Is every website a cloud service?

No. A website can run on many kinds of hosting. The delivery model and service characteristics determine whether the underlying arrangement is cloud computing.

Can a cloud application work offline?

Some applications retain local capabilities and synchronize later. That behavior belongs to the application; network access is still needed for its cloud interactions.

Are cloud computing and AI the same thing?

No. Cloud services can run ordinary software, and AI can run on local equipment. Cloud resources are one way to develop and deliver AI systems.

AI Glossary / MLOps and deployment / Beginner

Cloud computing

Definition

Cloud computing provides computing resources as network-accessible services that users can obtain and release on demand. Resources such as processing, storage and applications come from managed pools, with usage measured and capacity able to expand or contract.

What it means

Instead of acquiring and operating every machine needed for a task, an organization can request suitable resources through a service. An AI team might temporarily obtain accelerators for training, store a dataset or run a model-serving application. A household might use a cloud application for shared documents without managing its servers.

The defining idea is the delivery and management model, not simply that a computer is far away. Renting a fixed remote machine with entirely manual provisioning does not necessarily exhibit the usual cloud characteristics.

NIST describes five essentials: users can provision resources themselves, access them over networks, draw from pooled resources, adjust capacity rapidly and have usage measured. “Rapid” does not mean instantaneous or unlimited. Quotas, available hardware and application design still constrain what a customer can obtain and use.

How it works

A service exposes an interface through which a customer requests an application or computing resource. The provider's management systems allocate capacity and track its use. The customer configures the parts under their control, connects data and users, and releases resources when they are no longer required.

The service model changes that division of work. Infrastructure as a service supplies fundamental computing resources, leaving the customer substantial responsibility for operating systems and software. Platform as a service supplies an environment for deploying applications. Software as a service supplies an application that users access and configure within the service's limits.

Deployment is a separate dimension. A public cloud serves a broad customer population. A private cloud serves one organization, potentially at its own premises or through a third party. Community clouds serve organizations with shared concerns. A hybrid cloud connects distinct cloud infrastructures to support movement or coordination of applications and data.

Three practical examples

A retailer temporarily increases processing capacity for a seasonal forecasting job, then releases the additional workers.
A small consultancy uses a managed document application so staff can collaborate without administering the underlying application servers.
A research team stores experiment outputs centrally and provisions a controlled computing environment for each approved training run.

Think of it this way

Think of a serviced workshop where users reserve workbenches, equipment and storage as needed. Some rent only an equipped space; others pay for an operated service. The arrangement changes what each customer must organize.

The comparison explains access and responsibility, but cloud resources are software-controlled and can be shared in ways a workshop cannot. Network dependence, data movement and application architecture also make the economics different.

Worked example

Separating shorter running time from lower cost

Suppose an invented analysis requires 400 worker-hours and can be divided perfectly across identical workers. Assume a teaching price of 2 currency units per worker-hour, with no startup delay.

Ten workers finish in 40 hours and cost 10 × 40 × 2 = 800 units. Forty workers finish in 10 hours and also cost 800 units. Extra capacity reduces elapsed time under these assumptions; it does not automatically reduce the processing bill.

If the team leaves those 40 workers allocated for another five hours, unused capacity adds 40 × 5 × 2 = 400 units. Storage and data transfer would be additional if billed separately.

Real jobs may divide unevenly, wait for data or incur coordination overhead. If the forty-worker run instead takes 12 hours, processing costs 960 units. A useful comparison therefore records completion time, total resource use and result correctness. These invented rates and idealized assumptions are not a provider quotation.

Illustrative case study

A fictional distributor moves its nightly demand analysis to a cloud platform. The first trial starts quickly, but its bill grows because completed experiments leave resources running and duplicate datasets remain in storage.

The team adds ownership labels, explicit shutdown rules and a reviewed retention policy. It also measures whether its data-loading stage keeps additional workers busy. Some experiments benefit from greater capacity, while others remain limited by that stage.

The distributor assesses the complete workflow rather than assuming that migration itself creates savings. Its decision includes operational work, recovery arrangements and the ability to export necessary data. The case is fictional.

Common misconceptions and mistakes

Cloud computing does not mean the absence of physical servers; someone operates them.
A private cloud need not be located inside the organization's building.
Metered use does not guarantee the lowest total cost or require one particular billing contract.
Scaling infrastructure does not automatically make an application scale effectively.
Using a managed service does not remove every customer responsibility for accounts, configuration or data handling.

Business use

Begin with the workload: its demand pattern, acceptable delays, data movement and skills needed to operate it. Compare a realistic service configuration against alternatives using the same performance and recovery requirements.

For AI projects, distinguish training resources from ongoing inference resources. Occasional large training runs and continuously available prediction services have different utilization patterns. Include experimentation, retained datasets and idle capacity in the estimate. Identify which team owns each resource and which responsibilities the provider actually assumes under the chosen service.

Check your understanding

A team doubles its worker count and halves running time at an unchanged hourly price per worker. Ignoring other costs, does its processing bill halve?

Answer: No. Worker-hours remain the same. The job finishes sooner, but the processing charge is unchanged under those assumptions.

Related terms

Infrastructure as a service
platform as a service
software as a service
elasticity
virtual machine
model serving
hybrid cloud

Learn next

Map a familiar application into its data, software, operating-system and hardware responsibilities. Compare how that map changes between infrastructure, platform and software services before calculating a workload-specific cost estimate.

Sources

Sources: Peter Mell and Timothy Grance, NIST SP 800-145, The NIST Definition of Cloud Computing: https://nvlpubs.nist.gov/nistpubs/Legacy/SP/nistspecialpublication800-145.pdf . The definition, five characteristics, service models and deployment models were reviewed. Operational examples and cost arithmetic are original illustrations, not current market prices or provider-specific commitments.

Previous: Cloud-based integration Next: Cloud computing architecture