Google Vertex AI: the practical guide for SMEs

What is Google Vertex AI?

Vertex AI is Google Cloud's managed AI platform. Through it you use Google's Gemini models via API without running infrastructure yourself: prompt in, answer out, billed only for the tokens you use. In early 2026 Google bundled its generative AI under the name "Gemini Enterprise Agent Platform" – the product is still called Vertex AI and the API remains aiplatform.googleapis.com.

What sets it apart from the direct Gemini API (ai.google.dev, via API key): Vertex runs inside your Google Cloud project. Authentication via IAM instead of a shared key, billing through your existing cloud invoice, SLAs, private networking – and the ability to keep processing within the EU.

And Vertex isn't only Gemini: via Model Garden you get the same enterprise terms for partner models too – Anthropic's Claude, Meta's Llama, Mistral and others – as managed APIs. One contract, one access management, many models.

Why Vertex AI for SMEs? The GDPR argument

The biggest hurdle for AI in the German Mittelstand is rarely the technology – it's data protection. "Are we even allowed to send customer data to a US model?" stalls almost every project.

Vertex answers this through data residency – and this is where the crucial detail sits. Google separates two layers: where data rests (storage) and where it is processed (ML processing, i.e. the actual inference). For GDPR, processing matters most.

You control both through the endpoint. Call a regional endpoint such as europe-west3 (Frankfurt) or the EU multi-region (eu), and the processing of your prompts and responses stays in the EU. Google does not train its models on this data, and the Google Cloud DPA under Art. 28 GDPR applies, including the Standard Contractual Clauses.

The main pitfall: the global endpoint (global) does NOT. It routes your request dynamically to wherever capacity is free – worldwide, with no residency guarantee. Google now even recommends the global endpoint as the default for availability. So for real EU data residency you have to deliberately choose a regional or EU endpoint – and check that your model is available there. That switch is the heart of the matter.

The endpoint decides data residency

CurrentGemini models with EU data residency

These generally available (GA) Gemini models can run with EU data residency via a European regional endpoint (e.g. europe-west3). Prices are on-demand list prices per 1M tokens in US dollars (standard context up to 200K tokens).

Important for budgeting: the Gemini 2.5 family costs the same on a regional EU endpoint as it does globally – no surcharge. For the newer GA Gemini 3 models, however, Google introduces a premium for non-global (regional/EU) endpoints from 1 July 2026 – around 10% above the global price. Since GDPR residency requires exactly such a regional endpoint, a residency-compliant Gemini 3 deployment pays that surcharge.

As of 7 July 2026. Prices and availability change – the Vertex pricing page is what counts. Very large contexts (above 200K tokens) are billed higher for Gemini 2.5 Pro. Official Vertex pricing ↗

CurrentPrice range: cost per 1M tokens

What does Vertex AI cost?

By default Vertex bills on demand: you pay per token, split into input and output. No base fee, no idle cost – send nothing and you pay nothing. Summarizing two A4 pages with Gemini 2.5 Flash costs fractions of a cent.

The pleasant part for the EU: the Gemini 2.5 family costs exactly the same on a regional endpoint in Frankfurt as on the global one – unlike some competitors, you pay no surcharge for EU processing. The exception is the new GA Gemini 3 models, for which regional endpoints cost about 10% more from July 2026.

Two levers cut the bill further: context caching (recurring prompt parts are billed cheaper) and batch processing (asynchronous, around 50% cheaper). For predictable, high load there is Provisioned Throughput. For most SMEs: start on demand, measure your real usage, then optimize.

CurrentCost calculator: what does this cost per month?

Step by step: your first call on Vertex AI

How to get from zero to your first answer – in about 15 minutes. For standard use you don't need an access request.

1. 1
   Create a Google Cloud project & enable billing

   Sign in to the Cloud Console, create a project and enable billing. Everything else hangs off this project.

2. 2
   Enable the Vertex AI API

   Enable the Vertex AI API (aiplatform.googleapis.com) – in the console with a click or via gcloud services enable aiplatform.googleapis.com.

3. 3
   Choose an EU region as the endpoint

   For data residency, use a regional endpoint such as europe-west3 (Frankfurt) or the EU multi-region (eu) – NOT "global". Only then does processing stay in the EU for certain. Check that your model is available in that region.

4. 4
   Set up authentication (ADC)

   Locally: gcloud auth application-default login. In production: a service account with the "Vertex AI User" role (roles/aiplatform.user); the client library picks up the credentials automatically (Application Default Credentials) – no API key needed.

5. 5
   Install the SDK & send your first call

   pip install google-genai, create the client with vertexai=True, your project and location="europe-west3", and call generate_content – see the code below.

from google import genai

# Vertex AI · regionaler Endpoint in Frankfurt = Verarbeitung in der EU
client = genai.Client(
    vertexai=True,
    project="dein-projekt-id",
    location="europe-west3",  # Frankfurt – NICHT "global"
)

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Erklär in einem Satz, was Vertex AI ist.",
)

print(response.text)

The key to EU residency is location="europe-west3" (or the EU multi-region) instead of "global". Authentication runs through IAM/ADC, not an API key – that's the difference from the direct Gemini API.

Vertex AI, Bedrock, Azure or the direct API? A clear comparison

Vertex isn't always the best choice. The short decision guide:

The models are converging; the platforms differ in contracts, identities and regions. What decides is rarely "which model" but "where are you already – and where does processing stay?".

GDPR checklist for Vertex AI

To keep your Vertex deployment GDPR-compliant, watch these points:

• ✓Use a regional endpoint (e.g. europe-west3/Frankfurt) or the EU multi-region (eu) – NOT the global endpoint. And check that your model is available in the chosen region.
• ✓Data processing: the Google Cloud Data Processing Addendum (Art. 28 GDPR) applies, including the Standard Contractual Clauses for third-country transfers – confirm that Vertex AI is listed as a "Covered Service".
• ✓No training: Google does not use your prompts and responses to train its foundation models (Google Cloud Service Specific Terms).
• ✓Decide caching & retention deliberately: Gemini caches inputs for up to 24 hours by default. For zero storage, disable caching and enable Zero Data Retention (ZDR).
• ✓Know abuse monitoring: for standard models Google only logs flagged prompts (up to 90 days, in your region) – an opt-out is available via a form. For stricter requirements, combine ZDR and the abuse opt-out.
• ✓Harden security: enable CMEK (your own keys), VPC Service Controls and Access Transparency; minimize data in prompts and add the AI processing to your record of processing activities (a DPIA if needed).

CurrentLatest Vertex AI & Gemini news

What's happened around Vertex AI and Gemini recently:

• 3 March 2026Gemini 3.1 Flash-Lite in public preview on Vertex AI — the cost-efficient tierSource: Google ↗
• 19 February 2026Gemini 3.1 Pro in preview (Model Garden) on Vertex AISource: Google ↗
• 17 December 2025Gemini 3 Flash in public preview on Vertex AISource: Google ↗
• 18 June 2025Gemini 2.5 Flash & 2.5 Pro reach GA on Vertex AI; Flash-Lite in previewSource: Google ↗

Frequently asked questions about Vertex AI