The whole as-a-service vocabulary is a question about where one company's duty of care ends and yours begins, and IaaS draws that line as low as it can be drawn while still being a rental. The provider keeps the buildings, the power, the physical machines, the hypervisors and the network fabric working. What reaches you is a catalogue of programmable parts: instances created and destroyed on demand, block devices you attach to them, private networks, firewall rules, snapshots, load balancers, all created and destroyed by API call rather than by conversation. Amazon EC2 with EBS volumes inside a VPC is the reference implementation of the idea; Google Compute Engine, Microsoft Azure virtual machines, Hetzner Cloud and DigitalOcean Droplets sell the same category with different catalogue breadth. Anything above that boundary is yours, and the boundary is contractual rather than technical. Providers publish shared responsibility models precisely so that an unpatched machine of yours is unambiguously your incident, never theirs.
The freedom is genuine, and so is the work it hands back. An API-provisioned estate can reproduce anything a data centre used to run, which is why organisations moving existing systems land here before anywhere else: the shapes are familiar, the migration is mostly a copy, and none of the software needs rewriting. What arrives with it is a permanent operations job. Somebody has to apply kernel updates and reboot in a sensible order, size the instances, decide what happens when one dies, take backups off the platform and then actually restore one to prove it works, rotate credentials, read the logs and answer the alert. Tooling such as Terraform or the provider's own templates makes the estate reproducible, which helps enormously and removes none of the duty. Read the rest of the cloud market as a long negotiation over exactly this burden. Managed databases, PaaS and serverless each take another slice of it away in exchange for less control, more lock-in and a higher price per unit of compute, and that bargain is usually a good one for anybody whose product is not infrastructure.
- Compute instances replace the tower in the cupboard: a virtual machine with chosen cores and memory, billed while it exists rather than when it is bought.
- Block storage replaces the disk: a volume attached to one instance, resizable, snapshot-able, surviving the instance it was mounted on.
- Object storage replaces the file share for anything the application only reads and writes wholesale, such as uploads, exports and static assets.
- Virtual networks replace the switch and the cabling, with subnets, routing and security groups written as configuration instead of plugged in by hand.
- Managed load balancers and floating addresses replace the appliance in front of the rack, and make a failed instance a routing change.
- Images and snapshots replace the imaging trolley, letting a rebuilt machine come back identical instead of approximately right.