Google Cloud Adds Brazil AI Capacity and Sets Local Processing for October

Brazilian companies can now run GPU-powered Google Cloud servers in São Paulo, and Google says a specified Gemini Enterprise web workload will gain in-country processing on October 15. The changes are part of a Brazil push that also includes a plan to double the company’s technical infrastructure there by 2030. The distinction for customers is what runs locally today, what arrives later, and which AI work the processing commitment covers.

Three different pieces of a local AI setup

The hardware is the part already available. Google Cloud’s G4 virtual machines, powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, are now offered in its São Paulo cloud region. A virtual machine is rented computing capacity; the GPUs provide processing power for demanding AI tasks. Regional access puts that capacity closer to the businesses and data Google wants to serve.

The processing change has a later date and a defined scope. Starting October 15, Gemini Enterprise on the web will support in-country processing for Gemini 3.5 Flash in Brazil. Google says it already supports customer-data storage there for agent workloads. Keeping stored data in a country and processing a model request there are different steps; this announcement specifies the latter for the named model and web product.

The largest commitment is farther out. Google expects to double its technical infrastructure in Brazil by 2030, adding localized compute, storage and high-performance AI processing. It has not presented that future capacity as already built. Together, the servers, processing change and buildout address different parts of running enterprise AI near Brazilian users and data.

The gap between trying agents and governing them

Google is pitching that local foundation alongside tools for managing AI agents, software that can carry out multistep tasks. A Google Cloud-commissioned survey found that 62% of the Brazilian organizations it surveyed were implementing or accelerating agent adoption. Only 17% had consolidated it with the governance needed to use agents across multiple core processes. Those are survey findings, not a measure of how Google’s new offerings will perform.

62%Implementing or accelerating

Share of surveyed Brazilian organizations implementing or accelerating AI agent adoption, according to the Google Cloud-commissioned study.

17%Consolidated with governance

Share that had consolidated adoption with governance for agents across multiple core processes in the same study.

Google introduced three Gemini Enterprise capabilities aimed at that operating problem. Projects gives individuals or teams a persistent space built around selected files and documents. Skills Registry lets an organization set policy for which reusable agent instructions and resources an agent, person or team may use. Agent Sandbox provides an isolated environment where agents can run code or use a browser. Each addresses a different question: what context an agent keeps, what it may do, and where it can execute work.

More than a server rollout

The Brazil announcement also reaches into data and security. Google says a new AlloyDB architecture is designed to support millions of collaborating AI agents while handling live queries and analytics without disrupting core production systems. That scale is a design claim, not a demonstrated customer result. For Brazilian organizations, Wiz plans a complimentary assessment that scans externally exposed systems, checks whether risks are exploitable and prioritizes findings for repair.

Google is adding people to the effort, too. It says it is expanding its Brazil-based forward-deployed engineering team, whose members work inside customer technical teams on applications. It also announced 10,000 free certification vouchers for developers in Brazil. Both steps target the work around deployment, rather than changing where an AI request is processed.

Google points to existing Brazilian customer work as evidence of demand: it says Bradesco uses Gemini Enterprise for fiscal, accounting and contract analysis, cutting document review from an hour to five minutes. That company-supplied example describes one workflow, not the effect of the newly announced capacity. The immediate test is whether customers can pair the available São Paulo hardware with workable controls; October 15 will bring a clearer test of the promised local processing.