Concept 6 of 7

The five layers of the gen AI landscape

3 questions test this

Google describes the market in five layers. The reason to know them is not taxonomy, it is that the layer you buy at decides what you are responsible for, what you pay for, and how quickly you can move.

Infrastructure

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 here for a gen AI project, and it matters anyway. It is why a hosted model is available at the price and latency it is, and it is the layer that decides whether a region can serve a given model at all.

Models

The foundation models themselves. Gemini, Gemma, Imagen and Veo from Google, alongside open and third party models offered next to them.

Buying at this layer means calling a model directly and building everything around it. Maximum flexibility, and you own the plumbing, the retrieval, the evaluation and the safety controls.

Platforms

The tooling around models. Deploying them, tuning them, grounding them, evaluating them, and governing what they do. Gemini Enterprise Agent Platform is Google's, and Model Garden, Agent Search and the tuning and evaluation tooling sit within it.

This is where most engineering teams work. The platform handles what is common to every project so a team spends its time on what is not.

Agents

Systems that pursue a goal rather than answer a question. An agent plans steps, calls tools, reads results and adjusts. Building one means deciding what it may do, not only what it may say.

The business implication is the important part. An agent that can act creates risk a chat interface does not, because a prompt injection now reaches something that changes data rather than something that talks.

Applications

Finished products people use without knowing a model is underneath. The Gemini app, Gemini for Google Workspace, and the Customer Engagement Suite are applications in this sense.

Fastest to adopt, least to configure, and the layer where a whole organisation can get value in weeks rather than quarters.

Choosing a layer

The question is what you are actually trying to differentiate on.

If the answer is a workflow specific to your business, buy the application and spend your effort elsewhere. If it is a customer experience nobody else can offer, you need the platform and probably agents. If it is the model itself, you are in a very small group and should be sure.

The common error runs one way. Teams build at the model layer when an application would have done, then spend a year rebuilding retrieval, evaluation, and safety controls that were already available a layer up.

Common misconceptions

Buying at a higher layer means giving up control.

It means giving up control of things most organisations were never going to do well. What you keep is control of your data, your access rules and your use case, and those are the decisions that actually differentiate the result.

An agent is just a model with a longer prompt.

An agent plans a sequence, calls tools, reads what comes back and decides what to do next. That ability to act is what separates the layer, and it is why an agent raises security questions a chat interface never did.

The layers are alternatives and you pick one.

They stack. An application runs on an agent, which runs on a platform, which serves a model, which runs on infrastructure. The choice is how far down you go, not which one you use.

3 questions test this concept

A retailer wants a customer service assistant that answers from its own returns policy, checks an order in its warehouse system, and issues a refund where the policy allows it. Which layer of the gen AI landscape does this describe?

  • AModels, because a foundation model produces every response.
  • BAgents, because the system plans steps, calls tools and acts rather than only answering.
  • CInfrastructure, because it requires compute to serve requests.
  • DApplications, because customers interact with it directly.
Check whether it stuck.

One per page, with a worked explanation.

Start the set
Related material
Book
AI Engineering, On what a team actually owns at each layer.