Provision a GPU. Start working in minutes.
Pick hardware, attach storage, launch a Linux instance, SSH in. Same loop as the major clouds, hosted in Newcastle, billed by the minute in Australian dollars.
From signup to first job in under ten minutes
Create an account
Email and password for now; SSO and SCIM are on the roadmap. You'll add a payment method before launching paid instances, but verification is free.
Launch an instance
Pick a hardware profile and image (Ubuntu 24.04 with CUDA 12 by default; bring your own image if you prefer). Your public SSH key gets baked in at boot.
Connect and work
SSH in with the address shown in the dashboard. Persistent storage is mounted at /data. Usage shows up in the cost panel within a minute.
What's in the rack
| Card | VRAM | Interconnect | TDP | Best for |
|---|---|---|---|---|
| B200 SXM6 Frontier | 192 GB HBM3e | NVLink 5 (1.8 TB/s) | 1000 W | Frontier training, FP4 inference for 100B+ models |
| H200 SXM5 | 141 GB HBM3e | NVLink (900 GB/s) | 700 W | Mainstream training and long-context inference (70B+) |
| H100 SXM5 | 80 GB HBM3 | NVLink (900 GB/s) | 700 W | Cost-effective training and batch inference |
| RTX PRO 6000 Blackwell Server Edition | 96 GB GDDR7 | PCIe Gen 5 | 600 W | Inference workhorse with FP4 — 6× faster than L40S |
| AMD Instinct MI325X On request | 256 GB HBM3e | Infinity Fabric (896 GB/s) | 1000 W | Memory-heavy inference on ROCm, vLLM, SGLang |
Network
Dedicated 25 Gbps internal fabric between nodes inside the facility. 10 Gbps public uplink with carrier redundancy via two distinct providers. Internal latency under 0.2 ms.
Power and cooling
N+1 power redundancy. Direct-to-chip liquid cooling on B200 and H200 trays; rear-door heat exchangers across the rest of the floor. Designed for AI workload power density from day one — racks happily run at 50 kW+.
Provision via HTTP or CLI
Everything you can do in the dashboard is available via API. Same authentication, same primitives.
Launch an instance (curl)
curl https://api.compute.newcastlerising.com.au/v1/instances \
-H "Authorization: Bearer $NC_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"hardware": "h200-141gb",
"count": 1,
"image": "ubuntu-24.04-cuda-12",
"ssh_keys": ["ssh-ed25519 AAAA..."]
}'Python
from newcastle_compute import Client
nc = Client(api_key="...")
instance = nc.instances.create(
hardware="h200-141gb",
count=1,
image="ubuntu-24.04-cuda-12",
)
print(instance.ssh_command)
SDKs for Python and Go are published as newcastle-compute. A static CLI binary is available for macOS, Linux, and Windows. Full reference on the docs page.
Things people ask first
How do I know how much I'm spending?
Usage shows up in the dashboard within roughly a minute of running. We also expose a /v1/usage endpoint that returns the current month's spend per instance, so you can stop yourself before you scare yourself.
Can I reserve capacity?
Yes. 30-day reservation: 25% discount. 12-month reservation: 40% discount. We hold the specific GPU for you and you can use it (or not) as you like during the term.
What happens if there's an outage?
Live status on the status page. SLA credits apply automatically once we miss 99.9% in a calendar month. We post a public post-mortem within 5 business days of anything serious.
Can I mount persistent storage?
Yes — NVMe-backed volumes attach to instances, survive shutdown, and bill at AUD $0.12 per GB-month. Local scratch space on each node is included.
Bring my own image?
Yes. Custom QCOW2 / VHD images can be uploaded. We also support standard Docker images via --container at launch time.
What about egress?
Traffic inside Australia is included. Outbound to other regions is metered separately at AUD $0.06/GB. There's no charge to bring data in.