Prepare and test your skills
Prepare and test your skills
Assessing model modality and technical constraints is the first filter for choosing a foundation model for a business use case. Foundation models (opens in a new tab) vary in what data they can process and how much data they can handle at once. A model is a candidate only when its modality matches the use case data and its context window and performance fit the workload size and speed needs.
Model modality is the type of content a foundation model can accept as input and produce as output, such as text, images, audio, video, or code. A text-only model accepts and returns only text, which fits summarization, question answering, and drafting. A multimodal model accepts or produces more than one type of content, for example accepting an image and returning a text description or accepting text and returning an image. The business use case determines the required modality, so a use case that must read scanned invoices needs image input while a use case that must create marketing copy needs text output. If the modality does not match, the task cannot be completed no matter how well the model performs in other areas.
The context window is the maximum amount of information, measured in tokens, that a model can consider in a single request. Tokens are the small pieces of content that a model processes, so a longer document, longer conversation history, or larger set of reference material consumes more tokens. A use case that must reason over a long contract, a full codebase, or a lengthy chat history requires a model with a context window large enough to hold that content. When input exceeds the context window, the content must be shortened, split into multiple requests, or summarized first, which adds complexity and can lose information. Larger context windows generally increase cost and response time, so the goal is to choose a window that fits the use case rather than the largest available.
Performance metrics describe how a model behaves when it runs and whether it is practical for a given workload. The metrics that matter most for model selection are latency, throughput, quality, and reliability:
These metrics often trade off against each other. A larger, more capable model may give higher-quality output but respond more slowly and cost more per request, while a smaller model may respond quickly and cheaply but handle only simpler tasks.
Matching modality and constraints to a use case is an ordered check that keeps only models that pass modality, input size, and performance targets. The process flows from the data type to the data volume to the operating targets:
A decision tree that starts with the business use case and applies ordered filters: modality match, context window capacity, and performance targets. Failures at any pass-or-fail filter branch to a rejected outcome; only models passing all filters are compared and the least expensive suitable one is chosen.
Exam tip: Modality and context window are pass-or-fail filters, while performance metrics are compared among the models that pass — a model with excellent output is not valid if it cannot accept the input type or fit the input in its context window.
Evaluating operational reliability and cost-efficiency means choosing a deployment approach that stays available under changing demand while keeping total cost sustainable. Reliability comes from scalable resources, multi-region design, and continuous monitoring, while cost-efficiency comes from matching pricing models and discounts to actual usage. Together they allow consistent model deployment without over-provisioning or unexpected spending.
Reliable deployment is the ability of cloud resources to handle workload fluctuations, support high availability, and recover from failure without interruption. Google Cloud Observability supports reliability by collecting monitoring, logging, and performance profiling data from workloads, data, and processes to reveal trends and demand patterns. Multi-region resources add reliability because services such as Cloud Storage with dual- or multi-region buckets handle data replication automatically, so data remains available without manual migration. When planning across regions, reliability decisions affect cost because cross-regional traffic, new resources in the target region, and network egress charges must be assessed with the Google Cloud Pricing Calculator (opens in a new tab).
Autoscaling adjusts the number of resources based on workload fluctuations, so applications have capacity when demand rises and avoid unnecessary cost when demand falls. Three Google Cloud options divide this work by platform:
By using autoscaling, teams improve performance during peaks, reduce waste during idle periods, and use cloud resources more efficiently.
Pricing models and discounts determine the billing rate paid for cloud resources, with a central FinOps team typically optimizing rates while product teams optimize usage. Committed use discounts (CUDs) provide discounts in exchange for committing to a level of usage over a one-year period or a three-year period, and significant savings can come from using CUDs on top of negotiated enterprise discounts. Product teams can also lower rates with sustained-use and preemptible capacity and by choosing regions that balance cost, latency, and compliance. BigQuery shows how pricing differs by workload because it offers on-demand pricing and edition-based pricing tied to commitments and autoscaling, where the right edition lowers cost per slot for high-volume analytics.
| Pricing option | Commitment required | Where discount applies |
|---|---|---|
| Enterprise discounts | Minimum total spend on Google Cloud at a reduced rate | Organization billing account by negotiation |
| Resource-based CUDs | Minimum quantity of Compute Engine resources for 1 or 3 years in a specific region | Specific region, shared across projects in a Cloud Billing account |
| Spend-based CUDs | Minimum spend on a particular product for 1 or 3 years | Billing account level, regionally or globally depending on product |
| Spot VMs | No commitment, uses unused VM capacity for fault-tolerant workloads | Specific VMs at more than 80% lower cost than regular VMs |
Pricing models change over time, so architectures should be reviewed regularly for new features that offer better performance or lower cost.
Exam tip: Choose resource-based CUDs when committing to a quantity of resources in one region, and spend-based CUDs when committing to spend on a product at the billing-account level.
Monitoring, tracking, and planning costs over time depends on visibility into spending, clear ownership of resources, and a long-term growth strategy. Cost Management tools such as budgets and alerts let teams set budgets and receive alerts when budgets are exceeded, which helps track spending and find areas for reduction. Tagging and labeling organize resources by project, department, or other dimensions, so teams can monitor spending patterns for specific resources and find high usage or potential savings. Product-specific techniques further reduce waste, including Object Lifecycle Management and Autoclass in Cloud Storage, autoscaling and custom machine types in Compute Engine, cluster autoscaler and node auto-provisioning in GKE, and capacity-based pricing in BigQuery. Sustainable growth planning connects these controls to business needs through business requirements, trends from observability data, and analysis of the cost and carbon impact of expanding an existing environment versus provisioning a new one.
Examining security requirements and customization strategies means deciding how a foundation model must protect data and when it must be adapted for specialized needs. Security requirements control how data is protected, which regulations must be met, and where data can be stored and processed. Customization strategies decide when a general-purpose model is enough and when adaptation such as fine-tuning is needed for domain vocabulary, accuracy, or controls.
The shared responsibility model (opens in a new tab) describes how security duties are divided between Google Cloud and the customer organization. Google Cloud is responsible for protecting the cloud itself, including the underlying hardware, storage, and network infrastructure. The customer is responsible for securing what it brings to the cloud, including its data and how data is accessed and used. When selecting a foundation model, this division means the organization must still control identities, access, configuration, and data handling around the model, even when Google secures the platform underneath it.
Exam tip: The customer always remains responsible for securing its data, while Google Cloud is responsible for securing the underlying infrastructure.
Compliance, security, and privacy protocols are the enterprise rules that determine whether a foundation model can be used for a regulated workload. Many workloads are subject to industry and government standards such as HIPAA, HITECH, GDPR, PCI DSS, FedRAMP, CJIS, and CCPA, and can require attestations of compliance by both the customer and the cloud provider. An organization must confirm which Google Cloud services are covered for its regulation, understand its own customer responsibilities, and review certifications in the Compliance resource center. The Google Cloud trust principles describe Google's commitment to protecting customer data privacy, including not using customer data for advertising or other purposes beyond delivering cloud services.
Essential protocols for protecting sensitive and regulated model data include:
Data residency and location controls determine where model data is allowed to be stored and processed. Geography matters for compliance because regulations may require data to stay in specific countries or regions. Organizations use resource location constraints, a type of organization policy, to ensure workloads can only be deployed in pre-approved regions. Model selection must also account for the regionality of different Google Cloud services, because some services are regional, zonal, or multiregional and the deployment target affects both compliance and availability.
A secure governed foundation is the baseline of resources, configurations, and capabilities that lets an organization adopt Google Cloud and deploy workloads, including generative AI workloads, at scale. The enterprise foundations blueprint, formerly known as the security foundations blueprint, provides deployable Terraform assets and design guidance for consistent governance, security controls, visibility, and access to shared services. The blueprint implements a defense-in-depth model that combines policy controls, architecture controls, and detective controls. Policy controls enforce acceptable configurations through infrastructure-as-code validation and organization policy constraints, architecture controls configure networks and the resource hierarchy by best practice, and detective controls detect anomalous behavior through Security Command Center and integration with a security operations center. Strong governance prevents cloud sprawl and keeps model use reliable over time through consistent naming, labeling, tags, and network protections such as VPC Service Controls, Cloud VPN or Cloud Interconnect, and firewall rules.
Customization strategies decide whether to use a foundation model as-is or adapt the surrounding environment and the model itself for specialized domain needs. The enterprise approach is to deploy the recommended baseline first and then customize the environment to address specific business requirements. For AI, that means developing and deploying machine learning models in a secure and governed environment using Gemini Enterprise Agent Platform, where features like model explainability and fairness metrics help address responsible-AI concerns. Fine-tuning or other domain adaptation becomes necessary when a general model cannot meet requirements for specialized vocabulary, accuracy, data boundaries, or regulatory controls.
Initiatives should be prioritized by risk and regulatory mandate in an ordered sequence:
This sequence ensures that basic protections and mandatory compliance are in place before specialized customization expands what the model is allowed to do.
A text-only foundation model accepts and returns only text, which fits summarization, question answering, and drafting, while a multimodal foundation model accepts or produces more than one type of content such as text, images, audio, video, or code. For example, a multimodal model can accept an image and return a text description or accept text and return an image. If the model modality does not match the use case data, the task cannot be completed no matter how well the model performs in other areas.
Choose a foundation model by first keeping only models whose modality matches the required input and output types, then keeping only models whose context window can hold the largest expected request measured in tokens. Next define acceptable latency, throughput, quality, and availability and compare the remaining models against those targets. Select the model that satisfies modality, context window, and performance needs, preferring the least expensive or simplest option that still meets them.
Resource-based Committed Use Discounts (CUDs) require committing to a minimum quantity of Compute Engine resources for one year or three years in a specific region, while spend-based Committed Use Discounts (CUDs) require committing to a minimum spend on a particular product for one year or three years at the billing-account level. Use resource-based CUDs when committing to a quantity of resources in one region and spend-based CUDs when committing to spend on a product at the billing-account level.
Under the shared responsibility model, Google Cloud is responsible for protecting the cloud itself, including the underlying hardware, storage, and network infrastructure. The customer remains responsible for securing its data and how data is accessed and used, including identities, access, configuration, and data handling around the foundation model.
An enterprise is deploying a mission-critical generative AI application on Google Cloud that experiences continuous, high-volume user traffic with strict low-latency requirements. The company needs predictable monthly expenses and guaranteed model capacity to prevent request throttling during peak business hours.
Which foundation model consumption and pricing model should the enterprise select?
Last updated