Innovating with Google Cloud Artificial Intelligence · Topic 4 of 4
Custom Models, BigQuery ML and AI Hypercomputer
Build custom models with your own data using Agent Studio and AutoML on Agent Platform, run ML with SQL in BigQuery ML, and the parts of AI Hypercomputer.
- 4 min
- Your own data makes models specific to your business and differentiating.
- Agent Studio: prompts and model customization. AutoML: custom models without code. BigQuery ML: models with SQL.
- BigQuery ML trains where the data lives, using SQL skills.
What it is
When pre-built AI is not enough, organizations create custom models using their own data, which can give a competitive edge no rival can copy. Google Cloud offers several routes: Agent Studio and AutoML on Gemini Enterprise Agent Platform, and BigQuery ML, which builds models with SQL inside the data warehouse. Underneath runs AI Hypercomputer, Google's integrated AI infrastructure.
How it works
Why use your own data
Generic models know general things. Your data captures your products, customers and processes. A model adapted to that data gives answers and predictions specific to your business, which is where differentiation comes from.
Routes to a custom model
| Route | What it does | Who it suits | Example |
|---|---|---|---|
| Agent Studio on Agent Platform | A workspace to design and test prompts and customize foundation models such as Gemini with your own examples | Teams adapting generative AI to their style and content | Tune a model to write support replies in the company's tone |
| AutoML on Agent Platform | Trains a custom model on your labeled data with little or no code; the platform chooses and tunes the model | Teams with data but limited ML expertise | Classify product images into the company's own categories |
| BigQuery ML | Create and run ML models using standard SQL inside BigQuery | Data analysts who know SQL and whose data is in BigQuery | Forecast monthly sales from historical tables |
| Custom training on Agent Platform | Data scientists write their own training code and frameworks | Teams with ML engineers and unique needs | A proprietary risk model |
"Analysts who know SQL, data already in BigQuery" is BigQuery ML. "Limited ML expertise, custom categories" is AutoML.
BigQuery ML
BigQuery ML lets users create and execute machine learning models in BigQuery using standard SQL queries. The business value:
- No data movement: models train where the data already is, which is faster and keeps governance in one place.
- Uses existing skills: analysts use SQL instead of learning Python and ML frameworks.
- Flexibility to experiment: try several models quickly on the same data.
- Built-in model types: for example regression, classification, time series forecasting and recommendation.
The idea in one line: CREATE MODEL trains a model from a query, and ML.PREDICT uses it to score new rows.
AI Hypercomputer
AI Hypercomputer is Google Cloud's integrated supercomputing system for AI. It is not a single machine; it combines three parts:
| Component | What it includes | Business benefit |
|---|---|---|
| Performance-optimized hardware | TPUs (Google's Tensor Processing Units, designed for AI) and GPUs, with fast networking and storage | Train and serve large models faster |
| Open software | Popular open frameworks and standards such as JAX, PyTorch and TensorFlow, plus Google's optimizations | Use familiar tools; avoid lock-in |
| Flexible consumption | Options such as on-demand, committed use discounts, Spot VMs and Dynamic Workload Scheduler | Control cost by matching how you pay to how you use capacity |
The result is better performance and efficiency for AI workloads: more training and inference done per dollar.
TPUs and GPUs
| TPU | GPU | |
|---|---|---|
| Designed by | Google, specifically for ML workloads | Originally for graphics, now widely used for AI |
| Strength | Large-scale training and serving of ML models | Flexible acceleration for many AI and compute tasks |
You do not need hardware detail for the exam; know that both accelerate AI and that TPUs are Google's own design.
A worked scenario
A retailer's analysts want to forecast demand for each store. Their sales data is in BigQuery and they know SQL, so they use BigQuery ML. The marketing team wants product descriptions in the brand's voice, so it tunes Gemini with past descriptions in Agent Studio. The data science team trains a large recommendation model on AI Hypercomputer using TPUs, with Dynamic Workload Scheduler to obtain capacity cost-effectively.
How the exam asks about this
- "Analysts want to build a model using SQL on data in BigQuery." BigQuery ML.
- "Create a custom image model with little ML expertise." AutoML on Agent Platform.
- "Adapt a foundation model with the company's own examples." Agent Studio on Agent Platform.
- "Which components make up AI Hypercomputer?" Hardware (TPUs and GPUs), open software, flexible consumption.
What to remember for the exam
- Your own data makes models specific to your business and differentiating.
- Agent Studio: prompts and model customization. AutoML: custom models without code. BigQuery ML: models with SQL.
- BigQuery ML trains where the data lives, using SQL skills.
- AI Hypercomputer: performance-optimized hardware, open software, flexible consumption.
- TPUs are Google-designed AI accelerators.
AutoML still needs good labeled data. Low-code does not mean no data preparation.
Common mistakes
Moving BigQuery data out to train a simple model. BigQuery ML avoids the move.
Thinking AI Hypercomputer is only hardware. Software and consumption models are part of it.
Choosing custom training for a team with no ML skills. AutoML or BigQuery ML fits better.
Check yourself
Which product lets analysts create ML models with standard SQL?
BigQuery ML.
What are the three parts of AI Hypercomputer?
Performance-optimized hardware such as TPUs and GPUs, open software, and flexible consumption models.
Which option trains a custom model with little or no code?
AutoML on Gemini Enterprise Agent Platform.
Topic 4 of 4 in Innovating with Google Cloud Artificial Intelligence
Test yourself on Innovating with Google Cloud Artificial Intelligence
The full Google Cloud Digital Leader question bank has 512 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Google Cloud. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.