# Anyscale and CoreWeave Achieve 600 TB Video Captioning in 95 Minutes

> In an impressive feat, Anyscale and CoreWeave have managed to caption 600 TB of video data in just 95 minutes, redefining the standards for large-scale processing. This breakthrough emphasizes the critical role of advanced GPU workloads in meeting the escalating demands of AI-driven industries.

**Source**: wf.coreweave.com | **Published**: 2026-08-25 | **Type**: case_study

## Key Facts

- Anyscale and CoreWeave reduced GPU setup time from weeks to under 24 hours, enhancing agility.
- Scaling from 600 GB to 600 TB maintained efficiency, indicating strong infrastructure resilience.
- CoreWeave's AI Object Storage achieved 40 GB/s write speeds, showcasing superior data handling capacity.
- Ray Data improved caption output by 3.7x per GPU-hour, revealing significant competitive advantages.
- LOTA's caching reduced model load times to 2 seconds, highlighting critical operational efficiencies.

## Summary

\## Summary
CoreWeave partnered with Anyscale to deploy a video captioning pipeline capable of processing 600 TB of video data in just 95 minutes. The challenge was to efficiently manage GPU workloads for large-scale data processing, which they achieved by leveraging CoreWeave's infrastructure and Anyscale's orchestration tools. The result was a dramatic reduction in setup time, with the entire system operational within 24 hours.

\## Background
CoreWeave operates in the cloud computing industry, specializing in GPU infrastructure. Before deploying the video captioning pipeline, they faced significant delays in setting up GPU resources for large-scale data processing, often taking weeks to establish the necessary infrastructure.

\## Challenge
The primary challenge was to efficiently process vast amounts of video data for captioning while minimizing the time and resources required to set up the necessary GPU infrastructure.

\## Solution
CoreWeave and Anyscale implemented a two-layer video captioning pipeline. CoreWeave provided the AI infrastructure, including a dynamic cluster of 1,600 NVIDIA GPUs and AI Object Storage, while Anyscale managed the workload layer using Ray for task scheduling and data streaming. The deployment process was streamlined, taking less than 24 hours from account signup to a running production job.

\## Results
The video captioning pipeline processed 600 TB of video data in just 95 minutes. The system achieved a throughput of 12,666 captions per second, with the processing efficiency remaining steady even as the GPU cluster scaled from 256 to 1,600 GPUs. The CoreWeave AI Object Storage handled writes at 40 GB/s, and the latency for data retrieval was maintained at a p50 of 9.1 seconds.

\## Key Insights
Efficient orchestration across multiple infrastructure layers is crucial for large-scale data processing. Utilizing managed services like CoreWeave's AI Object Storage can significantly reduce setup time and prevent storage bottlenecks, enabling faster processing of large datasets.

\## Customer Testimonial
No direct quote is available from the source material.

## Entities

- **Companies**: Anyscale, CoreWeave
- **Products**: CoreWeave AI Object Storage, Ray OSS, Ray Data, Qwen3-VL-8B
- **Technologies**: GPU, Kubernetes, AI, LLM, NVMe

## Key Concepts

GPU workload, video captioning, data processing, infrastructure scaling, Ray OSS, CoreWeave AI Object Storage, Kubernetes, multimodal data

## Definitions

- **GPU**: A Graphics Processing Unit, used for accelerating data processing tasks, particularly in AI and machine learning.
- **Kubernetes**: An open-source platform for automating the deployment, scaling, and management of containerized applications.
- **Ray OSS**: An open-source framework for building and running distributed applications, particularly for AI and machine learning workloads.
- **CoreWeave AI Object Storage**: A storage solution designed for AI workloads, providing high-speed data access and management without egress fees.
- **LLM**: Large Language Models, which are AI models trained on vast amounts of text data to understand and generate human-like text.

## Use Cases

- Video captioning at scale
- AI infrastructure management
- Data processing for multimodal datasets
- Real-time video analysis
- Synthetic data generation
- Efficient resource allocation in AI workloads

## Frequently Asked Questions

**What is the significance of using GPUs for data processing?**

GPUs are essential for handling large-scale data processing tasks, especially in AI applications, due to their ability to perform parallel computations efficiently. This allows for faster processing times compared to traditional CPU workloads.

**How does CoreWeave AI Object Storage improve performance?**

CoreWeave AI Object Storage enhances performance by providing a high-speed data access layer that eliminates egress fees and utilizes local caching to reduce latency. This ensures that data is readily available for processing without bottlenecks.

**What role does Ray OSS play in the video captioning pipeline?**

Ray OSS manages the workload layer of the video captioning pipeline, scheduling tasks and scaling resources dynamically to match demand. This optimizes GPU utilization and ensures that the processing pipeline remains efficient.

**What are the benefits of using Kubernetes in this context?**

Kubernetes automates the deployment and management of containerized applications, which simplifies the orchestration of complex workloads. It allows for efficient resource management and scaling, crucial for handling large GPU clusters.

**How quickly can a video captioning pipeline be set up?**

With the right infrastructure in place, such as CoreWeave and Anyscale, a video captioning pipeline can be set up from account signup to production job in under 24 hours, significantly reducing the typical setup time.

## Links

- [Read on Welcome.AI](https://welcome.ai/content/anyscale-and-coreweave-achieve-600-tb-video-captioning-in-95-minutes)
- [Original source](https://wf.coreweave.com/blog/video-captioning-at-scale-600-tb-in-95-minutes-with-anyscale-on-coreweave)
- [CoreWeave](https://welcome.ai/company/coreweave): Featured company

---

Source: Welcome.AI | https://welcome.ai/content/anyscale-and-coreweave-achieve-600-tb-video-captioning-in-95-minutes