Governance & Risk

Governance is treated as first-class: the node is a shared, multi-tenant resource, so explicit rules are what let it serve many use cases plus open startup and researcher access fairly and safely. This page describes how workloads, data, and risk are managed on the Hub.

Workload governance framework

Workloads on the shared node are managed through a workload governance framework – the set of rules that determines who gets GPU time, when, and at what priority, so that many tenants can run side by side without contention.

The framework defines:

• Quota policies – how much GPU time, memory, and storage each tenant may hold.
• Priority tiers – which workloads take precedence, and which are protected from preemption.
• Scheduling rules – how inference and training windows are placed so they do not collide.

A GPU workload calendar operationalises this: perpetual inference services run continuously on time-sliced partitions, while training and fine-tuning are scheduled into off-peak windows so live services are never degraded by heavy batch jobs.

Data governance

Data is governed alongside compute. For each workload the Hub manages:

• Suitability – whether the data behind a use case is available, curated, and representative enough to train and serve a reliable model.
• Legal basis & consent – use under the Nigeria Data Protection Act (NDPA) and the terms the data was collected under, especially for health and financial records.
• Retention – how long data is held and when it is cleaned up, rather than keeping it indefinitely.

Because several use cases are constrained more by data than by compute, data pipelines and governance – not GPU procurement – set the pace for those workloads. See Data Readiness.

Access control & sovereignty

• Identity & access – authenticated, role-based access, so a tenant reaches only its own workloads and data.
• Data residency – proprietary data stays on national infrastructure via on-node storage rather than leaving the country.
• Auditability – access and usage are logged, so it is always possible to answer who ran what and who accessed which data.

These are the same controls the Hub expects of the systems it hosts – see Securing AI-Powered Solutions and AI Governance.

Managing risk: pre-deployment gates

Risk is managed by gating deployment on evidence, not by hoping for the best. Before a workload serves real users the Hub applies proportionate gates:

• Safety-critical workloads – clinical and other high-stakes use cases must pass alignment and safety evaluation before deployment; a wrong answer here can cause real harm, so this is non-negotiable. See Responsible AI.
• Quality bar – models are evaluated against task-specific and local-language benchmarks before release. See Model Lifecycle Support.
• Phased rollout – workloads start on a pilot footing and scale as evidence and controls mature, rather than launching at full production immediately.