Overview
Domain 5 surveys how Oracle packages AI on its cloud. At the top are applications with embedded AI, such as Oracle Fusion Cloud Applications. Below them sit the AI services: OCI Generative AI and pretrained services for language, speech, vision and documents that developers call through APIs without building models.
Next come the machine learning services for teams that build their own models, chiefly OCI Data Science and Oracle Machine Learning inside Oracle AI Database, plus labelling support through OCI Data Labeling. At the base is infrastructure: GPU compute shapes, OCI Supercluster with RDMA cluster networking, and high-performance storage. The section also covers responsible AI, including Oracle's stated principles and how bias, transparency and human oversight apply when organisations deploy AI.
How it's tested
Expect selection questions built around team skills, data ownership and effort. A scenario describes what a company wants to achieve and who will do the work, and you choose a pretrained service, a model-building platform or infrastructure. Responsible AI items describe a risk and ask which principle or control applies.
-
Build or consume
You decide between a pretrained OCI AI service and building a custom model in OCI Data Science or Oracle Machine Learning.
-
Size the infrastructure
A training or inference workload is described and you choose GPU shapes, Supercluster networking or managed services.
-
Name the principle
A fairness, transparency or oversight concern is described and you match it to the responsible AI principle it reflects.
A data science team wants to train a churn model in Python using open source libraries, track experiments and deploy the model as a REST endpoint. Pretrained AI services cannot learn churn from the team's own data, so they do not fit. OCI Data Science provides notebook sessions, a model catalog and model deployments for exactly this workflow, so it is the correct choice.
Objectives 4
Official objectives from the Oracle outline, in outline order.
5.1 Discuss OCI AI servicesPretrained and generative services called through APIs.
- What it means
- OCI AI services are managed, pretrained capabilities that developers use through APIs, SDKs or the console without machine learning expertise. They include OCI Generative AI, Language, Speech, Vision, Document Understanding and others such as Digital Assistant.
- What to understand
- These services ship with models Oracle has already trained, so you send data and get results such as sentiment, transcripts or detected objects. Several also allow custom models trained on your labelled data, for example custom image classification in Vision or custom text classification in Language, still without writing training code. They are billed per use and scale automatically. They suit common tasks where speed to value matters more than full control.
- How it shows up
- Expect a scenario where a team without data scientists needs a standard AI task delivered quickly.
- Common confusion
- Assuming pretrained services cannot be customised at all; several accept your labelled data.
5.2 Discuss OCI ML servicesPlatforms for building, training and deploying custom models.
- What it means
- OCI's ML services are for teams that build their own models. OCI Data Science is a managed platform with notebook sessions, conda environments, a model catalog, jobs, pipelines and model deployments. Oracle Machine Learning runs algorithms inside Oracle AI Database so data does not move.
- What to understand
- OCI Data Science suits Python and open source workflows, including AutoML through Oracle's Accelerated Data Science SDK. Models are versioned in the catalog and served as HTTP endpoints through model deployment. Oracle Machine Learning offers in-database algorithms through SQL, Python and R, OML Notebooks and an AutoML user interface, which suits data already stored in the database. OCI Data Labeling creates labelled datasets for training.
- How it shows up
- Questions contrast Data Science with Oracle Machine Learning based on where the data lives and team skills.
- Common confusion
- Choosing OCI Data Science when the requirement stresses keeping data inside the database.
5.3 Discuss OCI AI infrastructureGPU shapes, Supercluster and supporting storage.
- What it means
- OCI provides bare metal and virtual machine GPU compute shapes with NVIDIA and AMD accelerators for training and inference. OCI Supercluster joins large numbers of GPU nodes with RDMA cluster networking so distributed training can scale.
- What to understand
- Large model training splits work across many GPUs, which constantly exchange data, so network latency becomes the bottleneck. RDMA over Converged Ethernet lets nodes move data between memory directly with very low latency. Bare metal shapes avoid virtualisation overhead. Smaller GPU shapes suit inference and fine-tuning, while the largest clusters target foundation model training. High-throughput storage, including local NVMe and file systems, keeps GPUs fed with data.
- How it shows up
- A scenario about training a very large model across many GPUs expects Supercluster with RDMA networking.
- Common confusion
- Assuming more CPUs can replace GPUs for deep learning training at scale.
5.4 Explain responsible AIPrinciples and practices for trustworthy AI.
- What it means
- Responsible AI means designing and running AI systems that are lawful, ethical and robust. Oracle's course frames this through principles such as human oversight, fairness, transparency and explainability, privacy, safety and accountability.
- What to understand
- Bias can enter through unrepresentative training data, flawed features or the way outputs are used. Mitigations include diverse data, bias testing, explainable models, documentation and a human able to review or override high-impact decisions. Transparency means users know they are dealing with AI and can understand the basis of a decision. Privacy and security controls protect training and inference data. These principles apply to generative systems as well as predictive ones.
- How it shows up
- A described harm, such as unfair loan denials, is matched to the principle or control that addresses it.
- Common confusion
- Treating responsible AI as only a legal compliance task rather than a design and monitoring practice.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
AI services layer
Pretrained APIs for language, speech, vision, documents and generative AI.
- Remember
- No model building needed.
- Tested as
- Build-or-consume scenarios.
- Trap
- Overlooking that some allow custom training on your labels.
-
OCI Data Science
Managed platform for building and deploying custom Python models.
- Remember
- Notebooks, catalog, jobs, pipelines, deployments.
- Tested as
- Custom model scenarios.
- Trap
- Thinking it is a pretrained API.
-
Model catalog
Stores and versions trained models with metadata.
- Remember
- Feeds model deployment.
- Tested as
- Lifecycle questions.
- Trap
- Confusing it with OCI Generative AI's model list.
-
Model deployment
Serves a catalogued model as an HTTP endpoint.
- Remember
- Managed scaling for inference.
- Tested as
- Deploy step questions.
- Trap
- Assuming notebooks are production endpoints.
-
Oracle Machine Learning
In-database ML through SQL, Python and R.
- Remember
- Data stays in Oracle AI Database.
- Tested as
- Data-movement scenarios.
- Trap
- Choosing Data Science when moving data is forbidden.
-
OCI Data Labeling
Service for creating labelled datasets.
- Remember
- Feeds custom models.
- Tested as
- Custom training prerequisites.
- Trap
- Thinking it trains models itself.
-
GPU shapes
Compute shapes with NVIDIA or AMD accelerators.
- Remember
- Bare metal and VM options.
- Tested as
- Infrastructure selection.
- Trap
- Using CPU shapes for large training.
-
OCI Supercluster
Large-scale GPU clusters joined by RDMA networking.
- Remember
- Built for distributed training.
- Tested as
- Foundation model training scenarios.
- Trap
- Assuming standard VCN networking is enough.
-
RDMA cluster networking
Low-latency direct memory transfer between nodes.
- Remember
- Removes the inter-GPU communication bottleneck.
- Tested as
- Why Supercluster scales.
- Trap
- Linking it to internet bandwidth.
-
Human oversight
A person can review or override AI decisions.
- Remember
- Key for high-impact outcomes.
- Tested as
- Responsible AI items.
- Trap
- Thinking full automation is always preferred.
-
Fairness and bias
Avoiding unjust outcomes for groups.
- Remember
- Bias often starts in training data.
- Tested as
- Responsible AI items.
- Trap
- Assuming removing one sensitive column removes all bias.
-
Transparency and explainability
Users understand that AI is involved and why it decided.
- Remember
- Supports trust and appeals.
- Tested as
- Responsible AI items.
- Trap
- Equating it with publishing source code.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Layers of the OCI AI stack
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| AI in applications | AI built into Oracle Fusion and industry apps. | Consumed as features, no development. | Ignoring it when the user simply needs an app capability. |
| AI services | Pretrained and generative APIs. | Call an API, no training code. | Expecting full control over the algorithm. |
| ML services | OCI Data Science, Oracle Machine Learning. | Build and train your own models. | Using them for a standard task an API already does. |
| AI infrastructure | GPU compute, Supercluster, storage. | Raw capacity you manage. | Choosing it when a managed service would do. |
OCI Data Science vs Oracle Machine Learning
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| OCI Data Science | Notebook-based Python platform. | Open source libraries, data loaded into the platform. | Picking it when data must stay in the database. |
| Oracle Machine Learning | Algorithms run inside the database. | No data movement, SQL access. | Expecting arbitrary open source frameworks to run in-database. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Build when you could consume
- What candidates pick
- OCI Data Science for document text extraction
- Why it looks right
- Custom sounds more powerful.
- Why it's wrong
- A pretrained service such as Document Understanding already extracts text and fields, with no model building required.
Clue to watch for Team has no data scientists or tight deadlines.
-
Data movement overlooked
- What candidates pick
- Export tables to OCI Data Science
- Why it looks right
- Data Science is the best-known ML platform.
- Why it's wrong
- If policy keeps data inside the database, Oracle Machine Learning trains models in place without exporting it.
Clue to watch for Phrases like must not leave the database.
-
CPUs for large training
- What candidates pick
- Many large CPU shapes
- Why it looks right
- CPU capacity is cheaper and familiar.
- Why it's wrong
- Deep learning training depends on parallel matrix work that GPUs perform far faster than general-purpose CPUs.
Clue to watch for Billions of parameters or large image sets.
-
Network ignored
- What candidates pick
- Standalone GPU VMs in separate regions
- Why it looks right
- More GPUs anywhere sounds like more power.
- Why it's wrong
- Distributed training needs nodes to exchange gradients constantly, which requires low-latency cluster networking in one location.
Clue to watch for Training split across many nodes.
-
Removing a column fixes bias
- What candidates pick
- Drop the gender column
- Why it looks right
- It looks like a direct fix.
- Why it's wrong
- Other features can act as proxies for the removed attribute, so bias testing and monitoring are still needed.
Clue to watch for Fairness complaint after deployment.
-
Notebook as production
- What candidates pick
- Serve predictions from a notebook session
- Why it looks right
- The model already runs there.
- Why it's wrong
- Production inference should use model deployment, which provides a managed, scalable HTTP endpoint.
Clue to watch for External applications need predictions.
Answer framework
Use these steps to pick the right part of Oracle's AI stack for a Domain 5 scenario.
- Check whether an Oracle application already provides the capability as a built-in feature before considering any development.
- If development is needed, ask whether a pretrained or generative OCI AI service handles the task through an API.
- When the business needs a model trained on its own data, choose OCI Data Science or Oracle Machine Learning based on where data lives.
- For training or hosting very large models, select GPU shapes and Supercluster with RDMA networking rather than CPU compute, because low-latency links between GPUs keep distributed training efficient.
- Review the design against responsible AI principles, adding human review, bias testing and transparency where decisions affect people.
Practice question
Want more like this? The 1Z0-1122-26 practice test has questions for every domain.
Checklist
0 of 11
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Spend five to seven hours here. Start with the OCI AI portfolio lessons in Oracle's course and draw the stack as four stacked boxes: applications, AI services, ML services and infrastructure. Under each box, list the products and one sentence on who uses them.
If you have an OCI tenancy, even a free tier account, open the console and browse the Analytics and AI menu. Seeing where Data Science, Language, Vision and Generative AI live makes the portfolio easier to recall. Open a Data Science project page and note the notebook, model and deployment sections.
For infrastructure, learn the reasoning rather than shape names, which change often: GPUs for parallel training, bare metal to avoid overhead, RDMA cluster networking for multi-node jobs.
For responsible AI, write each principle with one example of a harm it prevents. Then work through practice questions focused on build-versus-consume choices, since they are the most common item type in this section, and review every miss against your stack diagram.
FAQ
How many questions come from Get Started with OCI AI Portfolio?
Oracle does not publish weights for 1Z0-1122-26. Secondary sources estimate around six of the 40 questions for this portfolio section.
Do I need to memorise GPU shape names?
No. Understand GPU versus CPU, bare metal versus VM and why Supercluster uses RDMA networking. Shape names change too often to be core exam content.
What is the difference between AI services and ML services?
AI services are pretrained APIs you consume. ML services are platforms where you build and train your own models.
Is OCI Generative AI part of this section?
It appears here as part of the portfolio, but the details of models, clusters and Select AI are in Domain 6.
Which responsible AI principles should I know?
Human oversight, fairness, transparency and explainability, privacy, safety and accountability, framed as AI that is lawful, ethical and robust.
Can pretrained services be customised?
Some can. Vision and Language, for example, let you train custom models on your labelled data without writing training code.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 5 knowledge
You have mapped the OCI AI stack, ML platforms, GPU infrastructure and responsible AI. Practise choosing the right layer for each scenario.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the 1Z0-1122-26 AI tutor Get any concept explained your way