AI is also an infrastructure problem
When a company decides to use AI, the conversation usually starts with the tool, the number of users and the promise of productivity. The physical part comes later: computing capacity, data centers, connectivity and energy. Brazil's Energy Research Company (EPE) put that base at the center of a panel on the advance of AI in Brazil.
The event took place on September 16, at Maravalley, in Rio de Janeiro. EPE published its account on September 21. At the meeting, the company's president, Thiago Prado, took part in a discussion on energy, connectivity and data centers as the physical base of AI at scale.
Scale has an invisible layer
According to EPE, as AI applications and models advance, the layers that support them become more evident: data centers, chip capacity and energy. The company also says that connecting electricity-intensive loads, such as AI data centers, requires expanding the transmission grid and technical, economic and socio-environmental planning.
The fact doesn't allow concluding that there will be an immediate price increase or energy shortage for small and midsize businesses. A scaling plan needs to consider services and vendors that carry physical costs, availability and regional concentration.
The cost doesn't end with the subscription
As a planning hypothesis for an SMB, the dependency can show up in usage limits, latency, downtime, storage, connectivity or variable billing. If the project only considers the tool's monthly fee, the calculated margin may be incomplete.
This applies especially to automations that run frequently, process long documents, query external databases or support operations that can't stop. The risk isn't just paying more. It's designing a process that doesn't work when the contracted capacity, the connection or the vendor fails.
The inventory also helps choose where redundancy makes sense. Not every task needs a second tool or a complex architecture, but critical processes need an alternative proportional to the impact of an interruption. In some cases, that means a temporary manual procedure; in others, a second source, a usage limit or a rule to reduce the volume processed.
The decision for the company
Before scaling, leadership can list what the process consumes: model calls, storage, integrations, network, review time and support. Then create three scenarios, current use, expected growth and peak, and record cost, response time, dependencies and contingency plan.
The test doesn't need to predict the future of global infrastructure. It needs to show whether the routine stays economically viable when use grows. The best decision may be to expand, limit, change the architecture or keep the experiment small until the numbers make sense.
Sources
Empresa de Pesquisa Energética, 21 Sep 2026. https://www.epe.gov.br/pt/imprensa/noticias/epe-fala-sobre-energia-e-avanco-da-inteligencia-artificial-no-techgrid-2026