# Ray Can Schedule Every GPU You Own. Something Still Has to Feed Them.

**Author:** WEKA

**Published:** August 24, 2026

![WEKA and Anyscale logos on a dark, abstract background.](https://cdn.sanity.io/images/ult5g8gw/production/1ee21c2d5ddf2a992a17ee01a4a7e9a19d97ae64-1800x1013.png)

## TL;DR

A forward-looking approach at how WEKA and Anyscale are closing the storage gap in distributed AI.

At WEKA, we pay close attention to GPU compute trends, because the data infrastructure has to keep pace. Right now, that compute is concentrating around [Ray](https://www.ray.io/), the framework that's become the default for teams running AI at scale, from multimodal data curation and training to fine-tuning, RL, and online inference, across as many nodes as the job needs.

Multi-node Ray clusters leverage shared, persistent storage for checkpoints written across every job, object I/O, KV cache replicas, and model weight syncing. Most of that lands on whatever file service was easiest to mount, not on something purpose built for an AI-first access pattern.

That's why we've been building [WEKA® NeuralMesh™](/product/neuralmesh) to sit underneath it: one solution, purpose-built, to accelerate ingest, training, and serving, so checkpoints and cache reads never wait on storage that was built for something else.

Anyscale and WEKA consistently hear the same pain point from customers: resolving GPU utilization bottlenecks span every part of the stack, from storage to compute, and often multiple interfaces between the two. Together, NeuralMesh and Anyscale provide both sides of the solution spanning ingest, training, and serving, so checkpoints and cache reads can scale in lock step with compute orchestrated by Ray. .

## A Ray cluster is only as fast as its slowest read

Object storage was built for durability and cheap capacity, not for the concurrency, small files, and metadata load a Ray pipeline needs. So when I/O gets throttled, a checkpoint pauses, a loader stalls, or a KV cache gets re-pulled on every replica, the cost is often misdiagnosed as inefficient pipeline code. This shows up as idle GPU time and decreases overall compute utilization. The common fix, only looking at the pipeline, treats the symptom instead of the cause.

This is the same story across the AI infrastructure landscape right now: the industry keeps misdiagnosing a complex, cross-node data delivery problem as purely a compute challenge.

## What we're validating together

WEKA and Anyscale are proving NeuralMesh as a faster, more resilient storage layer for Ray workloads running on Anyscale to address one of the industry’s largest problems: bottlenecks spanning both the storage and compute layers. The goal is a supported configuration to unlock the benefits of both tools for our customers: large-scale compute orchestration, faster checkpoints, and faster recovery when a node fails, with as few changes to Ray application code to get there.

> “AI teams need to move from pilot to production faster than ever, and every team is battling data bottlenecks across every layer of the stack and underutilized GPUs that impact cost, reliability, and time to market. We're excited to prove what Anyscale and WEKA's high-performance data platform can accomplish together to accelerate that journey for our customers.”   
>   
> - Christian Stano, Field CTO, Anyscale

NeuralMesh is built for exactly the kind of workload Ray powers: one framework spanning ingest, training, and serving. And because the same architecture runs on premises, in a customer's own cloud account, or inside a GPU cloud, what gets proven in a pilot is what runs in production, which matters most for teams that don't want to re-tune storage every time capacity moves.   
  
**Ready to see NeuralMesh in action? [Reach out today](/contact)!**