InsightsGoogle Cloud
Why we build AI on Google Cloud
Google builds the chips, the Gemini models and the agent platform, writes no-training and indemnity terms into the contract, and ships the security controls.
- Written by
- CloudMagic engineering team
- Published
- Reading time
- 12 minutes
The short answer
We build on Google Cloud because one company designs the chips, trains the Gemini models and runs the agent platform they are served through, and its terms say it will not train on your data without your permission. Model Armor, Agent Identity, VPC Service Controls and customer-managed keys come from the same company that runs Google Threat Intelligence and Mandiant, and answers can draw on your own data, Google Search and Google Maps together. Before you commit, check that every feature you depend on is generally available in the location you need.
On Google Cloud, the company that trained the model also designed the chip it runs on, runs the platform your agents live on, writes the no-training and copyright terms into your contract, ships the controls that screen and isolate each agent, and investigates the attacks those controls are built against. That is why we build our clients' agents there. For a CIO it comes to three things: one agreement covers every layer an agent depends on, one company answers when two layers disagree, and the controls switch on in one place. This post gives the Google facts behind each reason, the checklist we apply on day one, and what to check before you commit. Product details were checked against Google's pages on November 14, 2026.
On names: Google's product page now calls Gemini Enterprise "Gemini for enterprise", but the docs and contracts still say Gemini Enterprise, so we do too. Gemini Enterprise Agent Platform, formerly Vertex AI, is where custom agents are built and run.
Why do we build AI on Google Cloud?
Google designs the chips, trains the Gemini models and runs the agent platform, and it writes its commitments on training data and copyright into the contract. The security controls, grounding services and residency options come from the same company.
| Reason | What Google publishes |
|---|---|
| One company builds the chips, models and agent platform | Ironwood TPUs; Gemini 4 Argon, Carbon, Hydrogen; Gemini 3.8 Flash; ADK, Agent Runtime, Agent Identity, Agent Gateway; 50+ managed MCP servers |
| The contract covers training and copyright | No training on your data without permission; indemnity for Gemini Enterprise, grounding and generally available models |
| Security is part of the stack | Model Armor in every Gemini Enterprise edition at no extra cost; Agent Identity; VPC Service Controls; CMEK; Google Threat Intelligence and Mandiant |
| Answers draw on your data and the world | Grounding with Google Search and Google Maps alongside your sources; BigQuery beside the models |
| Residency and published scale | US and EU multi-regions, allowlisted in-country regions, air-gapped options; nearly 32 billion tokens a minute |
Why does it matter that Google builds the chips, the models and the platform?
One company answers for every layer an agent runs on, and all of it sits under your Google Cloud agreement.
Google's Tensor Processing Units are, in Google's words, "the engine behind Gemini and Google's flagship billion-user products, including Search, Photos, and Maps." The seventh generation, Ironwood, is generally available with 9,216 liquid-cooled chips per pod. The eighth generation, with separate chips for training and inference, is announced and listed as coming soon.
The models are Google's own. Gemini 4 Argon and Gemini 4 Carbon are Google's frontier models for deep work on complex tasks. Gemini 4 Hydrogen is built for speed in everyday coding, agent workflows and tool use, and Gemini 3.8 Flash is Google's workhorse model. All four are generally available on Gemini Enterprise Agent Platform. In the Gemini Enterprise app, Carbon and Hydrogen stay off until an administrator turns them on.
On the platform, agents are built in code with the Agent Development Kit (ADK), hosted on Agent Runtime, and governed with Agent Identity, Agent Registry and Agent Gateway. Google runs more than 50 managed Model Context Protocol (MCP) servers, generally available or in preview, so agents reach Google services through a standard interface. Google also gave the Agent2Agent (A2A) protocol, which lets agents built by different teams work together, to a neutral open-source foundation in June 2025, so no single company governs it.
Will Google train on our data, and who carries the copyright risk?
Google's Service Specific Terms say Google will not use customer data to train or fine-tune any AI or ML model without your prior permission or instruction. Google's generative AI indemnity covers Gemini Enterprise, grounding with Google Search and Google Maps, and the Agent Platform API used with generally available model versions.
On Gemini Enterprise Agent Platform the training restriction applies to every managed model, generally available or in preview. Gemini Enterprise's security documentation says the same for the app: prompts, responses and connected enterprise data are not used to train Google's foundation models.
The indemnity has two parts, one for Google's use of training data and one for the generated output of Google's foundation models. Google's list names the Agent Platform API "used with generally available versions" of Gemini, Imagen and Veo, so a model in preview sits outside it, which is one reason we keep preview models out of production. Your lawyers should read the obligations in Google's Service Specific Terms; this is our engineering reading, not legal advice.
How is security built into the AI stack?
The controls an agent needs are Google Cloud features: a screen on prompts and responses, an identity per agent, a network perimeter and keys you hold. They come from the company that also runs Google Threat Intelligence and Mandiant.
Model Armor screens prompts and responses for prompt injection, sensitive data leaks and harmful content, and Google supports it on all Gemini Enterprise editions at no additional cost. The console setting covers the Gemini Enterprise assistant, Google-made agents and Workflow Builder agents. It does not screen custom ADK or A2A agents, so every custom agent we build calls Model Armor from its own code, as our post on stopping prompt injection in AI agents shows.
Agent Identity gives each agent its own cryptographic identity, which Google says cannot be impersonated and does not allow long-lived service account keys. Agent Gateway sits between agents and the tools they call and applies your access policies to that traffic.
VPC Service Controls puts a perimeter around the Gemini Enterprise app and its connected data; Google recommends a dry run before enforcing it. Customer-managed encryption keys (CMEK) let you hold the key for data at rest, and they protect only apps and connectors created after the key is registered.
Security Command Center's AI Protection keeps an inventory of your AI assets and, in the Premium tier, flags Agent Runtime agents with more permissions than they need. Google has also announced, as coming in preview, discovery of unmanaged agents and MCP servers on Cloud Run and GKE.
Google Threat Intelligence combines Google's view of attacks with Mandiant's, and Mandiant's M-Trends 2026 report draws on over 500,000 hours of investigations in 2025. One finding is why we switch the controls on before an agent sees real data: the median time between an intruder's first access and the hand-off to a second threat group fell from more than 8 hours in 2022 to 22 seconds in 2025.
Can answers draw on our own data, Google Search and Google Maps at once?
Yes. On Gemini Enterprise Agent Platform a model can be grounded in Google Search, Google Maps and your own data sources in the same request, and both Search and Maps grounding are on Google's indemnity list.
Grounding with Google Search connects the model to results from Google's search engine, and the response links to its sources. Grounding with Google Maps adds information on over 250 million places. Here is what one request looks like at Larrow Market, the made-up grocery chain from our storm-planning demo; the scenario and the figures are invented to show the shape.
A dispatcher asks: "The interchange at junction 14 closed an hour ago. Which of our depots on the northern route still hold pallet jacks for the Tuesday run?" Your own data answers the stock question: the agent's query against the stock table in BigQuery returns three depots with pallet jacks on hand, under the dispatcher's own permissions. Grounding with Google Maps supplies the place data for those depots and for the closed interchange, so the answer can say which two sit on the open side of it. Grounding with Google Search supplies this morning's closure notice from the highway authority's site, with the link. The answer the dispatcher reads names the two depots, shows each source beside the fact it supports, and stops there. The dispatcher decides what to move, and the agent drafts the message to the depot for the dispatcher to send.
Your own data is often already in BigQuery, in the same cloud as the models, and Gemini in BigQuery features are included in BigQuery pricing. Gemini Enterprise can also query BigQuery, AlloyDB, Cloud SQL and Spanner in place through federated query mode, in preview, using each user's own credentials, so an answer draws only on records that person can already see.
Can we keep data in our region, and will the platform keep up?
Yes, within limits you check feature by feature. Gemini Enterprise keeps data at rest in the US or EU multi-regions or allowlisted in-country regions, and Gemini Enterprise Agent Platform keeps ML processing in the US or the EU on jurisdictional endpoints.
Google recommends the global location unless you have security or regulatory reasons to choose another, because global gets the latest models and features. On the Agent Platform, requests to a global endpoint can be processed in any Google Cloud location and carry no residency guarantee.
For stricter rules, Google Cloud Data Boundary applies residency and access controls across more than 25 regional and regulatory boundaries. Google Cloud Dedicated runs on isolated infrastructure operated locally and is designed to keep running for up to twelve months if disconnected from Google. Google Distributed Cloud in air-gapped mode is accredited up to US Department of Defense Impact Level 6 and can run Gemini inside your own perimeter.
On scale, Google's CEO said in the Q3 2026 earnings remarks that Google's first-party model APIs were processing nearly 32 billion tokens a minute, and in the Q2 2026 remarks that nearly 90% of the Fortune 100 were using Gemini Enterprise.
What does this mean for how CloudMagic builds?
This is the checklist we apply on day one, in the order the settings have to land. The first four are expensive to change later.
| Order | Control | What we set | What we check before a user sees the agent |
|---|---|---|---|
| 1 | Location | Licenses bought in the project and location the app will live in: global for the newest models and features, US or EU multi-region where you need CMEK or Access Transparency (Google's log of actions its staff take on your content); an in-country allowlist request filed before anything else, because the build waits on it | The location on the license matches the location on the app |
| 2 | Keys | Your CMEK key registered before any app or data store exists | Every app and connector was created after the key, because anything created before it stays unprotected |
| 3 | Perimeter | VPC Service Controls in dry run, then enforced, with the data sources and egress domains it allows set in the managed organization policy constraints | The dry-run log shows no denied call the agent needs |
| 4 | Identity | The identity provider fixed now, because changing it later wipes your users' chat history | A test user who should not see a document cannot reach it through the agent |
| 5 | Screening | Model Armor on in the Gemini Enterprise console, and called from every custom agent's own code | The misuse set is refused and logged on both paths |
| 6 | Identity per agent | Agent Identity on each code agent, so it runs under its own attested identity with no long-lived service account key | No service account key sits in the repository or the runtime configuration |
| 7 | Evidence | Cloud Trace on with Google's OpenTelemetry attributes, Cloud Audit Logs on, the prompt-logging decision recorded | The trace for one test case shows every tool call and what the model received |
| 8 | Testing | The answer key written from your own past cases before the build; every case run at least four times beside a prompt-only version on the same model, with misuse tests | Every gate in our post on how we test an AI agent before it does real work passes |
Our agents prepare, draft and recommend, and a named person approves anything that leaves the system.
Workflows that need custom code follow our 30-day plan for one workflow on Gemini Enterprise, which shows where each setting lands week by week. Every engagement starts with a call with one of our engineers. Our two-week offers, such as Gemini Enterprise for a company's first teams, a contract first pass for legal and HR answers, start with a free readiness call and a statement of work that is already written. Custom builds start with a fixed-fee discovery that can end in a recommendation not to build, and after launch Agent Quality keeps every agent tested each month. Ask us whether Google funding applies to your project.
What should you check before you commit?
Check the launch stage and location support of every feature your plan depends on. Several facts above carry conditions.
| Check | What Google's pages say | What we do |
|---|---|---|
| Is it generally available? | Preview features come under Pre-GA terms; the indemnity lists generally available models; Access Transparency does not cover data in preview features | Keep preview features out of production |
| Do you need an in-country region? | It needs an allowlist from your account team, and the newest Gemini 4 models are served from the global region only, after you confirm a warning | File the request before the build starts; choose each agent's model against your residency rules |
| Does residency cover every feature? | No. Custom skills keep data at rest locally without in-region ML processing; Agent Runtime code execution keeps data at rest in the US multi-region but not the EU | Read the locations page row by row |
| What does web grounding keep? | For Grounding with Google Search, Google stores queries derived from prompts and the context sent with them, not associated with you or your users, for up to three days to debug the service, and that storage cannot be switched off; Google recommends Web Grounding for Enterprise where you need zero data retention | Use Web Grounding for Enterprise, which is also on the indemnity list, for regulated workloads; otherwise write the three-day retention into the data flow your compliance team approves |
| What is announced but not shipped? | Eighth-generation TPUs are coming soon; unmanaged agent discovery is coming in preview; federated query mode and Model Armor's inline integration with Agent Runtime are in preview | Plan on what is generally available today |
Frequently asked questions
Can we keep all our AI data in the EU?
For most features. Gemini Enterprise supports the eu multi-region and the Agent Platform keeps ML processing in the EU on jurisdictional endpoints, but some features, such as Agent Runtime code execution, differ.
Are we tied to Google's models?
Model Garden on the Agent Platform offers more than 200 models, Google's own and others, and our agents use open protocols, MCP for tools and A2A between agents. We start with Gemini because the indemnity covers generally available Gemini versions.
Who responds if one of our agents is attacked?
Your incident response provider, or Mandiant if you hold a Mandiant Retainer, which Google offers with 2-hour response times. We work alongside them and hand over the agent's logs and runbook.
Sources
- TPUs, Google Cloud
- Google models, Google Cloud
- Gemini Enterprise release notes, Google Cloud
- Introducing Gemini Enterprise Agent Platform, Google Cloud
- Google-managed MCP servers, Google Cloud
- Google Cloud donates A2A to an open-source foundation, Google for Developers
- Service Specific Terms, Google Cloud
- Zero data retention on Gemini Enterprise Agent Platform, Google Cloud
- Gemini Enterprise security overview, Google Cloud
- Generative AI indemnified services, Google Cloud
- Generative AI indemnification, Google Cloud
- Model Armor, Google Cloud
- Model Armor in Gemini Enterprise, Google Cloud
- Agent Identity, Google Cloud
- VPC Service Controls, Google Cloud
- CMEK, Google Cloud
- Access Transparency, Google Cloud
- AI Protection, Google Cloud
- Securing AI, Google Cloud
- M-Trends 2026, Google Cloud
- Mandiant Retainer, Google Cloud
- Grounding with Google Maps, Google Cloud
- Gemini Enterprise locations, Google Cloud
- Agent Platform data residency, Google Cloud
- Sovereign cloud, Google Cloud
- Q3 2026 earnings remarks, Google
- Q2 2026 earnings remarks, Google