Representative system image
File storage
Storage that behaves like a shared drive. The commonest starting point, and what most training frameworks expect to read from.
Scale-out enterprise file platforms, on-premises or consumed as a service.
AI infrastructure
AI storage
Accelerators are the expensive part; storage is what stops them being wasted. Under-specified storage is the commonest reason an AI cluster underperforms its brochure figures.
Start here
Why does AI need storage as well as GPUs?
Three common situations, and what each one actually needs.
A GPU can only work on data that has already reached it. If the storage behind the cluster cannot keep up, the accelerators stall — and an idle accelerator costs exactly the same as a busy one. This is not a theoretical concern: it is the single most common disappointment in a first AI deployment.
The practical rule. Storage is sized from three numbers: how much data you hold, how fast the GPUs consume it, and how much you write back. Nobody can size it from the GPU count alone, and anyone who offers to is guessing.
Private AI assistant
- Your documents
- Retrieval store
- GPU inference
Training
- Training data
- High-throughput tier
- GPU cluster
- Checkpoints
Retrieval at scale
- Object store
- Index & embeddings
- Serving layer
Definitions for every term are in the glossary.
Categories
What we source
Representative system image
Object storage
Storage designed for enormous numbers of items rather than a tidy folder structure. Where large training corpora and document sets usually end up.
S3-compatible scale-out object platforms with AI pipeline acceleration.
Representative system image
Block storage
The traditional enterprise storage your databases and virtual machines already use. Relevant because AI rarely arrives on a blank site.
Disaggregated scale-out block platforms with mission-critical resiliency.
Representative system image
Scale-out platforms
Storage you extend by adding units rather than replacing the whole thing. Important because AI data volumes are almost never estimated correctly first time.
Modular architectures where capacity and throughput grow independently.
Representative system image
Dataset storage
The working copy of your training data, sized so the accelerators are never left waiting for the next batch.
High-throughput tiers matched to GPU count and job profile.
Representative system image
Model repositories
Somewhere to keep the model weights themselves — versioned, so you can go back to the one that worked.
Versioned artefact storage with access control.
Representative system image
RAG and vector data
The document store behind a private AI assistant. Usually far smaller than people expect, but read constantly.
Low-latency storage for retrieval-augmented generation workloads.
Representative system image
Checkpoint storage
Training writes its progress to disk periodically so a failure does not cost you a fortnight. Those writes are large, sudden and unforgiving.
Burst-capable write tiers sized to model size and checkpoint interval.
Representative system image
Storage services
Sizing, migration and integration — working out what you actually have before deciding what to buy.
Creative Compute data platform assessment and migration support.
Tell us what your data actually looks like
How much, what kind, and what you intend to run against it. We will size it properly rather than selling you a tier you will not use.