What challenges did TwelveLabs face before AI implementation?

Scaling video AI from research to global production required infrastructure for training multimodal models and running inference on petabytes of video data. Traditional search methods fail on video archives with millions of hours because the data volume and complexity are too massive.

What was the impact of the AI solution?

The architecture enabled production deployment and unit-economics improvements at petabyte scale. With TwelveLabs models available in Amazon Bedrock, developers can build video AI applications while maintaining control over their data.

Source

ProductionEvidence: Medium55/100

TwelveLabs builds multimodal video understanding using Amazon Bedrock and SageMaker HyperPod

Use case typeContent discovery assistantUpdated Jul 15, 2026

TwelveLabs developed multimodal video foundation models and deployed them on AWS to support training and high-volume inference at global production scale. Its Marengo and Pegasus models power semantic video search, summarization, and structured outputs over massive archives.

Organization: TwelveLabs
Industry: Tech & Comms
Location: United States
Published: July 2026

Reported outcomes

Strategic outcomes

Scale & capacityProduction deployment at petabyte scaleBetter decisions & insightMade video archives searchable and actionableCost efficiencyImproved unit economics

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: TwelveLabs
Provider: AWS
Maturity: Production
Linked source: AWS Solutions Case Study

TwelveLabs developed multimodal video foundation models and deployed them on AWS to support training and high-volume inference at global production scale

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 3 of 3

1Content discovery assistant
2Multimodal analytics
3Search modernization

Scaling video AI from research to global production required infrastructure for training multimodal models and running inference on petabytes of video data.
Traditional search methods fail on video archives with millions of hours because the data volume and complexity are too massive.

TwelveLabs built a unified AWS production stack with EKS, EC2 GPU instances, S3, and S3 Vectors to store video assets and embeddings together.
The company trained models across thousands of GPUs with Amazon SageMaker HyperPod and used vector search for sub-second retrieval across billions of embeddings.

Technologies

Amazon Bedrock Amazon SageMaker HyperPod Amazon EKS Amazon EC2 Amazon S3 Amazon S3 Vectors

The architecture enabled production deployment and unit-economics improvements at petabyte scale.
With TwelveLabs models available in Amazon Bedrock, developers can build video AI applications while maintaining control over their data.

Architecture

TwelveLabs trains multimodal video foundation models on Amazon SageMaker HyperPod, runs production workloads on Amazon EKS and Amazon EC2 GPU instances, stores raw video and processed multimodal data in Amazon S3, and uses Amazon S3 Vectors to co-locate searchable embeddings with video assets for sub-second semantic retrieval.

Sources & evidence1

Evidence: Medium55/100Evidence strength