CoreWeave and Anyscale Caption 600 TB of Video in 95 Minutes

A 1,600-GPU test of multimodal data processing

CoreWeave and Anyscale published an engineering case study on August 25, 2026 describing a large-scale video-captioning pipeline that processed 600 TB of video in 95 minutes.

The system used up to 1,600 NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, 200 AMD Turin 9655P nodes, Ray Data, CoreWeave Kubernetes Service and CoreWeave AI Object Storage. Qwen3-VL-8B handled caption and annotation inference.

This is an engineering benchmark and case study, not a new foundation model release. The performance figures are reported by CoreWeave and Anyscale for their specific pipeline and infrastructure configuration.

From 600 GB to 600 TB

The teams first tested a 600 GB workload on 256 GPUs and compared a manually managed Ray Core design with Ray Data.

CoreWeave reports that Ray Data used 73% of the GPU time required by the Ray Core version and produced 3.7 times more captions per GPU-hour in that smaller comparison.

The team then expanded the dataset to 600 TB by generating synthetic video variations. At peak, 35,820 CPU cores transformed 43,000 original clips into 70 million clips for the scale test.

What happened at 1,600 GPUs

For the full-scale captioning run, CoreWeave says its object storage delivered roughly 110–120 GB/s across about 46 CPU nodes while the pipeline produced captions with Qwen3-VL-8B.

Captioning finished in 95 minutes.

Reported total throughput increased from 938 captions per second in the smaller test to 12,666 captions per second at the larger scale, while caption-per-GPU-hour efficiency remained roughly steady.

The authors attribute the scaling result to streaming execution, better CPU/GPU pipelining, autoscaling, local caching and storage throughput that kept the accelerators fed.

Model loading and storage were part of the bottleneck

The case study highlights an infrastructure problem that becomes visible at large scale: loading the same model onto hundreds or thousands of GPUs can create a storage thundering herd.

CoreWeave says its Local Object Transport Accelerator cached model data on node-local NVMe. In the reported test, a 17 GB Qwen3-VL model loaded at 8.68 GB/s per engine from local cache and reached an engine in about two seconds.

Those numbers are environment-specific and should not be generalized to other clouds, models or storage layouts without independent testing.

Why this matters for AI training data

Training and fine-tuning multimodal systems increasingly require large volumes of video, audio, LiDAR or image data to be decoded, filtered, annotated and normalized before model training begins.

This case study shows that the data-preparation stage itself is becoming a major GPU workload. It also illustrates why scheduling, storage locality, CPU preprocessing and fault handling can determine useful accelerator throughput as much as the model kernel does.

The practical takeaway is not that every 600 TB workload will finish in 95 minutes. It is that large-scale multimodal preprocessing can be designed as a streaming distributed system rather than a sequence of disconnected batch jobs.