The dark matter of enterprise data
Twelve Labs has raised 100 million dollars in a Series B co-led by NEA and Naver Ventures, with Amazon returning alongside Radical Ventures, Index Ventures and others, bringing the company's total funding past 200 million dollars. The pitch behind the round is a single, vivid idea from chief executive Jae Lee: the world's video is still mostly dark matter to machines. It sits in archives, drones and satellites, accessed only through filenames, folders, captions and transcripts, none of which describe what is actually happening on screen.
For enterprises, the framing lands because it names a problem they feel but rarely quantify. Organizations generate and store enormous volumes of video, from security footage and manufacturing lines to recorded meetings and field inspections, and almost none of it is queryable in any meaningful way. The text corpus of a company has been mined for a decade. Its video corpus remains opaque, searchable only by whatever a human happened to type into a filename. Twelve Labs is betting that closing that gap is a category sized opportunity, not a feature.
Two models, one architecture
The company's approach rests on two models with a clear division of labor. Marengo handles perception, indexing what appears in footage, and Pegasus handles reasoning, answering questions about it. Together they are designed to treat as much as two hours of video as a single context, so a query can span a full recording rather than isolated clips. Lee describes the goal as making every second of video addressable, searchable and usable by agents, which is a more ambitious claim than transcript search dressed up in new language.
The distinction between perception and reasoning matters technically. Transcribing speech or tagging objects is a solved problem in isolation; understanding a two hour recording well enough to answer a specific question about cause, sequence and consequence is not. By splitting the work into an index and a reasoning layer, Twelve Labs is trying to build something an agent can call as a tool, the way it would call a database. If that architecture holds up at enterprise scale, it becomes infrastructure. If it does not, it is a better video search box, which is useful but not a platform.
Why Amazon came back
Amazon's participation is the most strategically loaded part of the round. This is a repeat investment, and Amazon used it to make AWS the company's preferred cloud, with new Twelve Labs models tuned for AWS Trainium chips and launching there first. That is the familiar hyperscaler pattern: capital that arrives bundled with infrastructure commitments and silicon alignment. For Amazon, backing a video understanding leader deepens the AI catalog it can offer enterprise customers without building the capability in house.
For Twelve Labs, the arrangement is a double edged asset. Preferred cloud status and Trainium optimization lower costs and buy distribution through the largest cloud channel in the market. They also tie the company's fortunes to one platform's roadmap and pricing. We have watched enough startups discover the limits of a hyperscaler embrace to flag the risk plainly. The upside is real and immediate; the constraint shows up later, when a company that wants to be neutral infrastructure finds its most important relationship pulling it toward a single ecosystem.
The use cases hiding in the archive
The commercial case for video understanding is easiest to see in industries that already drown in footage. A manufacturer can ask why a line stopped and get an answer grounded in the actual video, not a maintenance log written after the fact. A media company can search decades of archive by what happens on screen rather than by metadata a human once typed. Security and compliance teams can query hours of recordings for a specific event without scrubbing through them frame by frame. In each case the value is not novelty; it is turning a cost center into a queryable asset.
What makes the timing credible is the shift toward agents. An agent that can plan and act is only as capable as the tools it can call, and until now video has been a tool it could not use. Give an agent a reliable way to search and reason over footage, and a class of workflows that were manual by necessity becomes automatable. That is the bet the funding underwrites: that video understanding graduates from a standalone product into a capability agents reach for by default, the way they already reach for text search.
Video understanding is not video generation
Twelve Labs is careful to distinguish its work from the video generation race that has captured most of the attention and capital. The frontier labs are competing to synthesize video from prompts. Twelve Labs is competing to make existing video comprehensible. Those are different problems with different customers. Generation is largely a creative and marketing story. Understanding is an enterprise data story, closer to search and analytics than to media production.
That separation is a smart piece of positioning, because it keeps Twelve Labs out of a direct fight with the best funded companies in AI. Rather than trying to out generate labs spending billions, it is claiming a lane those labs have mostly skipped. The risk is that a general purpose frontier model eventually absorbs video understanding as a capability, the way large models have swallowed other specialized tasks. Twelve Labs' defense is depth: two purpose built models, long context, and enterprise integration that a generalist would have to rebuild. Whether depth is a durable moat against scale is the open question every specialized AI company now faces.
The frontier the labs skipped
The most interesting thing about this round is what it says about where value is migrating. As the largest models converge on similar text and image capabilities, the differentiated opportunities are moving to the modalities and workflows the giants have not prioritized. Video is the richest of those, the densest source of real world data and the least accessible to machines today. A company that owns the layer between raw footage and reasoning agents is positioned in exactly the gap the frontier left open.
For technology executives, the practical signal is to inventory the video their organizations already own and treat it as latent data rather than storage overhead. The tooling to make it queryable is arriving, backed now by a well capitalized specialist and the largest cloud provider. The companies that move early will turn archives they currently pay to store into assets they can interrogate. The rest will keep filing footage under whatever a human named it, which is to say they will keep it dark, right up until a competitor turns the lights on.
