Foundations of Cloud Computing for AI

Introduction

Cloud computing has become a foundational infrastructure model for artificial intelligence and machine learning (AI/ML). Modern AI systems depend on large datasets, specialized processors, distributed software, rapid experimentation, and reliable production services. Few organizations can economically purchase and operate all the computing, storage, networking, and security capabilities required for these activities. Cloud computing addresses this challenge by providing configurable resources over a network, usually on demand and with usage-based pricing.

Cloud computing is not simply “using someone else’s computers.” It represents a change in how computing resources are designed, acquired, managed, measured, and consumed. Organizations can move from owning fixed infrastructure to accessing a flexible pool of virtualized and automated resources. This shift is particularly important for AI because model training workloads may require thousands of processor or accelerator hours, while model-serving workloads may fluctuate significantly with user demand.

This chapter introduces the conceptual and operational foundations required to understand cloud computing for AI. It explains the evolution of cloud computing, essential terminology, the characteristics of utility computing, elasticity, resource pooling, measured service, geographic organization, shared responsibility, capital and operating expenditure, and the role of cloud platforms in digital transformation.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain the historical evolution from centralized computing to cloud computing.

2. Define essential cloud-computing terminology and distinguish major service models.

3. Explain utility computing, elasticity, resource pooling, and measured service.

4. Describe regions, availability zones, and their significance for resilience and latency.

5. Analyze the CAPEX–OPEX shift associated with cloud adoption.

6. Explain the shared-responsibility security model.

7. Relate cloud capabilities to AI development, deployment, and digital transformation.

8. Identify major cloud limitations, governance concerns, and responsible adoption practices.

1. From Centralized Computing to Cloud Platforms

The history of cloud computing is a history of abstraction and resource sharing. Early organizations operated mainframe computers that were expensive and accessed through terminals. Users shared processing capacity, but the system was usually located in one organization and managed centrally.

The personal-computer era moved computation toward individual devices. This increased user autonomy but often created fragmented infrastructure, duplicated software, and inconsistent data management. Client–server computing later divided applications between user devices and centralized servers. Organizations then built data centers containing racks of physical servers, storage systems, networking equipment, power systems, and cooling infrastructure.

Virtualization was a major transitional technology. A hypervisor can divide a physical server into multiple virtual machines (VMs), each with an operating system and allocated virtual resources. Virtualization improved server utilization, simplified provisioning, and enabled workloads to be moved between physical hosts. However, virtualization alone is not equivalent to cloud computing. Cloud computing additionally emphasizes self-service, network access, automation, pooled resources, rapid provisioning, and measured consumption.

The growth of the internet and large-scale data centers enabled infrastructure to be delivered as a service. Cloud providers developed standardized facilities containing large numbers of servers, storage devices, and networking components. Software interfaces, commonly called application programming interfaces (APIs), allow customers to create and manage resources programmatically. This automation is essential for repeatable AI experiments and scalable services.

Containerization further increased portability and deployment speed. A container packages an application and its dependencies while sharing the host operating system kernel. Container orchestration platforms can schedule containers across clusters, restart failed instances, and scale services. Serverless computing extends abstraction further by allowing developers to submit functions or workloads without managing servers directly. These developments illustrate a general progression: physical hardware is abstracted into virtual, programmable, and increasingly automated services.

2. Essential Cloud Terminology

A cloud provider is an organization that supplies computing services through a network. A cloud consumer or customer uses those services. A cloud resource may include a virtual machine, object-storage bucket, database, network, container cluster, accelerator, or managed AI service.

A data center is a physical facility containing computing, storage, networking, power, cooling, and security systems. A region is a geographically distinct area containing one or more data centers or availability zones. An availability zone (AZ) is an isolated location or failure domain within a region. Provider-specific definitions differ, so users should consult the provider’s documentation rather than assuming that all zones have identical physical arrangements.

Virtualization presents logical computing resources independently of their physical implementation. A virtual machine emulates a computer through software. A container isolates an application process and its dependencies, usually with less overhead than a VM.

Infrastructure as a Service (IaaS) provides fundamental resources such as virtual machines, block storage, and virtual networks. The customer generally manages the operating system, applications, and much of the configuration.

Platform as a Service (PaaS) provides a managed environment for deploying applications. The provider manages more of the operating system and platform, while the customer concentrates on code, data, and application configuration.

Software as a Service (SaaS) delivers a complete application to end users. The provider normally manages the application, platform, infrastructure, and updates. Cloud-based collaboration software is a common example.

For AI, Machine Learning as a Service (MLaaS) refers to managed capabilities for data preparation, experiment tracking, model training, deployment, monitoring, and sometimes access to pretrained models. MLaaS can reduce infrastructure-management effort, but it does not remove the need for data governance, model evaluation, or responsible AI practices.

A workload is a set of computational activities, such as training a neural network or serving predictions. Provisioning is the process of allocating resources. Orchestration coordinates multiple resources and services. Scaling changes capacity to match workload requirements. Horizontal scaling adds or removes instances; vertical scaling changes the capacity of an existing instance.

A service-level agreement (SLA) is a contractual statement about service commitments, such as availability targets, support, or remedies. An SLA is not a guarantee that a customer’s application will always be available: application design, configuration, dependencies, and customer responsibilities also affect outcomes.

3. Cloud Computing as Utility Computing

Utility computing treats computing as a service analogous to electricity or water. Users obtain capacity when needed and pay according to consumption or subscription terms. The analogy is useful because the customer does not normally own the generating infrastructure. However, it is not exact: cloud resources are configurable, service quality varies by design, and customers remain responsible for architecture and governance.

The National Institute of Standards and Technology (NIST) characterizes cloud computing through five essential characteristics:

• On-demand self-service
• Broad network access
• Resource pooling
• Rapid elasticity
• Measured service

These characteristics distinguish a cloud service from a conventional outsourced data center.

On-Demand Self-Service

Customers can provision resources automatically, often through a web console, command-line interface, or API, without requiring manual intervention from provider staff. An AI engineer can create a GPU instance, attach storage, run training, and release the instance through scripts or workflow tools.

Broad Network Access

Cloud services are accessible over networks using standard mechanisms. Access may occur through web applications, APIs, secure private connections, or software development kits. Network access makes collaboration and distributed deployment possible, but it also increases the importance of identity management, encryption, and network security.

Resource Pooling

Provider resources are pooled to serve multiple customers. Physical and virtual resources are dynamically assigned according to demand. Customers usually have location independence at a high level, although they may select a region or zone for legal, latency, or resilience reasons.

Rapid Elasticity

Capacity can be increased or reduced quickly. To an appropriate extent, available capacity may appear nearly unlimited to a customer. Elasticity is especially valuable when an AI inference service experiences unpredictable demand or when training requires temporary accelerator capacity.

Measured Service

Cloud systems automatically measure resource consumption. Common meters include virtual-machine time, storage volume, data transfer, database operations, API requests, and accelerator usage. Measurement supports billing, capacity planning, chargeback, and operational analysis.

4. Elasticity, Scalability, and Availability

These terms are related but not interchangeable.

Scalability is the ability of a system to handle increased workload by adding resources or improving resource capacity. Elasticity is the ability to adjust capacity dynamically as demand changes, including reducing resources when demand falls. A system may be scalable but not elastic if it can grow only through a slow manual process.

For example, an AI model-training job can be made scalable by using multiple GPUs. If those GPUs are allocated automatically for a training run and released afterward, the environment is also elastic. An inference system may use horizontal autoscaling: it launches additional serving instances when request volume rises and removes them when volume declines.

Elasticity does not mean infinite or instantaneous capacity. Quotas, regional shortages, accelerator availability, startup time, data-transfer limits, and application bottlenecks constrain practical elasticity. Autoscaling also requires suitable metrics and safe policies. Scaling based only on CPU utilization may be inappropriate for a GPU-intensive model-serving workload.

Availability describes whether a service is operational and accessible. Reliability concerns the probability that a system performs correctly over time. Resilience is the ability to tolerate and recover from failures. Elasticity supports capacity adaptation, but it does not automatically provide resilience. A poorly designed application can scale rapidly and still fail if it depends on one database, one zone, or one unreplicated storage service.

5. Resource Pooling and Multi-Tenancy

Resource pooling allows providers to achieve high utilization by serving many customers from shared infrastructure. This is called multi-tenancy. Virtualization, isolation mechanisms, identity controls, network segmentation, encryption, and monitoring help separate customer environments.

Pooling creates economic benefits. If different customers have demand peaks at different times, the provider can use the same physical capacity more efficiently than if each customer maintained a private data center sized for its maximum demand. This principle is called statistical multiplexing.

Pooling also creates risks. A hardware failure, provider service disruption, misconfiguration, or noisy-neighbor effect may affect multiple workloads. A noisy neighbor is a workload whose resource consumption degrades the performance of others sharing infrastructure. Providers use quotas, scheduling, isolation, and dedicated options to manage such risks.

For sensitive AI workloads, customers should evaluate isolation requirements, encryption options, confidential-computing features, access controls, and contractual commitments. Resource pooling does not inherently make cloud computing insecure, but it changes the security model and requires disciplined configuration.

6. Measured Service and Cloud Economics

Measurement is both a technical and managerial feature. A cloud bill may include compute time, memory, storage capacity, input/output operations, network transfer, managed-service requests, and specialized accelerator time. Prices and billing units vary by provider and service; therefore, cost analysis must use current provider documentation.

A simple illustrative estimate for a workload is:

Total cost = compute cost + storage cost + data-transfer cost + managed-service cost

This is a cost model rather than a universal billing equation. It assumes that the listed categories capture the relevant charges and that each component is calculated using the provider’s applicable price, discounts, and billing rules.

For compute, if a resource is charged at a constant rate, then:

Compute cost = hourly rate × number of resources × hours used

The rate must have units such as currency per resource-hour. For example, if four instances cost $0.50 per instance-hour and run for 6 hours, compute cost is $0.50 × 4 × 6 = $12. This excludes storage, transfer, taxes, discounts, and other charges.

Measured service enables FinOps, a discipline that combines finance, technology, and operations to manage cloud value. Techniques include budgets, alerts, tagging, departmental chargeback, rightsizing, scheduling nonproduction resources, selecting suitable pricing commitments, and monitoring idle accelerators. Cost optimization must not compromise privacy, availability, reproducibility, or model quality.

AI introduces special cost concerns. GPU and other accelerator resources may be much more expensive than general-purpose CPUs. Data movement can be substantial because datasets, model checkpoints, and logs may be large. Repeated experimentation can generate hidden costs if training jobs are not tracked and terminated. A technically successful project can therefore be economically unsustainable without measurement and governance.

7. Regions, Availability Zones, and Failure Domains

Cloud infrastructure is geographically organized to support performance, compliance, and resilience. A region is selected according to several factors:

• Distance from users and therefore network latency
• Data-residency and regulatory requirements
• Availability of desired services or accelerators
• Pricing and tax considerations
• Disaster-recovery strategy
• Connectivity to other systems and data sources

Within a region, availability zones are designed to reduce the likelihood that a localized failure affects all deployed resources. A resilient application may distribute instances across multiple zones and use replicated or managed data services. However, zone separation does not eliminate all common-cause failures. A regional outage, software defect, identity-service failure, or incorrectly deployed configuration can still affect multiple zones.

A multi-region architecture deploys services in more than one region. It can improve disaster recovery and reduce latency for geographically distributed users, but it introduces complexity involving data replication, consistency, cost, legal restrictions, and operational coordination.

A useful design principle is to identify the failure domain relevant to each component. For example, placing two application servers in the same zone may provide redundancy against process failure but not against a zone outage. Resilience claims should be based on actual architecture rather than labels such as “high availability.”

8. Shared Responsibility for Security

Cloud security is a shared responsibility between provider and customer. The provider is generally responsible for security of the cloud: physical facilities, hardware, core networking, and the underlying managed service. The customer is responsible for security in the cloud: identities, permissions, data, workload configuration, and application behavior. The exact boundary varies with the service model.

With IaaS, customers typically manage guest operating systems, patches, firewall rules, applications, and data. With PaaS, the provider manages more of the platform, but customers still control application code, identities, and data. With SaaS, the provider manages most infrastructure and application operation, while customers remain responsible for user accounts, access settings, data governance, and appropriate use.

Shared responsibility does not mean that responsibility is divided equally or that the provider automatically protects every customer choice. Common customer failures include publicly exposed storage, excessive permissions, hard-coded credentials, unencrypted sensitive data, unpatched virtual machines, inadequate logging, and failure to test recovery procedures.

For AI systems, security must also address training-data poisoning, unauthorized model access, prompt injection where relevant, leakage of sensitive data through logs or outputs, theft of model artifacts, and attacks on inference endpoints. Strong identity and access management, least privilege, encryption in transit and at rest, secrets management, vulnerability management, network segmentation, audit logging, backup testing, and incident response are essential controls.

Privacy and governance extend beyond technical security. Organizations should establish lawful data-use purposes, retention rules, data classification, access review, model documentation, human oversight, and procedures for responding to data-subject or regulatory requirements. Cloud location does not by itself determine legal compliance; contractual arrangements, processing activities, sector rules, and applicable jurisdictions must be assessed.

9. CAPEX–OPEX Shifts

Traditional infrastructure commonly requires capital expenditure (CAPEX): substantial upfront spending on servers, storage, networking, facilities, and equipment with useful lives extending over several years. Cloud services are commonly associated with operating expenditure (OPEX): recurring spending for consumed services, subscriptions, support, and operations.

The shift has several consequences. CAPEX can provide long-term ownership and predictable physical capacity, but it requires forecasting demand and financing assets before value is realized. OPEX can reduce upfront barriers and align spending more closely with usage, allowing experimentation and rapid expansion. However, variable usage can produce unexpectedly high bills, and long-term cloud consumption may exceed the cost of owned infrastructure for stable workloads.

The distinction should not be treated as absolute. Cloud providers invest in CAPEX, while customers may purchase reserved capacity or dedicated hardware. Businesses must evaluate total cost of ownership, including personnel, facilities, licensing, migration, data transfer, security, downtime, and opportunity cost.

For AI, cloud economics are workload-dependent. Short-lived experiments, irregular training, and rapidly changing projects often benefit from on-demand resources. Stable, predictable, high-utilization training or inference workloads may justify reserved capacity, private infrastructure, or a hybrid approach. The correct decision follows workload analysis rather than enthusiasm for either ownership or outsourcing.

10. Cloud Computing’s Role in AI and Digital Transformation

Cloud computing supports the AI lifecycle in several ways. It provides scalable data storage, distributed processing, CPU and GPU resources, managed notebooks, experiment tracking, model registries, deployment platforms, monitoring, and integration with business systems.

During data preparation, cloud object storage can hold large datasets, while distributed processing services transform data. During experimentation, teams can provision reproducible environments and access accelerators without purchasing permanent hardware. During training, orchestration tools can schedule jobs and store checkpoints. During deployment, managed endpoints can expose models through APIs and scale according to demand. During operations, monitoring can track latency, errors, drift, resource utilization, and cost.

Cloud platforms also support collaboration. Data scientists, software engineers, operations teams, and business stakeholders can work with shared repositories, controlled environments, and automated pipelines. Infrastructure as code expresses infrastructure configuration in version-controlled files, improving repeatability and reviewability.

Digital transformation is broader than migration to cloud servers. It involves redesigning processes, products, decision-making, and organizational capabilities around digital technologies. Cloud enables transformation by reducing the time needed to acquire infrastructure, connect services, test ideas, and scale successful products. It does not guarantee transformation. Organizations may simply reproduce inefficient processes in the cloud, a condition sometimes called “lift and shift without modernization.”

AI creates value only when connected to a meaningful business or social objective. A cloud-based model should therefore be evaluated using technical metrics such as accuracy, latency, and availability, as well as operational and organizational measures such as cost per prediction, user adoption, fairness, explainability, safety, and business impact.

11. Illustrative Case: Cloud Migration for an AI-Based Retail Forecasting Service

Background

Consider a fictional mid-sized retailer, Northstar Retail, operating stores and an online marketplace. It wants to forecast product demand to reduce stockouts and excess inventory. The existing system uses a small on-premises server cluster and runs a weekly batch process.

Problem

The retailer’s data volume is increasing, and the forecasting team wants to test multiple algorithms and train models more frequently. The on-premises cluster has insufficient accelerator capacity and is heavily used by other business applications. Purchasing enough hardware for peak experimentation would be expensive and would leave resources idle at other times.

Analysis

The workload has two distinct patterns. Training is periodic, compute-intensive, and suitable for temporary scalable resources. Inference is more predictable for daily replenishment but may require rapid responses during seasonal events. Data includes sales transactions, customer information, supplier data, and potentially sensitive behavioral attributes. Security, access control, data residency, and cost management are therefore important.

Approach and Solution

Northstar selects a cloud region consistent with its legal and operational requirements. It stores curated training data in encrypted object storage, with access controlled through role-based identities. A managed processing service creates training datasets, while GPU-enabled instances are provisioned for experiments and released after jobs finish. Training configurations, code, and environment specifications are version-controlled.

The organization deploys the selected model behind a managed endpoint. Instances scale horizontally according to request volume, while batch forecasts use scheduled jobs. Monitoring records prediction latency, error rates, data-quality indicators, model drift, and resource costs. Production and development accounts are separated, and budgets and alerts are configured. Backups and recovery procedures are tested rather than merely documented.

Outcome

The retailer can run more experiments without purchasing permanent accelerator infrastructure and can respond more effectively to seasonal demand. However, it discovers that data-transfer charges and unused development resources increase costs. It addresses these issues through data-local processing, automatic shutdown policies, tagging, and periodic architecture reviews.

Learning Points

The case demonstrates that cloud value comes from matching workload characteristics to cloud capabilities. Elasticity is useful because training demand is intermittent. Resource pooling reduces the need to own peak capacity. Measured service exposes costs but also requires active governance. Security remains a shared responsibility, and AI success depends on data quality, monitoring, and operational integration rather than compute capacity alone.

12. Challenges, Limitations, and Best Practices

Cloud adoption introduces several limitations. Provider dependence can create vendor lock-in, especially when applications rely on proprietary databases, APIs, or AI services. Portability may be improved through containers, open standards, abstraction layers, and documented exit plans, although complete portability is rarely free.

Network dependence can create latency, outage, and data-transfer concerns. Organizations should identify which workloads require local processing, caching, private connectivity, or edge deployment. Cloud services can also become complex: many configuration choices increase the possibility of misconfiguration and skills gaps.

Organizations should adopt a structured approach:

1. Classify data and workloads before selecting services.

2. Define business, technical, security, and compliance requirements.

3. Choose regions and architectures based on latency, resilience, and legal needs.

4. Apply least privilege and automate secure configuration.

5. Use infrastructure as code and version-controlled environments.

6. Establish budgets, tagging, monitoring, and ownership for resources.

7. Design for failure and test backup and recovery processes.

8. Measure AI quality, fairness, drift, latency, availability, and cost.

9. Document provider dependencies and migration options.

10. Review architecture continuously as workloads and regulations change.

Contemporary cloud developments include specialized AI accelerators, managed generative-AI services, confidential computing, serverless inference, edge-cloud integration, and sustainability-aware scheduling. These developments expand possibilities but also raise questions about energy consumption, concentration of infrastructure power, transparency of managed models, intellectual property, and the environmental cost of large-scale computation. Technical progress should therefore be evaluated alongside social, legal, economic, and environmental consequences.

Chapter Summary

Cloud computing evolved from centralized and virtualized computing into programmable, network-accessible services. Its defining characteristics include on-demand self-service, broad network access, resource pooling, rapid elasticity, and measured service. IaaS, PaaS, SaaS, and MLaaS represent progressively managed service models, with different distributions of customer and provider responsibility.

Elasticity allows resources to follow changing demand, while scalability describes the ability to handle growth. Resource pooling and multi-tenancy improve utilization but require isolation and governance. Regions and availability zones support latency, compliance, and resilience decisions, although they do not eliminate common-cause failures. Measured service enables billing and FinOps but makes cost visibility and control essential, particularly for expensive AI accelerators and data movement.

The CAPEX–OPEX shift can reduce upfront investment and accelerate experimentation, but cloud is not automatically cheaper. Workload patterns, utilization, data transfer, personnel, compliance, and lock-in must be considered. The shared-responsibility model requires customers to secure identities, data, configurations, applications, and AI workloads even when the provider operates the underlying infrastructure.

For AI and digital transformation, cloud computing supplies scalable storage, processing, accelerators, orchestration, deployment, monitoring, and collaboration. Its value depends on sound architecture, responsible governance, cost discipline, and meaningful organizational objectives. Cloud is an enabling foundation, not a substitute for engineering judgment or responsible AI practice.

Key Terms

Cloud computing: Delivery of configurable computing resources over a network as services.

Cloud provider: Organization that operates and supplies cloud services.

Virtualization: Abstraction of physical computing resources into logical resources.

Virtual machine: Software-defined computer environment with its own operating system.

Container: Isolated application package sharing the host operating-system kernel.

IaaS: Service model providing fundamental compute, storage, and networking resources.

PaaS: Managed platform on which customers deploy applications.

SaaS: Complete software application delivered as a service.

MLaaS: Managed services supporting machine-learning development and operations.

Elasticity: Ability to increase or decrease capacity in response to changing demand.

Scalability: Ability to handle increased workload through additional or more capable resources.

Resource pooling: Dynamic sharing of provider resources among multiple customers.

Multi-tenancy: Operation of multiple customer workloads on shared infrastructure with isolation.

Measured service: Automatic monitoring and metering of resource consumption.

Region: Geographically distinct cloud location containing one or more zones.

Availability zone: Isolated location or failure domain within a cloud region.

CAPEX: Upfront expenditure on long-lived assets such as infrastructure.

OPEX: Recurring expenditure for operating or consuming services.

Shared responsibility: Security model dividing duties between provider and customer.

FinOps: Collaborative practice for managing the financial value of cloud usage.

Review Questions

Short Answer Questions

1. What are the five essential characteristics of cloud computing?

2. Distinguish scalability from elasticity.

3. What is the purpose of an availability zone?

4. Give two examples of customer responsibilities under the shared-responsibility model.

5. Why are measured services important for AI workloads?

Long Answer Questions

1. Explain the evolution from physical data centers to virtualized and cloud-based infrastructure.

2. Compare IaaS, PaaS, SaaS, and MLaaS in terms of management responsibility and AI applications.

3. Discuss how regions and availability zones influence latency, compliance, and resilience.

4. Analyze the advantages and risks of shifting from CAPEX to OPEX.

5. Explain how cloud computing supports the complete AI lifecycle from data preparation to monitoring.

Analytical / Application Questions

1. A company runs occasional GPU-intensive training jobs but operates a steady, high-volume inference service. Recommend a cloud, private, or hybrid strategy and justify your decision using workload characteristics.

2. Design a multi-zone architecture for an AI prediction API. Identify likely failure domains and explain how your design addresses them.

3. An organization’s cloud bill is rising because of idle GPUs, repeated data transfers, and untagged resources. Propose a FinOps and architecture-improvement plan without reducing required model quality or security.

Multiple Choice Questions with Answers

Question 1

Which characteristic most directly allows a cloud customer to provision resources without manual provider intervention?

A. On-demand self-service

B. Resource pooling

C. Measured service

D. Multi-region deployment

Correct Answer: A. On-demand self-service

Question 2

Which statement best distinguishes elasticity from scalability?

A. Elasticity concerns only storage, while scalability concerns only compute.

B. Elasticity adjusts capacity dynamically as demand changes, while scalability concerns handling growth.

C. Elasticity requires owned hardware, while scalability requires cloud services.

D. Elasticity measures cost, while scalability measures security.

Correct Answer: B. Elasticity adjusts capacity dynamically as demand changes, while scalability concerns handling growth.

Question 3

Which cloud service model generally gives the customer the greatest control over the operating system?

A. SaaS

B. PaaS

C. IaaS

D. MLaaS

Correct Answer: C. IaaS

Question 4

What is a primary purpose of distributing application instances across availability zones?

A. To eliminate all software defects

B. To reduce exposure to a localized zone failure

C. To guarantee unlimited capacity

D. To remove the need for identity management

Correct Answer: B. To reduce exposure to a localized zone failure

Question 5

A resource costs $0.50 per instance-hour. What is the compute cost of four instances running for six hours, excluding all other charges?

A. $3

B. $8

C. $12

D. $24

Correct Answer: C. $12

Question 6

Under the shared-responsibility model, which activity is typically a customer responsibility?

A. Maintaining the provider’s physical data-center cooling system

B. Manufacturing the provider’s servers

C. Configuring customer identities and access permissions

D. Repairing the provider’s core networking hardware

Correct Answer: C. Configuring customer identities and access permissions

Question 7

Which situation is most likely to benefit from cloud elasticity?

A. A workload with completely fixed demand and no variation

B. An intermittent GPU training workload with unpredictable experiment frequency

C. A disconnected device that never communicates with a network

D. A physical security gate with no software component

Correct Answer: B. An intermittent GPU training workload with unpredictable experiment frequency

Question 8

What does measured service enable most directly?

A. Automatic elimination of security risks

B. Usage-based billing and resource-cost analysis

C. Guaranteed application correctness

D. Permanent ownership of provider hardware

Correct Answer: B. Usage-based billing and resource-cost analysis

Question 9

Which statement about CAPEX and OPEX is most accurate?

A. Cloud usage is always cheaper than owned infrastructure.

B. CAPEX refers to recurring usage charges, while OPEX refers to server ownership.

C. Cloud can reduce upfront infrastructure investment, but total cost depends on workload and operating factors.

D. OPEX eliminates the need for financial governance.

Correct Answer: C. Cloud can reduce upfront infrastructure investment, but total cost depends on workload and operating factors.

Question 10

Which practice best supports reproducible cloud-based AI experimentation?

A. Manually changing environments without documentation

B. Using version-controlled code, configurations, and infrastructure definitions

C. Sharing one unrestricted administrative account

D. Keeping training data and model versions unidentified

Correct Answer: B. Using version-controlled code, configurations, and infrastructure definitions