Google Cloud Vertex AI

Google Cloud Vertex AI is Google's unified machine learning platform: hosted Gemini and partner model APIs, custom model training, orchestration pipelines, notebooks, a model registry, feature serving, and monitoring on one surface. Google is renaming it the Gemini Enterprise Agent Platform; the Vertex AI name still dominates the SDK and practitioner tooling.

itArtificial intelligence and machine learning

Don't Panic — Google Cloud Vertex AI

Vertex AI, honestly, is Google renting you its machine learning plumbing: the models (including Gemini, the one from the adverts), the training compute, the serving machinery, and a small museum of MLOps services, all under one bill. Google now prefers to call it the Gemini Enterprise Agent Platform. It renamed the platform while most of the world was still learning the old name, which tells you something about who this relationship serves.

Before it existed, running ML on Google Cloud meant stitching together separate services: one for training, one for hosting, notebooks somewhere else entirely. Vertex AI's whole pitch is that these are one platform now. The pitch mostly holds.

Three ideas carry the platform.

The two faces. One face is the modern one: call Gemini and partner models through an API, pay per token, tune them if prompting disappoints. The other face is classic ML: submit your own training container, let Google run it serverless, register the result. Same bill, very different workflows. Knowing which face your problem belongs to is half the navigation.

The endpoint. Serving runs through a resource called an Endpoint: deploy a model there and it can take predictions. Multiple models can share one endpoint with a traffic split, which is how careful teams roll out new versions. This is also where the platform's most expensive surprise lives, so it gets its own paragraph.

Deployed means paying. A model that is deployed to an endpoint bills node-hours continuously, whether ten million requests arrive or zero. The minimum replica count is one, and there is no scale-to-zero in the standard path. Every idle test endpoint you have forgotten about is a subscription you did not mean to start. The single most valuable Vertex habit is undeploying things the moment they stop being interesting. The pricing page states it plainly: you pay for each deployed model even with no prediction, and you must undeploy to stop.

What will surprise you? The quota system, in one word: shared. Gemini capacity follows a Dynamic Shared Quota, a phrase meaning there are no fixed limits to check, so throttling can arrive while the dashboard peacefully shows headroom. Guaranteed throughput exists and costs extra. Teams discover this during their first 429 storm, usually at the worst possible hour.

A gentler surprise: the bill can show line items for services you never knowingly enabled, because AutoML jobs quietly incur pipeline-run fees and metadata storage. Read the pricing fine print before the first experiment, not after the first invoice.

Where next? The Cheatsheet compresses the two faces into tables, the Reference tab starts with Google's own documentation, and the Timeline tab explains why your bookmarks say Vertex AI while the console says something else. The quiz is kind: it will not ask you to name the rebrand from memory. It will ask what an idle endpoint costs, which is the question that actually keeps money in your account.

Don't Panic — Google Cloud Vertex AI

Relevant careers