InferaGrid
A sovereign distributed AI infrastructure network for Australia. It evaluates compute, power, network and location policy before scheduling work to an eligible GPU site.
Distributed Infrastructure, Coordinated as One AI Grid
Resources publish compute, power, network and site state. Workloads enter through the Gateway; InferaGrid selects one eligible execution site and records the decision context.
Select the eligible execution site
What InferaGrid Is
One coordination layer across distributed compute, energy, network infrastructure and qualified sovereign sites.
One reference node first
A repeatable reference architecture, validated on the bench and in one node before any site class is scaled.
One site per workload
Each job stays within one site. The control plane coordinates capacity and telemetry across sites.
Nodes are not facilities
A node may sit at an edge, colocation or distributed site; the node itself is not a Tier-rated facility.
Scenarios, not forecasts
Large deployment numbers describe possible future scale. They are not current capacity or commitments.
Workloads Follow the Data
Sensitive data stays in controlled environments. InferaGrid is for work that policy allows at distributed sites.
Your own environment
Use for identified personal, clinical or commercial data that must remain with you.
A certified colocation facility
Use for sensitive data needing certified physical controls. InferaStack designs dedicated GPU environments for NEXTDC facilities.
InferaGrid
Use for encrypted, non-identified inference and bounded batch jobs when policy allows. Distributed training is out of scope.
What a Node Will Offer
Reserved capacity for AI workloads that run all day — designed, not yet available.
Fixed-budget model service
Reserve capacity on a pre-deployed open model for always-on agents. Choose dedicated GPUs or a shared endpoint with a reserved share.
- OpenAI-compatible API through the InferaStack Gateway
- Envelope state and region of execution recorded per request
- Last-resort drain to hosted capacity when a node is curtailed
GPU VMs and managed workers
Run isolated GPU VMs, larger two-GPU workers, or flexible spot and batch jobs.
- Shared partitions or whole-GPU capacity
- Spot and batch are the curtailable classes when the envelope contracts
- No cross-site pooling — capacity is sold per site
How a Node Behaves
Control priority is fixed by design. The compute layer sits at the bottom of it.
Device protection
Battery management, power conversion, electrical protection and server thermal protection.
The owner's rules
The site owner's backup reserve and operating limits set by the site owner.
Available power
Live load, solar, storage and grid limits define the power available to compute. The ConnectVPP interface is still in technical validation.
Workload control
Reserved work comes first. Spot and batch yield first, then GPU power is reduced, and work drains only as a last resort.
Energy Integration Partner
ConnectVPP is InferaGrid's energy integration partner for the proposed interface between site energy controls and compute scheduling.
Power-aware compute orchestration
The partnership connects distributed-energy telemetry with InferaGrid's workload scheduler while preserving device protection, owner policy and backup reserve.
Live telemetry in
Site load, solar, battery state and grid limits define the power available to compute.
Owner control retained
Storage remains the owner's asset and available to its VPP operator when compute demand is low.
Status: the interface remains in technical validation and is not deployed. Gate 0 tests the control logic against a simulated envelope. See the energy reference architecture →
Reference Configuration
What Stage 0 proposes to validate. No node has been validated on hardware.
View the proposed Stage 0 configuration
| Stage | Configuration | What is measured | Status |
|---|---|---|---|
| Stage 0a · bench | 4× NVIDIA RTX PRO 6000 Blackwell Server Edition in a controlled bench environment | Inference, GPU partitioning, peer-to-peer links, telemetry, simulated power-envelope control loop | Proposed · Gate 0 pending |
| Stage 0b · reference node | One OEM-supported 8-GPU RTX PRO Server configuration | 72-hour measured workload; whole-node power 7.5–12 kW including cooling; thermal, acoustic, network, recovery | Candidate · OEM review required |
| Site classes | Colocation and data centre · enterprise and private AI · regional edge · qualified distributed energy sites | Qualification checklist, permitted service classes and data-handling default per class; residential remains a research pathway | Defined · first field site class chosen on Stage 0 evidence |
Figures are engineering targets and vendor specifications, not measurements. Whole-node power includes cooling and auxiliaries; GPU board power alone is lower. Scale figures elsewhere are long-term scenarios, not forecasts.
The Gated Pathway
Progress by acceptance results and paid reserved demand, not by calendar.
Bench
Stage 0a: four GPUs on a controlled bench. Inference, GPU partitioning, peer-to-peer links, telemetry and admission, drain, fallback and recovery against a simulated power envelope.
One reference node
Stage 0b: one OEM-supported 8-GPU node runs a 72-hour measured workload. Whole-node power, thermal, acoustic, network, topology and recovery evidence is accepted.
First qualified site
One separately approved field site, only after electrical, cooling, network, security, service, data-policy and energy-control acceptance.
Repeat, then decide
A small set of qualified sites must show consistent discovery, placement, drain and recovery. Broader deployment is a commercial decision taken on measured evidence and contracted demand.
Where Things Stand
Tested versus proposed, as of September 2026.
View current technical status
| Status | What |
|---|---|
| Development deployment | The Gateway runs in InferaStack's own AWS account and is not released. |
| Unit-tested software | Node invocation, envelope admission control and drain logic have not run on hardware. |
| Design only | Reference architecture and site classes defined; no site, acoustic, thermal or electrical acceptance yet. |
Choose Your InferaGrid Path
Both pathways are expressions of interest during technical validation. We will explain the current gate before discussing commitments.