Use-Case Feasibility Assessment

Not every AI idea belongs on shared compute, and the ones that do have very different demands. Before a use case is admitted, the Hub runs a feasibility assessment: teams arrive with a problem and a target population, and the Hub determines what model, what memory footprint, what GPU intensity, and what scheduling pattern make the workload viable – or whether the real constraint lies somewhere other than compute entirely. This is the first capability a use case meets, and it is what keeps the node from being over-committed.

Why it matters on a shared node

On dedicated hardware, feasibility is a private question – a team either has enough GPU or it does not. On a shared, multi-tenant node it is a collective one: admitting a workload that needs two GPUs resident around the clock removes that capacity from everyone else. The assessment therefore does not just ask "can this run?" but "what does this cost the shared pool, and is that cost justified by the use case's impact?"

What the assessment covers

• Model fit – which open-source model family suits the task, and at what parameter scale. The identified use cases span everything from lightweight forecasting models to sub-8B and 7–13B language models (cited up to ~33B for the most demanding).
• Memory footprint – VRAM required at FP16 versus quantised INT8, and whether the model fits alongside other tenants in the shared 1,128 GB of HBM3e. Quantisation is often the difference between a comfortable fit and a tight one.
• GPU intensity – low / moderate / high, and whether the workload is inference-only, training-heavy, or hybrid. This determines whether it can be time-sliced or needs scheduled training windows.
• Concurrency – realistic concurrent-session estimates rather than optimistic per-user averages. An earlier "2 sessions/user/month" figure was corrected because it understated real demand; planning now uses realistic concurrency (e.g. 500–1,000 concurrent sessions for agricultural advisory at pilot scale).
• Binding constraint – crucially, whether the true limit is compute, data, networking, or CPU / RAM. For several use cases it is not the GPU (see Data Readiness).

Worked results

The seven identified use cases were each put through this assessment. The output per use case – sector, model profile, VRAM, GPU intensity, and the Hub's specific role – is documented in Identified Use Cases, and the shared envelope they were assessed against is in Capacity. The consolidated finding: all seven fit in pilot configuration simultaneously, but full production across all at once would approach the node's ceiling, so rollout is phased.