# How Andromeda Scales Its GPU Market With WEKA

**Customer:** Andromeda AI

Andromeda has improved performance across its fleet, cut environment load times, cleared metadata bottlenecks, and lowered storage costs with WEKA.

## Results

- **6x** Faster Environment Startup (three minutes to 30 seconds)
- **Hours → Minutes** for Metadata Operations
- **400+GB/s** Throughput and 6M+ IOPS per Cluster in Production

Introduction

## Same Performance on 50+ Providers

Andromeda began as a single GPU cluster built to give Nat Friedman and Daniel Gross's portfolio companies a shared pool of compute. As demand grew, so did the need to work with more providers, as customers arriving from major hyperscalers expected a certain baseline of performance regardless of where their workloads ran. That expansion turned Andromeda into a market operating across more than 50 compute providers worldwide and managing tens of thousands of GPUs, with growth still ahead of it.

In order to ensure consistent performance across all customers, irrespective of the hardware involved, storage must be treated as core to every deployment, covering hardware, software, and the data layer in between. This requires storage that meets workload-specific performance requirements, scales as Andromeda adds providers, and can be operated efficiently by its infrastructure team.

Andromeda was in need of a storage layer built to go the distance and continue to raise the ceiling on behalf of the customers who depend on the platform every day.

The Challenge

## Metadata Bottlenecks Stalled GPU Clusters

Vishvajit Kher, Andromeda's lead architect, and Nicolas Joniaux, who leads AI infrastructure, own that promise day to day, and they were the ones searching for a storage partner that could keep pace with it.

Andromeda's clusters originally ran on an alternative storage platform. The trouble surfaced when the platform was getting slower under workloads that were, in Kher's words, "bottlenecked by storage." The culprit: metadata-heavy operations.

Andromeda's earlier architecture added its own constraint. The team had built around dedicated storage deployments, and when a given cluster didn't have one, there was no shared alternative to fall back on. Joniaux put it plainly: without that option, Andromeda was "essentially blocked in the process of utilizing the GPU cluster."

The Solution

## Dedicated or Converged Storage

In its initial test on one cluster, Andromeda observed improved performance using the same hardware, without adding new physical infrastructure.

Because WEKA® NeuralMesh™ is software-defined, Andromeda wasn't locked into a single deployment model. WEKA NeuralMesh Axon™ followed as a second option built on that same platform, fused directly into the GPU servers themselves, giving Andromeda the freedom to run dedicated or converged storage on whichever cluster made sense rather than standardizing on one architecture. That flexibility gives Andromeda more options for matching storage deployments to customer requirements.

### NeuralMesh: Every GPU Core Freed Up for Compute, Not Storage

NeuralMesh is WEKA's core storage platform, built to run on dedicated infrastructure that scales independently of the GPU cluster it feeds. The same architecture that ran Andromeda's very first CoreWeave cluster still runs its largest deployments today, without a rebuild in between.

For Andromeda, that separation matters most on clusters where customers want every GPU core dedicated to compute and expect the storage layer to stay up independent of the GPU nodes' own reliability, connecting across different protocols as needed.

```json
{
  "_key": "cf242bcf814d",
  "_type": "customerStoryMedia",
  "image": {
    "alt": "3D spiral metal loop",
    "asset": {
      "_id": "image-0fcf0c59275cc7a96e89303f2ceece54ec3c224f-3840x2160-jpg",
      "metadata": {
        "dimensions": {
          "_type": "sanity.imageDimensions",
          "aspectRatio": 1.7777777777777777,
          "height": 2160,
          "width": 3840
        },
        "lqip": "data:image/jpeg;base64,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"
      },
      "url": "https://cdn.sanity.io/images/ult5g8gw/production/0fcf0c59275cc7a96e89303f2ceece54ec3c224f-3840x2160.jpg"
    },
    "crop": null,
    "hotspot": null
  },
  "layout": "image",
  "markDefs": null,
  "secondImage": null,
  "video": null
}
```

### NeuralMesh Axon: High-Performance Storage With No New Hardware

NeuralMesh Axon is the GPU-native deployment of WEKA's NeuralMesh software platform. It fuses storage directly into a cluster's own GPU servers, turning NVMe drives that are already racked, powered, and paid for into a single high-performance data layer for training and inference. This provided Andromeda with a shared storage option for clusters that never had one, and a cost-effective solution for customers who wanted high performance without paying for dedicated infrastructure.

Joniaux summed up the shift in blunter terms: NeuralMesh Axon "changed the way we see GPU clusters," providing an additional shared storage option for clusters that meet the deployment requirements.

### Easy to Stand Up, Easier to Run

Kher's first WEKA deployment, done manually before the NeuralMesh Kubernetes Operator was implemented, took 10 minutes from bare metal to a running system. Automated through the operator today, that same setup extends across every provider Andromeda works with.

Day-to-day operations run largely on their own. Kher points to capabilities Andromeda didn't have to build or request: tiering, quotas, and multi-tenancy were already built into the platform. "Anything that I wanted, if I push it up there's already a solution there because they've thought about it," he said.

When something does need a human, WEKA's support closes the gap fast. Kher describes posting a support ticket and getting WEKA engineers on a call within minutes, at any hour: "It's a 24/7 team. I post a message and they get it." That responsiveness, he said, gives him the confidence to keep building on WEKA no matter which provider a customer's cluster runs on.

The Result

## 6x Faster Environment Startup

NeuralMesh and NeuralMesh Axon have sharply cut load and metadata times across Andromeda's fleet, with lower storage costs following as a result.

- **Environment startup times accelerated by 6x,** letting AI teams run more experiments every day.
- **Metadata operations cut from hours to minutes,** most notably for a biotech customer running roughly a billion files per directory.
- **First deployment stood up in 10 minutes,** now automated fleet-wide through the NeuralMesh Kubernetes Operator, sustaining over 400 GB per second of throughput and more than 6 million IOPS per cluster in production.
- **Meaningful reduction in storage costs,** savings Kher says translate into lower prices for the customers Andromeda serves.

> Responsiveness and quality of service is excellent with WEKA. Customers notice when an environment takes 30 seconds to load instead of three minutes. Same with metadata operations that used to take hours and now take minutes. We've also brought storage costs down, which means we can charge customers less. Those are the improvements that matter day to day.
>
> — Nicolas Joniaux, Head of AI Infrastructure • Andromeda

Looking Forward

## Bringing NeuralMesh Axon to More Clusters

Andromeda plans to expand support for NeuralMesh Axon in eligible environments, with deployment decisions guided by workload requirements and customer needs. Kher wants to build deep NeuralMesh expertise inside Andromeda itself, enough that the team resolves problems directly and files fewer tickets, then feeds what it learns back to WEKA's product team.

Andromeda brings together infrastructure across providers to meet customer workload requirements, with storage integration contributing to that experience.

[Download PDF](https://cdn.sanity.io/files/ult5g8gw/production/723bf94b59e386a4e4c0a31acde0ef68ec38bb2d.pdf)