Edge, IoT & Compute · Module
One3 Compute
Budgeted, auditable GPU jobs with hard caps and cost per job
One3 Compute is planned to give tenants GPU compute for fine-tunes, batch inference, evaluations and video renders without separate cloud accounts. Quotas are hard caps and every job's cost posts to AI FinOps and the ROI Ledger.
What it does
One3 Compute accepts GPU jobs from templates: adapter fine-tunes from AI Lab, batch inference and embeddings, Evaluation Studio runs and Video Factory renders, with Data Science Workbench, Forecast & Simulation and Business Twin runs next. It places each job on a GPU pool that matches the tenant's data residency and classification, enforces quotas as hard caps, registers outputs with lineage, and posts the cost.
Spend that cannot run away
Quotas apply per tenant, workspace, project or agency sub-account: GPU-time, concurrent jobs and a monthly spend cap. Before submission the job gets a cost and GPU-time estimate with an upper bound. If the upper estimate exceeds the remaining quota, the job does not start, and a quota raise is an approval routed by delegation of authority with requester and approver different. No tenant setting can turn a quota into a soft cap. Unusual spend rates are detected and non-critical jobs can auto-pause by policy.
Residency and isolation
GPU pools carry residency tags, classes, prices and availability across cloud providers, One3 reserved capacity and the tenant's own One3 Box GPUs. Jobs that read restricted or sensitive personal data run only in pools that match tenant residency, or on tenant boxes. Each job runs in its own sandboxed container with ephemeral scoped credentials, no shared volumes across tenants, default-deny egress and wiped scratch storage. Training datasets must pass consent and PII-scrub checks.
Lineage and chargeback
Outputs are registered as files and, for adapters, in the Model Hub registry with dataset versions, template, image digest and job ID. JobCost (GPU-time, storage, egress and pass-through provider cost) posts to AI FinOps, the Investment & ROI Ledger and usage metering. Chargeback reports cost by project, cost centre and agency client for rebilling. Preemptible capacity uses checkpoints, and batch windows can move non-urgent jobs into lower-cost periods within residency limits.
Features
What One3 Compute includes
First-party job types
AI Lab fine-tunes, batch inference and embeddings, evaluation runs and Video Factory renders from those modules or the API.
Job templates
Resource profiles, digest-pinned container images, parameter schema and checkpoint policy.
Quotas and budgets
GPU-time, concurrency and monthly spend caps with a hard stop and DoA-routed raise requests.
Cost estimate before submit
Estimated GPU-time and cost range; submission blocked when the upper estimate exceeds remaining quota.
JobCost posting
Actual GPU-time, storage, egress and provider cost to AI FinOps, the ROI Ledger and usage metering.
GPU pools
Pools per provider and region with residency tags, GPU classes, price and availability (providers to verify).
Residency and classification routing
Restricted or sensitive data runs only in residency-matching pools or on tenant boxes.
Scheduler
Priorities, fair share, per-tenant concurrency and preemptible capacity with checkpoint and resume.
Isolation
One sandboxed container per job, ephemeral credentials, default-deny egress and wiped scratch.
Artifacts and lineage
Outputs registered with input dataset versions, template, image digest and job ID.
Job monitoring
Status, logs, GPU utilisation, progress, cancel and notifications.
Dataset access gate
Only project-granted datasets; consent and PII-scrub checks for training data.
Spend anomaly, batch windows and failover
Auto-pause of non-critical jobs, cheaper scheduling windows and provider failover within residency.
Chargeback and tenant box GPUs
Cost by project, cost centre and agency client; One3 Box GPUs registered as a private pool.
Workflow
Job lifecycle with quota and residency
- 1
Submit from a template
- 2
Dataset access and consent checks
- 3
Estimate cost and GPU-time
- 4
Block and request a quota raise if the upper estimate exceeds remaining quota
- 5
Residency and classification policy selects an eligible pool
- 6
Run in a sandbox with checkpoints; stop at the hard cap
- 7
Register artifacts with lineage and post JobCost
AI
AI skills and Digital Employees
Cost and duration estimate
Estimates from template, dataset size, GPU class and history with an upper bound; the rule enforces it.
Job configuration assistance
Template choice, starting hyperparameters, checkpoint interval and residency-compatible pool.
Failure triage
Classifies out-of-memory, data error, driver mismatch, preemption and quota stops and proposes the smallest fix.
Chargeback narrative
Explains which team, agent, app and job drove the period's GPU spend.
Digital Employees that use this module
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
How is One3 Compute priced?
It is metered against a quota with hard caps. The candidate unit is GPU-time by GPU class; the unit and margin are set by pricing tests. No prices are in the plan.
Can a job overspend its budget?
No. A job whose upper cost estimate exceeds the remaining quota does not start, and a running job stops at the hard cap with a checkpoint.
Will our data leave India?
Jobs that read restricted or sensitive personal data run only in pools matching your residency, or on your own One3 Box GPUs.
Can agents submit GPU jobs?
Agents can draft template jobs within a granted project quota; a human approves submission.
See One3 Compute in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.