Google describes the gen AI market as five layers, from the hardware at the bottom to the finished products people use at the top. The layer an organisation buys at decides what it is responsible for, what it pays for and how quickly it can move, which is why the exam guide treats the five layers as a business question. The layers stack, so an application runs on an agent, an agent runs on a platform, a platform serves a model and a model runs on infrastructure.
The sections below take the layers from the bottom up, infrastructure, models, platforms, agents and applications, and close on how an organisation chooses the layer to buy at.
The infrastructure layer
Infrastructure is the compute, storage and networking that training and inference run on. For Google this means data centres, custom designed TPUs, GPUs and the AI Hypercomputer that arranges them as one system.
Almost nobody buys at this layer for a gen AI project. Infrastructure still decides two things a product team feels directly. It sets the price and the latency at which a hosted model is available, and it decides whether a given region can serve a given model at all.
The model layer
Models are the foundation models themselves. Google's are Gemini, Gemma, Imagen and Veo, and open and third party models sit alongside them.
Buying at the model layer means calling a model directly and building everything around it. A team that buys here has the most freedom on offer, and it owns the plumbing, the retrieval, the evaluation and the safety controls.
The platform layer
Platforms are the tooling around models, meaning everything a team needs to deploy them, tune them, ground them, evaluate them and govern what they do. Gemini Enterprise Agent Platform is Google's, and Model Garden, Agent Search and the tuning and evaluation tooling all sit within it.
Most engineering teams work at this layer, because the platform handles what every project has in common and leaves the team free to spend its time on the part that is particular to its own business.
The agent layer
Agents are systems that pursue a goal. An agent plans a sequence of steps, calls tools, reads what comes back and adjusts, so building one means deciding what it may do as well as what it may say.
The business implication is a change in risk. An agent that can act creates exposure a chat interface never had, because a prompt injection now reaches something that changes data.
The application layer
Applications are finished products people use without knowing a model is underneath. The Gemini app, Gemini for Google Workspace and the Customer Engagement Suite are all applications in this sense.
An application is the fastest layer to adopt and the one with the least to configure, and a whole organisation can see value from it within weeks.
Choosing a layer to buy at
The choice of layer follows from what an organisation is trying to differentiate on.
Where the differentiator is a workflow particular to the organisation, an application supplies the AI part and the effort belongs on the workflow. Where it is a customer experience no competitor can offer, the platform layer is the least the organisation needs and agents are usually part of it. Very few organisations differentiate on the model itself, and one that believes it does should be certain before committing.
The common error runs one way. Teams build at the model layer where an application would have done, and then spend a year rebuilding the retrieval, the evaluation and the safety controls that a layer above already had.