Rent the GPUs, skip the build
Bare-metal clusters and elastic cloud instances covering pretraining, fine-tuning, inference and scientific computing. Scale from one GPU to a thousand, billed by the hour or on a monthly plan.
Pick the card, we handle the cluster
All prices in NZD, per GPU per hour, excluding GST. Monthly and annual commitments get a lower rate — ask us for a quote.
H100 SXM5
Built for thousand-GPU pretraining. RDMA fabric with sub-2μs inter-node latency.
- Bare-metal, no virtualisation overhead
- NVLink + NVSwitch in-node fabric
- 200Gbps InfiniBand / RoCE
- NVIDIA AI Enterprise included
H800 / A800 Cluster
Fine-tuning and mid-size training, scaling weekly from 8 to 512 GPUs.
- Scale from 8 to 512 GPUs
- Weekly elastic scaling
- Self-healing scheduler
- Checkpointed job recovery
L40S Inference
Cost-efficient serving for LLM inference, rendering and video pipelines.
- Ideal for 7B–70B inference
- Minute-level provisioning
- Load-balanced inference gateway
- Per-second billing
RTX 4090 Elastic
Spin up a single card in under a minute for prototyping and small fine-tunes.
- Single-GPU to 8-GPU instances
- Billed by the second
- Jupyter / SSH ready images
- Best for prototypes and demos
Node-level comparison
Every node ships with dedicated physical resources — no noisy neighbours, no virtualisation tax.
| Configuration | GPU | Memory | CPU | System RAM | Local NVMe | Interconnect | NZD / GPU / hr |
|---|---|---|---|---|---|---|---|
| H100 SXM5Flagship training | NVIDIA H100 80GB SXM5 | 640GB HBM3 per node | 2 × Intel Xeon Platinum 8480+ | 2TB DDR5 | 30TB NVMe local · parallel file system | NVLink 900GB/s · 200Gbps InfiniBand | $4.60 |
| H800 / A800 ClusterElastic scale-out | NVIDIA H800 80GB / A800 80GB | 640GB HBM2e per node | 2 × AMD EPYC 7763 | 1TB DDR4 | 15TB NVMe local · 20GB/s parallel read | NVLink 400GB/s · 200Gbps RoCE | $3.20 |
| L40S InferenceProduction inference | NVIDIA L40S 48GB | 192GB GDDR6 per node | 2 × Intel Xeon Gold 6430 | 512GB DDR5 | 7.6TB NVMe local | PCIe Gen4 · 100Gbps RoCE | $1.90 |
| RTX 4090 ElasticStart small | NVIDIA RTX 4090 24GB | 24GB GDDR6X per GPU | AMD EPYC 7443 | 256GB DDR4 | 2TB NVMe local | PCIe Gen4 · 25Gbps network | $0.65 |
Pricing is indicative and excludes GST. Annual and multi-node commitments are quoted separately. Enterprise customers can bring their own licence.
Everything around the silicon
Bare-metal clusters
Dedicated physical resources with no virtualisation overhead — the shape that thousand-GPU pretraining actually needs.
RDMA high-speed fabric
200Gbps InfiniBand or RoCE with sub-2μs inter-node latency across every multi-GPU job.
Minute-level provisioning
Elastic instances are billed by the second — the right shape for fine-tunes and short-lived inference.
Direct-attached storage
A parallel file system plus cloud disk delivers up to 20GB/s of training-data read throughput.
Isolated training network
Training traffic stays on a network physically separated from general business and internet traffic.
Monitoring and checkpoints
Node health, GPU utilisation and job checkpoints surfaced in one dashboard, with automatic failure recovery.
From enquiry to running job
Scoping call
Tell us the model size, framework and timeline. We size the cluster with you.
Quote in 24 hours
A written proposal with node specs, hourly rate and commitment options.
Provisioning
Bare-metal clusters in around five minutes; elastic instances in under a minute.
Run and scale
SSH or Jupyter access, then scale nodes up or down as the job demands.
How many GPUs do you need?
Send us your workload and we'll come back with a sized cluster, a rate and a provisioning date.
Get a quote