Meridian Compute
Compute
High-performance compute,
cloud simplicity
Run and scale AI workloads on a platform that gives you the flexibility of the cloud and the performance of a supercomputer - for every stage of your AI pipeline.
Bare-metal level performance
Get maximum GPU utilization from your clusters. Instances don't virtualize GPUs or network interfaces, delivering performance on par with the best industry benchmarks.
Cloud flexibility
Spin up a single VM or scale to a multi-node cluster with the simplicity you're used to. Launch on-demand and preemptible instances, scale up or down, or replace a node - all via CLI or API.
Massive scale
Connect thousands of GPUs in a single cluster over optimized, non-blocking NVIDIA Quantum-2 InfiniBand fabric - built for frontier-scale training runs.
GPU lineup
Accelerated compute, powered by NVIDIA
From single-node instances to thousand-GPU clusters - pick the accelerator that matches your workload.
Rack-scale · Blackwell Ultra
NVIDIA GB300 NVL72
- 72 GPUs + 36 Grace CPUs / rack
- Liquid-cooled
- Frontier training
A fully liquid-cooled, rack-scale system with 72 Blackwell Ultra GPUs and 36 Grace CPUs, purpose-built for frontier model training, AI reasoning, and agentic AI at the highest scale.
Talk to sales →Large-scale · Blackwell
NVIDIA HGX B200
- 180 GB HBM3e / GPU
- 8-GPU HGX nodes
- LLM training + MoE
Built on the Blackwell architecture - a well-balanced platform for large-scale LLM training, mixture-of-experts workloads, and high-throughput inference.
See pricing →Memory-intensive · Hopper
NVIDIA HGX H200
- 141 GB HBM3e / GPU
- 4.8 TB/s memory bandwidth
- Large-model inference
The Hopper generation with 141 GB of memory per GPU - run large language models without quantization and serve memory-intensive inference workloads at scale.
See pricing →Battle-tested · Hopper
NVIDIA HGX H100
- 80 GB HBM3 / GPU
- Mature ecosystem
- Cost-efficient at scale
The proven Hopper GPU for cost-efficient inference, fine-tuning, and large-scale training, with a mature software ecosystem and battle-tested reliability at cluster scale.
See pricing →CPU instances
CPU compute for the rest of your pipeline
CPU-only instances on Intel Xeon and AMD EPYC keep GPU capacity free for the work that needs it.
AI applications and agents
Run app backends, serving logic, and orchestration layers that sit alongside your GPU workloads.
Data preprocessing and pipelines
Run tokenization, feature engineering, and data loading jobs on dedicated CPU compute.
Offline and batch inference
Process documents, run bulk evaluations, and serve non-latency-critical inference on CPU.
Automation and tooling
Run evaluation harnesses, scheduled jobs, ML pipeline scripts, and CI/CD workflows for AI.
Orchestration
Fully managed Kubernetes, pre-optimized for AI
Deploy, scale, and manage containerized AI workloads natively, without building or maintaining the orchestration layer yourself. GPU-aware scheduling, InfiniBand-aware networking, and node health management come built in.
Need direct, DevOps-level control over multi-node environments? Managed Kubernetes is also available as a standalone service.
$ meridian k8s create --name train-cluster \
--gpu-type hgx-h200 --nodes 8
$ kubectl get nodes -L gpu
NAME GPU
node-0 8x H200
node-1 8x H200
...Proof
Validated performance, out of the box
Quickstart
From zero to a running GPU in two commands
$ meridian instances create --gpu hgx-h100 --image cuda-12.4
$ ssh dev@<instance-ip>
nvidia-smi # your GPU, readyGetting started
Launch your first GPU instance from the quickstart, or reach out to discuss capacity, reserved pricing, or your specific workload requirements.