, there are obvious warning signs. The arrival of In-Q-Tel funding clearly points toward surveillance and intelligence applications, where misidentifications carry real-world consequences. With formal bias benchmarks still missing, TwelveLabs will need to prove its guardrails are as sharp as its search capabilities before expanding too deep into high-stakes verticals.