Compute as a Service

We run GPU clusters so you can train, serve, fine-tune and localize AI models, without becoming an infrastructure company first.

Explore the fleet

500

GPU servers in operation

4000+

Accelerators under management

GPUs per node

2019

Operating GPUs since

Four workloads, one operations team

The hardware is table stakes. What decides your cost per token is engine tuning, cache behaviour, fabric configuration and who picks up the pager at 3am.

AI Training
Multi-node training that stays up for weeks
Dense GPU nodes with high-bandwidth interconnect, wired for distributed training runs that cannot afford to restart from scratch.
AI Inference
Serving capacity behind an OpenAI-compatible endpoint
Production inference on vLLM, SGLang or TensorRT-LLM, metered per token and reachable through a single API surface.
Fine-tuning
From base model to a checkpoint you can actually serve
LoRA, QLoRA and full-parameter fine-tuning, followed by the quantization and evaluation work that makes the result deployable.
On-premise
Your models, your hardware, your network boundary
Localized deployment for teams whose data cannot leave the building, including non-NVIDIA accelerators.

What comes out of our clusters

Not a capability deck. Running systems and published artifacts you can inspect today.

You keep the software

LLMFabric runs our clusters, and it is the same Apache 2.0 software we install when a deployment has to live inside your own network. You can read it before you buy anything, and if you ever stop working with us the cluster keeps running without us in it.

Read the LLMFabric source

Runs on what you have
Five accelerator families, not an NVIDIA-only stack. That is what decides whether an on-premise estate or a domestic-hardware mandate is even possible.
- NVIDIA GPU
- AMD GPU
- Ascend NPU
- Hygon DCU
- MThreads GPU
Runs your serving stack
Engine choice and parameter tuning are automated per model and traffic shape, so the handover is a working configuration rather than a quickstart to copy.
- vLLM
- SGLang
- TensorRT-LLM
- Custom engines

Six years of running GPUs at scale

First place in the Filecoin Space Race, second in the Aleo incentivized testnet, both decided purely on sustained throughput. That operational practice is what moved to AI workloads.

2019to2026

Read the full timeline

Tell us what you are trying to run

Send the model, the concurrency target and the latency you need. We will come back with a sizing and a number, not a discovery call.

[email protected] Read our code

What are you trying to run?

Select all that apply.

Pick one or more above and we will prefill the enquiry for you.

Compute as a Service

Four workloads, one operations team

AI Training

AI Inference

Fine-tuning

On-premise

What comes out of our clusters

Quantized model releases

Checkpoints you can actually serve

Model inference

OpenModel

You keep the software

Runs on what you have

Runs your serving stack

Six years of running GPUs at scale

Tell us what you are trying to run