TwelveLabs raises $100 million to make video archives searchable

TwelveLabs, a San Francisco and Seoul company that builds models for understanding video, said on July 1 it raised $100 million in a Series B round. NEA and NAVER Ventures co-led it. Amazon, Radical Ventures, Korea Investment Partners, Index Ventures, Quadrille Capital, and Red Bull Ventures also put money in.

The press release sells the goal as "video superintelligence," and the founder's quote ends on "the road to Video Superintelligence starts here." Put the branding aside and look at what the company ships, because that part is more interesting than the tagline.

What the models actually do

TwelveLabs has two models. Marengo, now on version 3.0, is a video embedding model. It turns raw footage into vectors, so a clip, a spoken line, or an object on screen becomes something a database can index and a query can match. That is what makes an archive searchable without a person tagging every frame by hand.

Pegasus, on version 1.5, does the other half. It reads a video and emits structure: where scenes start and stop, which entities appear, what happens in each segment. TwelveLabs compares Pegasus to a markup language for video, which is a fair way to put it. It converts pixels into fields a program can read.

The reason any of this exists is a real limitation. A general-purpose language model cannot watch a two-hour video. It samples a few frames, misses everything between them, and starts from zero on the next question. TwelveLabs' argument is that video needs models born in video, not text models with a vision layer bolted on. As engineering, that position holds up, and it is the part of the story worth taking seriously.

The "superintelligence" claim, unpacked

What this round is meant to fund is an agentic layer on top of those two models. The company describes a system that keeps a persistent memory of every video it ingests and reasons across the whole store, rather than answering each query cold. Stripped of the branding, that is retrieval plus a memory index plus an orchestration layer, packaged as a product instead of a pile of parts.

It may be useful. It is not superintelligence, and the release backs the "world's most powerful video embedding model" line with no benchmark, no methodology, and nothing to compare it against. When a company claims a superlative and shows no measurement, that is marketing, not a result. Marengo may genuinely lead its category. The announcement just does not show the work.

The part with real weight

The concrete news is the Amazon relationship. AWS is TwelveLabs' preferred cloud, its models have been on Amazon Bedrock for over a year, and the two signed a multiyear commitment to move TwelveLabs' inference onto AWS Trainium, Amazon's in-house AI silicon. New models will ship on AWS first.

That is the sentence a competitor should read twice. Video inference is expensive because video is heavy, and a committed customer optimizing its workloads for Trainium hands Amazon a reference case for its own chips against Nvidia while handing TwelveLabs a cost structure it could not build alone. The funding number is $100 million. The move that lasts is where the compute runs.

TwelveLabs also shipped Rodeo, its first application-level product, earlier in June, a step from selling models to selling something a non-engineer can open and use. It is opening offices in New York and London on top of San Francisco and Seoul.

One number to treat with care

The round is a clean $100 million. The running total is not. TwelveLabs' own materials point to roughly $150 million raised after a $50 million Series A in 2024, while SiliconANGLE put the cumulative figure at "more than $207 million." Those do not reconcile, and the company did not publish a single reconciled number. Take the $100 million as solid and the total as approximate until someone states it plainly.