Google Cloud will add in four years the Brazilian capacity it took nine years to build

Axia Energia, the Brazilian customer Google cited at the summit, runs weather models on TPU clusters. By Google's account, it can predict severe weather up to ten days ahead and has cut service interruptions for millions of its customers year over year ^[15]. What one named user does, then, is heavy modeling on Google's accelerators. What the stage pitched was Gemini Enterprise. Milena Leal, Google Cloud's country manager in Brazil, said the expansion is meant to keep pace with its adoption and with demand for AI ^[6]. She said it will also add local storage and high-performance AI processing ^[6].

The accelerators have come a year at a time ^[12]. Flavio Franco runs Forward Deployed Engineering for Google Cloud in Latin America ^[12]. He said the company brought sixth-generation TPUs to Brazil at last year's summit and is adding Nvidia RTX Blackwell GPUs this year ^[12]. Google claims its new TPU family delivers three times the training performance and 80 percent better price-performance for inference ^[14].

The 2030 pledge is harder to plan against. Lopez measured it against Google's own history in Brazil: continuous building from 2017 to 2026, then a doubling of that capacity from 2026 to 2030 ^[2]. The plan repeats nine years of additions in four ^[1]. Spread evenly, Google would be adding about 2.25 times its average yearly build since 2017 ^[2].

In my view the unit is the weak point for a buyer, because only Google can count it. Lopez said that as processors get more powerful, the number of racks or buildings is no longer the primary indicator ^[9]. Franco said capacity will double in traditional infrastructure and in AI alike ^[13].

On Google's own costs, Lopez credited ReData, a recently enacted special tax regime for data centers, with significantly reducing operating costs, especially on component purchases ^[10].

The DCD account does not tie the plan to any named data-residency law, so the sorting has to start from a team's own obligations. I'd put each workload on two lines: whether a rule or contract requires its data to stay in Brazil, and whether it runs on hardware Google already lists in Sao Paulo. If the data can leave, Sao Paulo is one region among several, chosen on latency and price, and the pledge is background. When data must stay and the hardware is listed, this year's plan can go ahead on machines that exist. The hard case is data that must stay on hardware not yet offered locally. There the 2030 doubling is what the plan relies on, and the useful document is a capacity commitment from the account team with a region and a date on it. The price of that route is that the team has to forecast its own demand now, years before the capacity is due.