NEW DELHI, September 24, 2026, 10:24 p.m. IST — Google Cloud is expanding its AI infrastructure and agent-management stack in Brazil, pairing local processing for Gemini Enterprise with new GPU capacity, governed agent skills and security controls aimed at moving enterprise agents from pilot projects into production.
The company said at its Brazil Summit in São Paulo that Gemini Enterprise on the web will support in-country processing for Gemini 3.5 Flash beginning October 15. G4 virtual machines powered by Nvidia RTX PRO 6000 Blackwell Server Edition GPUs are already available in Google Cloud’s São Paulo region. Google also introduced Projects, Skills Registry and Agent Sandbox capabilities, alongside security updates for agent workflows.
The announcement matters beyond one cloud region because it brings several production concerns into the same release: where model requests are processed, where workloads run, how reusable agent capabilities are approved, and how autonomous actions are isolated and authenticated. For platform and DevOps teams, those controls are increasingly as important as model quality when an agent can execute code, browse the web or call business systems.
Local processing and local accelerators
Google said the October 15 change will add in-country processing for Gemini 3.5 Flash in Brazil to existing local storage support for agent workloads. The company framed the move as support for lower latency, disaster recovery and alignment with Brazilian data-sovereignty requirements.
That language needs a careful reading. Data residency is not a blanket guarantee that every Gemini Enterprise feature or model will remain in country. Google’s product documentation distinguishes data stored at rest from machine-learning processing and lists limitations for specific models and features in regional deployments. Teams with regulatory obligations should therefore validate the exact model, endpoint, grounding feature and logging path they intend to use rather than treating a regional setting as a universal control.
The G4 virtual machines are available now in São Paulo, according to Google. The instances use Nvidia’s RTX PRO 6000 Blackwell Server Edition GPUs and target high-throughput inference and real-time workloads closer to users and data in Latin America. Google separately said it expects to double its technical infrastructure in Brazil by 2030, covering compute, storage and high-performance AI processing. That is a long-term capacity plan, not a guarantee of immediate capacity for every machine type.
Google adds a control plane for agent skills and execution
The software announcements focus on the operational layer around AI agents. Projects gives individuals or teams a persistent workspace tied to a defined set of files or documents. Skills Registry provides a central catalogue where administrators can decide which packaged instructions, scripts and resources an agent, person or team may use. Agent Sandbox is an isolated managed environment where agents can compile and execute code, work through a command line, or operate a browser.
Together, those features resemble familiar platform-engineering patterns: versioned building blocks, policy-controlled reuse and isolated execution. The practical test will be whether organizations can integrate them with existing identity, audit, change-management and incident-response processes without creating a second control plane that drifts from the rest of the cloud estate.
Google’s wider Gemini Enterprise Agent Platform already includes Agent Identity, Agent Gateway, Agent Registry, runtime observability and production evaluation. Its July platform update described least-privilege identities for agents, policy enforcement for agent-to-tool traffic and continuous monitoring for behavioral drift. Readers building similar operational controls can also review GravityDevOps’ guides to LLMOps and retrieval-augmented generation.
Security updates mix general availability with previews
Google also announced changes under its Agentic Defense portfolio. Model Armor now supports native image screening, stronger detection of indirect prompt injection embedded in Google Workspace files and a generally available context window of up to 64,000 tokens. Exclude lists intended to reduce false positives and guardrail support for Anthropic models hosted through Gemini Enterprise are in public preview.
Auth Manager is generally available and is designed to handle credential management, token refresh and injection across complex OAuth flows. Support for JSON Web Tokens in agent-to-agent orchestration remains in public preview. Google also said Wiz will offer Brazilian organizations a complimentary external exposure assessment using its Red Agent system.
Those release labels are operationally significant. Preview controls should not be treated as production dependencies until teams have reviewed support commitments, regional coverage and failure behavior. Security scanners and guardrails also do not replace scoped identities, approval gates and durable audit logs, especially when an agent can make changes outside the model runtime.
What platform teams should verify
Before migrating an agent workload to the new regional path, teams should map each data flow separately: source documents, vector or database storage, prompt assembly, model processing, tool calls, telemetry and backup. They should then test which steps remain in Brazil, which features fall back to global services and what happens if the regional model or GPU capacity is unavailable.
Teams should also treat agent skills like deployable software. A useful baseline is to assign an owner, record provenance, review privileges, test in an isolated environment and define rollback before publishing a skill to a shared registry. Production observability should capture tool selection, latency, cost, policy denials and human approvals, without storing sensitive prompt content unnecessarily.
Google’s announcement confirms a meaningful regional expansion, but the broader analysis is still conditional: an integrated stack can reduce platform friction only if its residency, identity and security boundaries match the organization’s actual requirements. The immediate opportunity for Brazilian companies is lower-latency local compute and a clearer route to governed agent deployment. The remaining work is architectural verification, not simply enabling a feature.
Sources
Google Cloud detailed the Brazil expansion in its September 24 announcement. Availability and limitations were cross-checked against Google Cloud’s Gemini Enterprise data-residency documentation, its Agent Platform update, and the company’s security overview for Google Cloud and Wiz.
