NVIDIA B300 GPUs for AI — ready to train, in Sydney.

Book 1 to 8 Blackwell Ultra GPUs by the week, start from PyTorch, NeMo, vLLM or plain CUDA, and SSH straight in. You pay for the GPU-hours you actually use — no egress fees, no storage add-ons.

Request access See pricing

GPU NVIDIA_B300_SXM6 · 288GB_HBM3E

INTERCONNECT NVLINK · FULL_MESH

BOOK 1–8_GPUS · BY_THE_WEEK

LOCATION EQUINIX_SY5 · SYDNEY

Pick your GPUs.

Every configuration runs on an NVIDIA DGX B300 — eight Blackwell Ultra GPUs on NVLink — so a single GPU and a full node get the same hardware. Book one GPU to try an idea, or all eight to train.

Configuration	GPU memory	Interconnect	Good for
1× B300	288 GB HBM3e	—	Fine-tuning, evaluation, a large model for inference
2× B300	576 GB HBM3e	NVLink	Distillation, 70B-class LoRA fine-tunes
4× B300	1.15 TB HBM3e	NVLink	Mid-size training runs, high-throughput serving
8× B300 — whole node	2.3 TB HBM3e	NVLink, all eight GPUs	Full training runs, frontier-scale inference

Whole node: 2× Intel Xeon 6776P, 2 TB DDR5, local NVMe, 8× ConnectX-8 SuperNIC. Long-term dedicated nodes with your own operating system are also available — ask us.

Start from a ready-made stack.

Choose a starting image when you book and it's running on your GPUs when your booking starts — or bring your own container. Every booking includes the platform features below; the labelled ones are on the way.

pytorch Included

NGC PyTorch

NVIDIA's PyTorch container with the Blackwell-tuned NCCL. torchrun across your GPUs out of the box.

nemo Included

NVIDIA NeMo

NeMo framework for large-model training, fine-tuning and distillation pipelines.

vllm Included

vLLM inference server

An OpenAI-compatible endpoint on your own GPUs — serve open models like Qwen, Llama or DeepSeek privately.

cuda Included

CUDA base

A clean CUDA 12 development image. Bring your own stack, or any container image you like.

storage Included

Encrypted storage + erasure certificate

Your volume is encrypted at rest (LUKS2) and cryptographically erased when you hand back — with a certificate you can verify.

api Included

Console, API and API keys

Book, start containers, read usage and download invoices from the console or the API. Full-access or read-only keys.

sandbox Early access

Agent sandboxes

Isolated, disposable environments for running AI agents' code, with secrets the agent never sees. Ask to join the early-access program.

inference Coming soon

Managed inference endpoints

Pay-per-token APIs for open models, hosted in Sydney. Until then, the vLLM image gives you the same on your own GPUs.

confidential Coming soon

Confidential computing

Intel TDX with GPU confidential-computing mode, so memory is encrypted even from us. Being commissioned on this hardware — not enabled yet.

From sign-up to `nvidia-smi`.

No sales call needed for a first booking. A Terrabox engineer confirms new customers' first booking, usually within one business day.

01 · account

Create an account

Email, password and an authenticator app. Add a card — you're only charged for GPU-hours you use.

02 · book

Choose GPUs, dates and a stack

Pick 1–8 GPUs and the weeks you need. The calendar shows what's free; add your team's SSH keys.

03 · run

SSH in and train

Your GPUs, your encrypted volume, your containers. Hand back early and you pay less.

Simple pricing.

Per GPU-hour, metered from the time your GPUs are actually allocated, invoiced monthly in arrears in Australian dollars. Power, networking, storage and support are included — no egress charges.

by_the_week

Weekly bookings

Book 1–8 GPUs for one week or more. Billed per GPU-hour used, by card or invoice.

Rate: Per GPU-hour, on request
Minimum: One week

committed

Committed capacity

Reserve GPUs across a quarter for a lower rate and guaranteed weeks.

Rate: Discounted per GPU-hour
Term: Quarterly

dedicated

Dedicated node

A whole DGX B300 for months or years, bare metal with your own operating system and network.

Rate: Monthly, fixed
Term: 12 months +

Get a quote

Book from the console — or from code.

Everything in the console is an API call. Create a key under Environment → API keys, then book GPUs, start containers and pull usage from your scripts or CI.

auth: bearer API key, full or read-only sign-in: password + authenticator app billing: per GPU-hour, card or invoice

# book 4 B300s for a week, starting from PyTorch
curl -X POST https://console.terrabox.ai/api/v1/blocks \
  -H "Authorization: Bearer $TERRABOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"gpus": 4, "start_date": "2026-11-16",
       "days": 7, "template": "pytorch"}'

# when it starts: your GPUs, your encrypted volume
ssh -p 2222 you@ssh.terrabox.ai
nvidia-smi --query-gpu=index,name,memory.total --format=csv
torchrun --nproc_per_node=4 train.py

Security you can check, stated plainly.

Your data stays in Sydney, on hardware we own and run. Volumes are encrypted at rest and crypto-erased on hand-back, with a certificate. On shared nodes, customers get their own GPUs, volume and login and are separated by containers — not separate VMs. Confidential computing is being commissioned, not enabled, and we're engineered toward IRAP and APRA CPS 234 but not certified. We'll always tell you exactly what's on.

terrabox · syd Equinix SY5

DATA_LOCATIONSYDNEY, AUSTRALIA

TENANT_VOLUMELUKS2 AES-XTS

HAND_BACKCRYPTO-ERASE + CERTIFICATE

SIGN_INPASSWORD + 2FA

CONFIDENTIAL_TIERCOMMISSIONING · NOT ENABLED

$ ./scripts/tbxcert.py verify TBX-CE-2026-0001