Cloud providers rent compute, storage and managed services on demand.
Compute runs your code, from virtual machines to serverless functions.
Object storage such as S3 holds datasets and model artefacts.
GPU instances are available by the hour for training and inference.
Watching cost is a core skill, since GPU bills grow alarmingly fast.
Cloud providers rent compute, storage and managed services by the minute.
The appeal is starting without buying hardware and stopping without owning it.
Compute comes in layers, from virtual machines to containers to serverless functions.
A virtual machine gives control and leaves patching and scaling to you.
Serverless scales to zero and suits spiky low-volume work.
Containers on a managed platform are the usual middle ground for model serving.
Object storage such as S3 holds datasets and model artefacts cheaply and durably.
GPU instances are available by the hour, which is how most training gets done.
They are expensive, and an idle GPU costs the same as a busy one.
That is where cloud bills go wrong, since a forgotten instance runs all weekend.
Spot or preemptible instances cost far less and can be taken away mid-run.
Checkpoint frequently and they become a large saving rather than a risk.
Egress charges apply to data leaving the provider and surprise people regularly.
Set budget alerts on day one, before anything is deployed.
Managed services cost more per unit and less in time, which is usually the better trade early on.
One question on this card. Every option explains itself once you pick, right or wrong.
22 concepts in this module, then a graded test with every answer explained.