Cloud hosting and IaaS explained
Technically a cloud instance is a virtual machine, much like a VPS. Commercially and operationally it is a different product: capacity you create and destroy through an API, billed by the hour or the second, surrounded by managed services you can rent instead of run.
- You get
- Virtual machines on demand, plus managed databases, storage, queues and load balancers
- Billing
- Metered — compute by the second or hour, storage and egress separately
- Strength
- Elasticity and breadth of services; infrastructure as code
- Trap
- Egress charges and idle managed services, neither of which announce themselves
Cloud versus VPS: the honest distinction
A VPS is a machine you rent by the month. A cloud instance is capacity you call into existence and throw away, where the unit of work is an API request rather than an order form. The hardware underneath is the same kind of hardware.
What you are buying is not better virtualisation. It is three things:
- Elasticity. Ten instances for an hour costs what one instance for ten hours costs, which makes bursty and batch workloads economically sane.
- Managed building blocks. A replicated database, object storage, a queue, a load balancer, a cache — rented rather than installed and patched.
- Reproducibility. Because everything is an API, the whole environment can live in version control and be rebuilt from nothing.
If you are running one long-lived instance that you built by hand, you are using a cloud provider as an expensive VPS. That is a common and sometimes sensible choice, but be clear that it is what you are doing.
The billing model is the part that bites
- Egress. Data leaving the platform is metered, often at a rate far above what bandwidth costs wholesale. It is the single most common cause of a shocking cloud bill, and it is invisible until a site gets popular or someone hotlinks a video.
- Storage outlives compute. Deleting an instance frequently leaves its volume, its snapshots and its reserved IP billing quietly for months.
- Managed services bill while idle. A provisioned database costs the same asleep as awake.
- Per-request services. Fine at low volume; worth modelling before a launch.
Regions, zones and what redundancy actually requires
A region is a geographic location; availability zones are independent failure domains within it. Two instances in one zone share a power and network fate. Spreading across zones is the cheapest real redundancy money can buy on these platforms, and it is a configuration choice rather than an extra product.
Pick the region nearest your users, not nearest you. Round-trip latency across an ocean is physics and no amount of caching hides it for dynamic requests.
Where the lock-in lives
Virtual machines are portable. The managed services around them are not: a proprietary datastore, an identity system or a serverless runtime is where migrations stall. This is a reasonable trade for the leverage those services give — just make it deliberately, and keep the data in formats you can export.
When cloud is the right answer
Variable or unpredictable load; teams that want infrastructure in code; projects that genuinely benefit from rented managed services; anything that must survive the loss of a datacentre; and early-stage work where nobody knows what the capacity requirement will be.
It is the wrong answer for a steady, well-understood workload on a tight budget. For that, a VPS or a dedicated machine is usually several times cheaper for the same throughput.