TwelveLabs

The volume of video being produced and consumed globally has grown for two decades. In 2022, 65% of all internet traffic came from video. This increase has been driven largely by the decline in the cost of storing and delivering that footage. The price of Amazon S3 storage has fallen roughly 85% since the service launched in 2006, and the High Efficiency Video Coding standard, released in 2013, halved the bit rate needed to store video at the same visual quality. These cost declines have let governments and organizations vastly increase the amount of video they produce. In 2019, roughly 1 billion surveillance cameras were forecast to be installed globally by 2025. Meanwhile, YouTube's upload volumes grew from six hours of video uploaded every minute in 2007 to more than 100 hours per minute by 2013.

Yet the process of tagging video so it can later be searched and used remains largely manual. One employee at a professional sports franchise owner noted that it often took four minutes of human labor to tag just one minute of sports video. Querying the resulting archive can be just as slow. The NFL's video archive exceeded 200 petabytes as of January 2025, footage that would take over 175 years to watch continuously, and media producers can take hours or even days to locate and analyze the clips they need.

TwelveLabs develops video-native multimodal foundation models that let users search "entire video libraries using natural language" with no manual tagging needed. The company aims to unlock "the full potential of video for the world." Its method differs from what the company has called the dominant approach, which frames video understanding as an image or speech problem and, the company argues, loses the temporal context that understanding video requires. TwelveLabs describes itself as providing infrastructure for video intelligence, distributing its models through APIs that other companies build into their own products, while also selling an application and an agent for teams that want a finished workflow.

TwelveLabs was founded in 2021 by Jae Lee (CEO), Aiden Lee (CTO), Soyoung Lee (Head of Go-to-Market), Dave Chung (GTM), and SJ Kim. Jae Lee studied electrical engineering and computer science at UC Berkeley, and interned at Samsung and Amazon. He then returned to South Korea for mandatory military service, serving as lead data scientist at the Republic of Korea's Cyber Operations Command from 2018 to 2019. Aiden Lee studied at Brown University.

Jae Lee started the company with four friends he met at the Cyber Command, where they had been working on video understanding. The enlisted personnel at the Cyber Command lived on base, which gave the founders time to discuss their shared interests and ideate. Jae Lee's diagnosis was that language models had been trained on "the entire text that humanity has ever created" and still struggled with complex reasoning, yet video went largely unaddressed.

The team saw that the broader AI research community was split between those scaling language models and those adapting transformers for images, with essentially no one building purpose-built infrastructure for video. Rather than wrapping existing general-purpose models, the founders chose to build video-native foundation models from scratch. "We saw an opportunity," Jae Lee said, "that's when we started working on perceptual reasoning."

TwelveLabs participated in the 2021 Techstars Seattle Accelerator before raising a $5 million seed round led by Index Ventures in March 2022. Angel investors in that round included Dr. Fei-Fei Li of Stanford University and Alexandr Wang, then CEO of Scale AI, and the company also drew support from Aidan Gomez, a co-author of the paper that introduced the Transformer architecture and CEO of Cohere. A $12 million seed extension led by Radical Ventures in December 2022 added Nicolas Dessaigne (co-founder and former CEO of Algolia) and Lukas Biewald (CEO of Weights & Biases), alongside Jeffrey Katzenberg's WndrCo.

In December 2024, the company hired Yoon Kim as President and Chief Strategy Officer. Kim joined Apple in 2013 through its acquisition of Novauris Technologies, a speech recognition company where he was CEO, and managed Siri teams until 2018 before later serving as CTO of SK Telecom. Jae Lee described Kim as "instrumental in driving future growth with key acquisitions, expanding our global presence, and aligning our teams toward ambitious goals," and said that the company aimed to grow into adjacent verticals such as automotive and security.

Source: TwelveLabs

TwelveLabs provides a video intelligence platform built on two foundation models, Marengo and Pegasus. Marengo is an embedding model that turns video into searchable vectors, and Pegasus is a video language model that turns video into text and structured data. Developers reach the models through three APIs that the company's product menu lists as capabilities (Search, Analyze, and Embed), and TwelveLabs packages them into two applications, the Jockey agent and Compliance by TwelveLabs.

TwelveLabs Video Models

Source: TwelveLabs

Video, unlike text, has multiple modalities that interact with one another. Objects mix with audio, on-screen text, and motion, and the order in which they appear matters. A model has to decide not only which modalities are present onscreen, but when they start, when they end, and how they relate to one another. For instance, a person laughing while looking sad or saying something negative can only be read correctly by weighing facial expression against tone of voice and against the words themselves. Missing any single modality can flip the way the video is interpreted.

Early approaches to video understanding revolved around sampling still frames, encoding each one with an image model, and passing the resulting visual tokens to a language model alongside a text prompt. Flamingo, published in April 2022, used a Perceiver Resampler to compress visual features into a fixed set of tokens that a frozen language model could attend to through inserted cross-attention layers. While this dual-model approach could handle short clips, it degraded on causality and motion because each frame was encoded in isolation.

Gemini, announced in December 2023, brought reasoning within a single model and processes video by sampling one frame per second. At high media resolution, video costs roughly 300 tokens per second, so a one-million-token context window holds one hour of footage, and in standard 30-frames-per-second footage the model discards 29 of every 30 frames. This method forces a tradeoff between frame-level detail and temporal coverage. In May 2025, Apple researchers found that video models often answered benchmark questions without watching the video and performed similarly even when frames were reordered. Clearly, a model that scored the same on shuffled footage was not reasoning about sequence.

To preserve output quality while reasoning across hours of footage, TwelveLabs divided the task into two roles. A perception layer turns raw video into a searchable spatiotemporal representation, and a reasoning layer works over the stored evidence to produce output.

Marengo 3.5

Source: TwelveLabs

Marengo serves as a long-term memory layer, letting users index a video corpus once and answer questions quickly by retrieving the right indexed segments rather than re-running raw video through a model. Marengo 2.7, released in December 2024, introduced a multi-vector architecture that creates separate vectors for different aspects of a video, such as what things look like, how they move, and what is said, rather than compressing everything into a single vector. At launch, the company reported retrieval scores as much as 35% higher than those of external models on individual retrieval tasks.

Marengo 3.0, released in November 2025, added composed retrieval, which combines an image and a text query into a single query embedding to disambiguate visual attributes a text-only query would miss. On TwelveLabs' internal basketball benchmark, adding an image to a text query raised performance from 34.4 to 38.5 mAP. Version 3.0 also made the split explicit, returning separate visual, audio, and transcription embeddings for each segment, and added cross-language retrieval across 36 languages and cut embeddings to 512 dimensions, half the size of 2.7's, which lowers database costs and lets a fixed infrastructure budget index a larger library.

Marengo 3.5, released in August 2026, extends that design. It places video, audio, images, text, and documents in one 512-dimension embedding space, lets a single query mix text with an image, video, or audio clip, and removes 3.0's four-hour cap for asynchronous ingestion of long video. It also lets customers fuse timestamped metadata into the index at ingest, which the company reported raised average retrieval scores in its tests from 55.8 to 67.9.

Source: TwelveLabs

On TwelveLabs' eight-group retrieval benchmark, Marengo 3.5 scored 67.4 at 512 dimensions, against 62.4 for Google's Gemini Embedding 2 and 50.3 for Amazon's Nova Multimodal Embeddings, both at 3.1K dimensions. The gap in Marengo’s favor was widest on sports footage, where the company reported a score of 63.8 against 35.6 for Gemini Embedding 2.

Pegasus 1.6

Source: TwelveLabs

Pegasus started as a video question-answering system. Through version 1.2, users provided a single clip and asked a question, and the model returned an answer or summary. This worked well for users who already knew which clip they cared about, but the model had no way to sweep an entire video or library. Version 1.2 processed videos up to one hour long.

Pegasus 1.5, released in April 2026 at the NAB Show, raised that limit to two hours and changed what the model returns. Instead of answering questions about a predefined clip, the model partitions an entire video based on a user-defined schema and labels each segment with structured metadata, a feature the company calls time-based metadata. Given a basketball broadcast and a schema defining scoring plays, camera cuts, and non-gameplay footage as separate event types, Pegasus 1.5 returns timestamped segments for each event type in one request.

An example output for a scoring play runs from 0:46 to 0:52 and is labeled {"points_scored": "2pt", "shot_type": "layup", "scoring_team": "Loyola Ramblers"}. Users can define up to 10 event types per request, each with up to 20 custom fields that can be simple labels or full descriptions. In the company's testing, Pegasus 1.5 scored 0.43 against 0.34 for Gemini 3.1 Pro on segmentation quality, and 0.46 against 0.32 on multimodal prompting.

Pegasus 1.6, released in October 2026, is the company's first model built for egocentric, or first-person, video, the footage robotics labs collect to train physical AI systems. It also improves entity recognition and adds native image analysis through the same API, while keeping the two-hour input limit. The company said that time-based metadata in Pegasus 1.5 had "attracted several new customers."

Search, Analyze, and Embed APIs

Source: TwelveLabs

The Search API indexes video with Marengo into a TwelveLabs-hosted index. TwelveLabs handles vector storage, retrieval infrastructure, and search logic, so a customer queries its library in natural language without touching the raw embeddings. From Marengo 3.5 onward, the company offers hosted search through Jockey rather than as a standalone API.

The Embed API returns those embeddings to customers, who store them in their own vector databases and build their own retrieval systems. The Analyze API runs Pegasus to produce summaries, answers, and segmented metadata from video. Beyond its own API, TwelveLabs distributes Marengo and Pegasus through Amazon Bedrock, and it offers on-premise, hybrid, and cloud deployment through its sales team.

Jockey

Jockey is the agent layer TwelveLabs builds on top of Marengo and Pegasus. It began as an open-source reference project built on LangGraph, showing developers how the two models could be chained into an agent, and TwelveLabs has since rebuilt it as a first-party product in research preview. The Jockey product page describes the division of labor plainly: "Marengo finds it, Pegasus describes it, Jockey turns it into something you can query." Jockey's first capability is global entity understanding, which recognizes the same person across thousands of videos and photos in a library and groups everything they appear in. Marketing teams and creators can use Jockey as a connector inside Anthropic's Claude, and developers can build on it through an API.

Compliance by TwelveLabs

Source: TwelveLabs

Compliance by TwelveLabs, which became generally available in September 2026, reviews media for content that breaks broadcast and platform rules. The company calls it the first application built on its platform; Jockey, released earlier, remains in research preview. Customers author versioned rule packs, starting from more than 40 pre-configured regional packs, and every flag the product raises includes a timecode and a written justification.

It integrates NVIDIA's Synthetic Video Detector to score suspected deepfakes, and TwelveLabs targets a reviewer false-positive rate of 15% or lower. In internal evaluation, the company reported that the product's findings agreed with a manual review baseline 92.5% of the time and cut reviewer time by 80%. As of October 2026, the product ran only as a fully managed service.

Customer

TwelveLabs’ customers include organizations that hold large video archives and need to find, label, or reuse footage faster than people can tag it. It organizes its go-to-market around six verticals, which are media and entertainment, advertising, sports and broadcasting, government, security, and, since October 2026, physical AI.

The company does not disclose its full customer base, but its published case studies span those verticals, with SBS and UNICEF Korea in media, Dyn Sport and Maple Leaf Sports & Entertainment (MLSE) in sports, GS SHOP, AffiliateNetwork, and Mantis Solutions in advertising and commerce, and Protege as an AI data supplier. Several of those customers, including SBS, UNICEF Korea, and GS SHOP, are South Korean, and TwelveLabs operates offices in Seoul and Pangyo alongside San Francisco, New York, and London. TwelveLabs also sells through Amazon Bedrock, and GS SHOP uses Marengo through Bedrock rather than through TwelveLabs' API.

Market Size

The multimodal AI market was estimated at $3.9 billion in 2026 and is projected to grow to $13.5 billion by 2031. Within that market, video is projected to grow faster than any other data modality between 2026 and 2031, at a 39.8% CAGR, as real-time video analytics improve and live-streamed content volumes rise. A narrower estimate of the video analytics market, which covers software that interprets recorded and live footage, projects growth from $14.7 billion in 2026 to $41.4 billion by 2031.

Media archives are a core market for TwelveLabs. The media asset management market is projected to grow from $7.2 billion in 2025 to $8.8 billion in 2026, reaching $19.5 billion by 2030. The media and entertainment storage market is projected to expand from $18.7 billion in 2026 to $49.4 billion by 2035 as streaming services grow their archives and ultra-high-definition content libraries multiply. Each of these markets is a growing pool of footage that has to be found, tagged, or reused, which is the work TwelveLabs' models automate.

Advertising budgets are moving to connected TV. US connected TV ad spend is forecast to reach $38 billion in 2026, up 15% from 2025, and connected TV is on pace to exceed 40% of video ad spend by 2030. Sports and broadcasting face a similar volume problem. During the 2026 Milano Cortina Games, Olympic Broadcasting Services produced more than 6.5K hours of content. As of May 2026, sports was the content category growing fastest in the catalogs of major subscription video-on-demand platforms, and Paramount+'s sports programming grew 219% year-over-year after it acquired UFC rights.

TwelveLabs has two categories of competitors. First, the largest cloud providers sell general-purpose multimodal models and, since October 2025, video-capable embedding models, sometimes through the same marketplaces that distribute TwelveLabs. Second, venture-backed startups compete for the same archives as TwelveLabs, mostly with finished applications rather than model APIs. TwelveLabs differentiates itself from both camps with an architectural approach: the company holds that video should be understood once at ingest and kept addressable, rather than re-read by a general model at query time.

Frontier Model Providers

Google: Google builds the Gemini family of natively multimodal models, which accept video as an input alongside text, images, and audio. Gemini models with a one-million-token context window can process up to three hours of video at the default low media resolution, or one hour at high resolution. Each query re-reads the footage, which TwelveLabs argues makes Gemini an inspection tool rather than an index.

In March 2026, Google began selling an index-side product as well. Gemini Embedding 2, in public preview since March 2026, places text, images, video clips of up to 120 seconds, audio, and documents in one embedding space, the job TwelveLabs assigns to Marengo. Google's Gemini Omni, launched in May 2026, extended the family toward video generation.

Microsoft: Microsoft sells Azure AI Content Understanding, which processes videos up to two hours long and returns shot boundaries and schema-defined structured output ready for a retrieval index. Its Azure AI Video Indexer extracts insights from video and deploys to edge locations through Azure Arc, which matters for customers who cannot send footage to a public cloud.

Amazon: Amazon sells video understanding through Amazon Bedrock, where the Nova Lite, Nova Pro, and Nova Premier models accept video input and return text. Nova Multimodal Embeddings, available in Bedrock since October 2025, handles text, documents, images, video, and audio in one model for retrieval and semantic search, the job TwelveLabs assigns to Marengo.

Marengo and Pegasus have also been distributed on Bedrock since July 2025, and AWS was the first cloud provider to offer Marengo 3.0. Amazon participated in TwelveLabs' July 2026 Series B, which named AWS the company's preferred cloud provider and included a multi-year commitment to run TwelveLabs' video inference workloads on Amazon's Trainium chips. Amazon is simultaneously a major distribution channel for TwelveLabs, its compute supplier, an investor, and the vendor of a competing embedding model.

Video AI Startups

Reka AI: Founded in July 2022, Reka AI is a multimodal model lab whose Reka Vision platform is designed "to interpret, search, and reason across video and image content at scale." It raised a $110 million round in July 2025 led by NVIDIA and Snowflake, at a $1 billion valuation, and had raised $170 million in total as of October 2026. In June 2026, Reka merged with Moonvalley to build models for physical AI. That puts Reka in the same physical AI market TwelveLabs entered in October 2026, and the two share investors, NVIDIA and Snowflake, though Reka builds video search as one product inside a general multimodal lab rather than as the core of the company.

Coactive AI: Founded in 2021, Coactive AI sells a platform that lets enterprises search, classify, and moderate image and video libraries without manual tagging. It raised a $30 million Series B in May 2024 co-led by Cherryrock Capital and Emerson Collective at a $200 million valuation, and had raised $44.1 million in total as of October 2026 from investors including Andreessen Horowitz, Bessemer Venture Partners, and Greycroft. In June 2026, Coactive launched Context Studio, which turns advertiser briefs into contextual ad packages, moving it into the advertising budgets TwelveLabs also targets. Coactive sells a finished platform to data and marketing teams and treats images and video as one problem rather than specializing in video.

Moments Lab: Founded in 2016 as Newsbridge, the Paris-based Moments Lab indexes video archives for media companies, sports rights holders, and brands using its MXT-2 multimodal model. It raised a $24 million round in June 2025 led by Oxx, and had raised $37.4 million in total as of October 2026 from investors including Orange Ventures, Kadmos, Supernova Invest, and Elaia Partners; it has not disclosed a valuation. Moments Lab competes for the media and entertainment budgets that anchor TwelveLabs' traction, but sells a workflow application rather than model APIs, placing it closer to TwelveLabs' own applications than to Marengo or Pegasus.

Memories.ai: Launched in 2024 by Shawn Shen and Ben Zhou, who built the AI system behind Meta's Ray-Ban glasses, Memories.ai builds what it calls a Large Visual Memory Model, which stores and recalls visual data across long timeframes. It had raised $16 million as of March 2026 across two rounds, a seed in July 2025 led by Susa Ventures and an extension, with participation from Seedcamp, Fusion Fund, and Crane Venture Partners; no valuation has been disclosed. In September 2026, Qualcomm said that Memories.ai runs on-device models on its Snapdragon chips for smartphones and wearables. Memories.ai argues, as TwelveLabs does, that persistent visual memory beats re-reading footage at query time, but aims that architecture at wearables and robotics rather than enterprise archives.

TwelveLabs sells developer access to its models through usage-billed APIs and sells its applications by subscription. It offers three tiers of usage. The free plan is capped at 10 hours of indexing, the Developer plan is billed by usage, and the Enterprise tier carries committed-use contracts and custom pricing.

Source: TwelveLabs

Search API: Indexing a video into Marengo costs $2.50 per hour of video as a one-time fee, and TwelveLabs charges a recurring infrastructure fee of $0.09 per hour of video per month for embedding storage and management. Because the infrastructure fee is billed against the cumulative size of a customer's library rather than only new uploads, the cost of keeping video on TwelveLabs' platform compounds as an index grows. Search queries against an indexed library are billed at $4 per 1K queries. From Marengo 3.5 onward, hosted search is reached through Jockey.

Embed API: On Marengo 3.5, the Embed API is priced per one million tokens by input type, from $0.065 for audio to $0.26 for video, with images at $0.08, text at $0.20, and documents at $0.162.

Analyze API: Pegasus usage is billed at $1.75 per hour of input video and $7.50 per one million output tokens, and Pegasus 1.6 adds image input at $3 per one million tokens. The Segment capability, which uses Pegasus to break a video into defined windows, carries a multiplier that a standard Analyze call does not. Billed duration equals the video window multiplied by the number of segment definitions in the request, so a 60-minute video processed against four segment definitions bills as 240 minutes rather than 60. This structure rewards customers who scope their Segment calls tightly and penalizes broad, unbounded queries.

Jockey: Jockey is billed monthly or annually, with annual commitments priced 17% below the monthly rate, and is tiered by how much media a customer stores. The free tier provides 5 GB of media storage, the $20-a-month Plus tier provides 100 GB, and the $100-a-month Pro tier provides 500 GB and the highest rate limits. Every tier runs on the same underlying Marengo and Pegasus models.

TwelveLabs' serving costs are tied to AWS. The July 2026 Series B included a multiyear commitment to optimize TwelveLabs' video inference workloads on Trainium chips, and new TwelveLabs models launch first on AWS. The recurring infrastructure fee on hosted indexes bills the cost of storing embeddings to customers as their libraries grow.

TwelveLabs had more than 30K developers using its platform as of December 2024. Its published case studies report results at named customers. MLSE cut search and retrieval for its production team from 16 hours to nine minutes, and a senior director of media operations there said, "We can't imagine our content production process without TwelveLabs." GS SHOP, a Korean home-shopping retailer, reported that the number of customers placing orders rose 57.5% after it deployed Marengo, and UNICEF Korea cut retrieval time across an 8 TB archive by 95%.

In July 2025, TwelveLabs signed a memorandum of understanding with LG CNS to jointly advance video analysis technology and pursue new business together; LG CNS's public-sector clients include Korea's Ministry of Foreign Affairs and National Police Agency. Later that month, TwelveLabs' models became available in Amazon Bedrock, which powers generative AI for more than 100K organizations. In September 2026, Marengo became the first video model in Amazon Bedrock Managed Knowledge Base, with Iconik, Backlight's media asset management platform, among its first users. One unverified estimate indicates TwelveLabs generated $5.8 million in revenue in 2024, up 40.8% from 2023.

TwelveLabs raised a $100 million Series B in July 2026, co-led by NEA and NAVER Ventures, with participation from Amazon, Radical Ventures, Korea Investment Partners, Index Ventures, Quadrille Capital, and Red Bull Ventures. The round brought TwelveLabs' total funding to $210.1 million across nine rounds. As of October 2026, its investors also included In-Q-Tel, a not-for-profit strategic investment firm set up by the CIA in 1999 to source commercial technology for the US intelligence and defense communities.

Before the Series B, TwelveLabs raised a $3 million corporate round from Firstman Studios in October 2025 and an undisclosed round from NAVER Ventures in June 2025. In December 2024, it raised $30 million in strategic investments from Databricks, Snowflake, SK Telecom, HubSpot Ventures, and In-Q-Tel, bringing its total funding at the time to $107.1 million.

Six months earlier, in June 2024, it closed a $50 million Series A co-led by NEA and NVIDIA's venture arm NVentures, which followed a $10 million strategic round in October 2023 from NVentures, Intel, Samsung Next, and Korea Investment Partners. The company's seed financing totaled $17 million, made up of a $5 million round led by Index Ventures in March 2022 and a $12 million extension led by Radical Ventures in December 2022, following a $120K Techstars pre-seed investment in January 2021. TwelveLabs has not publicly disclosed a valuation for any of its rounds.

Several of TwelveLabs' investors also distribute or integrate its models: NVIDIA in October 2023 and the Series A, Databricks and Snowflake in December 2024, and Amazon in the Series B. Databricks paired TwelveLabs' multimodal embeddings with Databricks AI Search, letting customers build video search, recommendation, and analysis systems within its platform. The Snowflake integration stores TwelveLabs' video embeddings in Snowflake tables using the VECTOR data type, so Snowflake customers can run video search alongside the rest of their data.

Labeling Footage for Physical AI

Physical AI systems, from autonomous vehicles to robots, are trained and audited on footage at a scale that manual review cannot keep up with. A fleet of 62K autonomous vehicles can generate one petabyte of uploaded data per month. Regulators have layered a reporting clock onto that footage. NHTSA's Standing General Order 2021-01 requires manufacturers and operators of vehicles with Level 2 driver-assistance or automated driving systems to file an initial report on the most serious crashes within five days of learning of them. Locating the right clip inside a fast-growing, largely unlabeled archive against a regulatory deadline is one example of the search-and-retrieval problem Marengo and Pegasus were built to solve, and labeling that footage for model training is another.

TwelveLabs entered this market through training data. In October 2026, the company launched a physical AI offering built on Pegasus 1.6 that labels first-person footage for robotics and physical AI teams, covering action segmentation, dense captioning, quality scoring, search, and consent and compliance flagging. Its solution page pitches natural-language search for "rare events, duplicates, and long-tail edge cases" across a footage library, the same retrieval job fleet operators face. TwelveLabs has not named a customer in robotics or automotive, and specialists are entering the space. NomadicML raised $8.4 million in March 2026 for natural-language search over robotics and autonomous-vehicle fleet video.

Public-Sector Video Search

Government and security buyers are converging on a similar shift from manual monitoring to networked, searchable video. In 2020, 62% of US local police departments used body-worn cameras, a 43% increase over the share in 2016, and the global body-worn camera market is projected to grow from $1 billion in 2025 to $1.8 billion by 2030. Service contracts for these cameras bundle multi-year cloud storage and AI redaction, expanding the pool of stored footage agencies need to search for evidentiary purposes.

Broader video surveillance follows the same pattern. The video surveillance market was estimated at $83.5 billion in 2025 and is projected to reach $204.7 billion by 2033, an 11.7% CAGR, and surveillance-as-a-service is projected to grow at an 11.1% CAGR through 2031.

That shift toward outsourced analysis and storage creates demand for the search-and-analysis layer TwelveLabs sells. TwelveLabs said in its Series B announcement that it works with governments on mission-critical video workflows, and In-Q-Tel invested in its December 2024 round. Competition for these buyers is also well funded. Airis Labs launched with $60 million in May 2026 to sell analysis of social media, CCTV, and drone footage to public-sector agencies.

Amazon Transitions Away from TwelveLabs

Amazon sells a substitute for Marengo's embedding model. Amazon Nova Multimodal Embeddings, available in Bedrock since October 2025 and in AWS GovCloud since August 2026, is what AWS calls "the first unified embedding model that supports text, documents, images, video, and audio," the same cross-modal retrieval function as Marengo's Embed API. AWS's own guidance presents the two side by side. A March 2026 AWS machine learning post describes a single video-search architecture that "supports both embedding models with a unified interface," naming Nova Multimodal Embedding and TwelveLabs Marengo as alternatives.

Amazon has kept building video capability into Nova independent of its relationship with TwelveLabs. Nova 2 Lite, introduced in December 2025, processes text, image, and video inputs, and Nova 2 Omni is the first Nova model built for full multimodal generation.

Amazon has competed with products it distributes before. In 2019, open-source companies accused Amazon of "strip mining" their software by launching its own versions of popular open-source projects. Elastic sued AWS for trademark infringement and false advertising, MongoDB, MariaDB, and RedisLabs made similar complaints, and Amazon denied Elastic's claims.

TwelveLabs' position differs in two ways. Its models are proprietary rather than open source, and Amazon's equity stake gives it a return tied to TwelveLabs' success. The underlying exposure is the same: a platform owner selling a competing native product to the same customers. If Nova closes the performance gap, Amazon controls both the marketplace where TwelveLabs' models are sold and the Trainium infrastructure that runs its inference.

Open Reference Architectures Could Commoditize the Model Layer

NVIDIA's AI Blueprint for Video Search and Summarization is a GPU-accelerated reference architecture for video agents supporting natural-language search, visual question answering, clip retrieval, and long-video summarization. The pipeline splits video into 10- to 30-second chunks, captions each with a vision-language model, can merge the output with object detection and audio transcription, and writes the result to vector and graph databases queried later. This mirrors the understand-once, store-durably, retrieve-later sequence TwelveLabs describes as central to its approach.

NVIDIA distributes the blueprint as free reference code, and version 2.4.1 added support for smaller variants of Alibaba's open-weights Qwen3-VL model. The Qwen team reports that the flagship Qwen3-VL model performs on par with Gemini 2.5 Pro and GPT-5 at low reasoning settings on video benchmarks. An enterprise with existing GPU capacity can assemble a pipeline that does the same job as Marengo and Pegasus without paying a model vendor.

Production use already exists. Linker Vision built a city-scale deployment on the blueprint that serves 12 municipal departments and, as of May 2025, was scaling from 30K city cameras to more than 50K in 2026. The National Hockey League used VAST Data's InsightEngine with the blueprint to speed up vision AI workflows on its game footage. VAST Data is also the partner TwelveLabs selected in February 2026 for its first customer-managed deployment path.

The blueprint requires GPU capacity, deployment, and ongoing tuning that a managed service does not, and NVIDIA sells the GPUs it runs on regardless of which model a customer chooses. TwelveLabs has responded by moving up the stack. Jae Lee argued in July 2026 that "models commoditize" while "the intelligence layer that composes them does not." That position is tested hardest by the customers TwelveLabs is pursuing. Governments, defense agencies, and regulated industries with data-residency requirements are also the customers most likely to self-host a free reference architecture instead of paying for a managed model.

TwelveLabs builds foundation models that let organizations search, analyze, and reason over video natively, addressing a market where recorded video volume is growing faster than organizations can manually review or catalog it. The company has published customer results across media, sports, and commerce, raised a $100 million Series B in July 2026, and has since released Marengo 3.5, Pegasus 1.6, and Compliance by TwelveLabs as it shifts from a standalone model vendor toward a full-stack system for video. The core question is whether that platform can hold up against general-purpose multimodal and embedding models from larger labs and cloud providers, particularly Amazon and Google, while revenue outside its strongest verticals, media and sports, remains largely unestablished.