Cloud Computing Concepts, Service Models & Hardware Sizing

A model served from a laptop or a single on-prem server inherits all its weaknesses: power cuts, a flaky uplink, no room to grow, and a long wait to buy new hardware. Cloud computing rents compute, storage and networking on demand, billed by usage.

The NIST definition has five essential characteristics — learn them by name:

1. On-demand self-service — provision resources yourself, no human in the loop.
2. Broad network access — available over the network from standard clients.
3. Resource pooling — the provider serves many tenants from shared hardware.
4. Rapid elasticity — scale out and in quickly, even automatically.
5. Measured service — usage is metered and billed.

Moving down the table you give up control for convenience. Our StudyRoom backend is a PaaS deployment (Render); the hospital might choose IaaS in a nearby region for data-residency reasons.

Match hardware to the model, not to the hype:

• Tabular models (logistic regression, RandomForest, XGBoost) do little arithmetic per prediction. A small CPU instance serves them in ~10–25 ms for pennies.
• Large neural networks (vision transformers, LLMs) do billions of multiply-adds; GPUs do these in parallel and are worth their price — but only when busy. An idle GPU still bills by the hour, so GPU services should auto-scale to zero when quiet.

Latency has two parts: compute (how long the model takes) and network (distance to the users). For a Ugandan hospital, a server in Johannesburg adds ~40 ms per round trip; one in Ireland ~135 ms. Often the region matters more than the instance size.

Providers sell spare capacity as Spot (AWS), Preemptible/Spot (Google) VMs at 70–90% off — with a catch: they can be reclaimed with about two minutes' notice.

• Great for training and batch scoring: save checkpoints regularly, and if the VM disappears, resume from the last checkpoint on a new one.
• Wrong for real-time clinical serving: a reclaimed instance means triage stops answering. Serve on on-demand (or reserved) instances, possibly with Spot capacity added behind a load balancer only for overflow.

Cost arithmetic: spot rate = on-demand × (1 − discount); job cost = rate × hours.