GPU COMPUTE

Your next big idea.
Powered closer to home.

An Australian approach to AI compute: connect useful workloads with distributed GPU capacity and locally available solar and battery energy.

Explore compute demo Discuss access

Australian deployment focus Familiar API interface

hello_sunstack.py

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.sunstack.ai/v1",
    api_key=os.environ["SUNSTACK_API_KEY"],
)

response = client.chat.completions.create(
    model="qwen3.5-0.8b",
    messages=[{
        "role": "user",
        "content": "What will we build today?"
    }],
)

print(response.choices[0].message.content)

Integration example from the project prototype.Access and model availability are arranged with the team.

BUILT AROUND YOUR WORKLOAD

Useful compute.
Thoughtfully scheduled.

Workload fit comes first. Model size, latency needs, data requirements and energy flexibility all inform where and when a job should run.

Prototype demonstrated

From a prompt to a useful response.

The project prototype demonstrates streaming text and image-enabled inference through a familiar API interface. Talk to the team about models, access and service requirements.

Text generationImage understandingAPI integration

Local by design

A network designed around Australian homes and Australian execution. Deployment location and data handling are agreed as part of access.

Understand your usage

Bring job activity, compute usage and energy information into the same conversation, with metering built into the platform direction.

Use the energy window

Where your workload permits, schedule around available solar and battery energy instead of treating every hour as the same.

FIND THE RIGHT FIT

Different workloads. Different requirements.

Useful capacity starts with a realistic match. These considerations help scope a compute engagement; they are not a list of currently available service guarantees.

Workload	What matters	Scheduling fit
Interactive inference	Model memory, time to first token, response continuity	Requires an agreed availability and latency profile
Batch inference	Throughput, deadline, checkpointing and input size	A candidate for deferrable energy-aware execution
Model evaluation	Reproducibility, fixed model versions and comparable hardware	Often schedulable within a defined completion window
Fine-tuning	GPU memory, framework support and sustained runtime	Assessed against the model and available node hardware
Tightly coupled multi-GPU training	Interconnect, synchronisation, fault tolerance and topology	Requires a separate feasibility study; home nodes are not assumed to form one tightly coupled cluster

FROM REQUIREMENT TO RESULT

A clear path for your workload.

A proposed service workflow shaped around the job you actually need to run.

Define the job

Share the model, inputs, GPU memory needs, deadline and data-handling requirements.

Agree the environment

Confirm available hardware, execution location, dependencies and the intended service conditions.

Schedule & observe

Place eligible work on suitable capacity and follow its state and measured resource use.

Collect & review

Retrieve outputs, review usage and use the results to scope the next workload.

Know the conditions
before you run.

Distributed infrastructure makes the operating boundaries especially important.

Data handling

Agree what is uploaded, where it is processed, how long it is retained and how outputs are retrieved.

Availability

Clarify whether the job needs a continuous service or can pause, resume or wait for an energy window.

Execution assurance

Distinguish the prototype’s current controls from research into attestation and stronger isolation.

Usage & terms

Confirm how compute use is measured and the applicable commercial terms before running a real job.

COMPUTE, EXPLAINED

Questions before your first job?

LET’S BUILD WHAT’S NEXT

Tell us what you want to run.

Discuss your workload

Your next big idea.Powered closer to home.

Useful compute.Thoughtfully scheduled.