TwelveLabs Raises $100M and Signs AWS Trainium Deal to Scale Video AI

TwelveLabs raised $100M in its Series B on July 1, bringing total funding to roughly $150M since the company launched. NEA and NAVER Ventures co-led, with Amazon, Radical Ventures, Korea Investment Partners, Index Ventures, Quadrille Capital, and Red Bull Ventures also participating. TwelveLabs disclosed no valuation. Amazon came in as both a repeat investor and a strategic partner - a multiyear cloud commitment means TwelveLabs models will be optimized for AWS Trainium chips and launch on AWS before other platforms. CEO Jae Lee described the company's goal as building video superintelligence: making every second of video as addressable and searchable as text.

Half a Trillion Hours of Video Sit in Archives With No AI Access

"Video dark matter" is the phrase Lee uses for the problem. Factories, hospitals, satellite systems, security networks, broadcast archives, and sports organizations generate video every year that sits in storage with no way for AI systems to search or reason over it. Large language models made text programmable - embedding, retrieval, summarization, and search became infrastructure primitives developers could call from an API. Video never went through the same transition. Existing pipelines could transcribe audio, run frame-level object detection, or accept specific clips as multimodal input, but nobody had combined those signals into a unified, queryable representation of an entire video file.

For teams building agentic AI systems that need to reason over recorded content - security footage, training archives, call recordings, industrial sensor video - the gap shows up fast. Semantic search over text is a commodity now. Asking "find the moment in this 90-minute recording where the customer shows frustration while the presenter is on the pricing slide" still requires custom pipelines, manual review, or a model not designed for the job. TwelveLabs is positioning its stack as purpose-built video infrastructure rather than a vision layer bolted onto a general-purpose model.

Marengo 3.0 Indexes Visual, Audio, Speech, and On-Screen Text Into One Embedding

Marengo 3.0 is TwelveLabs's embedding model. Feed it a video and it maps all four modalities into a single vector space: what the camera captures, what the audio track contains, what speakers say, and any text visible on screen. A developer querying that embedding searches all four channels simultaneously. Separate extraction passes followed by a join are not required. A query for "presenter mentions pricing while the chart is on screen" returns timestamps because Marengo holds the visual, audio, and speech signals in one representation.

Pegasus 1.5 is the language model built on top of those representations. Where Marengo finds the right moment in a video, Pegasus answers questions about it in natural language, generates summaries grounded to specific segments, and produces descriptions tied to exact timestamps. Ingestion happens once. A video file processed by Marengo produces a durable representation that any downstream query or agent can access without re-ingesting the source - at archive scale, running the full ingestion pipeline on every query would make the system impractical. That design choice - ingest once, query forever - is probably what will matter most to developers already working with large video archives.

API access is live. TwelveLabs counts media companies, advertising agencies, government organizations, security firms, and sports teams among its customers - industries that hold video archives measured in decades with no viable way to make that footage computationally useful. A sports broadcaster can ask Pegasus to identify every defensive play in a 15-season archive by player and situation; a hospital compliance team can query 18 months of procedure recordings for specific protocol deviations. That breadth across verticals reflects the infrastructure framing: TwelveLabs is not optimizing for one use case.

Amazon Invested and Signed a Multiyear Trainium Commitment - New Models Launch on AWS First

Amazon joined the round as a repeat investor and formalized AWS as TwelveLabs's preferred cloud - a multiyear commitment covering training optimization on Trainium chips and first-launch access for new models. Launching on Trainium first gives Amazon two returns: TwelveLabs's inference spend stays on AWS infrastructure, and AWS customers get new video AI capabilities before they appear on other platforms. AWS is simultaneously building out its forward-deployed AI unit and investing heavily in custom silicon - TwelveLabs fits both bets.

NAVER Ventures co-leading the round suggests a second strategic angle. South Korea's largest internet company, NAVER also operates major video and content platforms - a stake in TwelveLabs makes sense as an infrastructure play for its own search and discovery problems. Korea Investment Partners also participated. Together AI's $800M Series C was built around open-source model serving at enterprise scale - TwelveLabs's raise is smaller but the bet is harder to replicate, since video understanding at this level requires purpose-built model architecture rather than better inference routing for existing models. AWS distribution gives TwelveLabs an enterprise sales channel it would otherwise have to build from scratch.