The interesting question about Brazilian AI data centers used to be whether the grid could carry them.
That question has not gone away. But it has been overtaken by a different one. What kind of water, electricity and renewable-energy footprint are the new sites actually booking, and how is that footprint being counted?
In the last few weeks, Brazil has produced three signals on that front. None of them are about whether AI capacity gets built. All of them are about how the capacity gets accounted for.
Microsoft cools with water, on purpose
Folha de S.Paulo reports that Microsoft began operating AI data centers in Brazil using evaporation cooling towers, an older technology that relies on water evaporating to remove heat from servers.
The trade-off is direct. Evaporation cooling uses less electricity than newer closed-loop or air-cooled designs. It also consumes far more water. Microsoft has acknowledged elsewhere that closed-loop cooling can cut on-site water use to near zero, but it requires more power and more capital up front.
In Brazil, the company has chosen the energy-cheaper, water-heavier path for at least part of its AI footprint. That choice is not a scandal. It is a disclosure problem.
A reader looking at Microsoft’s global sustainability report cannot easily tell that a Brazilian AI cluster is using more water per kilowatt than a comparable U.S. cluster. The numbers roll up. Local water draw becomes a footnote inside a global power-usage and water-usage figure.
ByteDance’s $2 billion electricity contract
Brazil’s second signal sits on the energy side.
Folha reports that Omnia and Casa dos Ventos have signed an electricity-supply agreement worth roughly $2 billion for a data center being built in Brazil for ByteDance, TikTok’s parent.
The contract is a long-dated power purchase deal anchored on Casa dos Ventos’s wind and solar pipeline. For ByteDance, it is the energy backbone for what will eventually be an AI-capable site. For Brazil, it is one of the first explicit cases where a hyperscaler-style AI build is being financed alongside a renewable supply commitment of similar magnitude.
This is not a virtue play. ByteDance does not need a $2 billion contract to look green. It needs a $2 billion contract because AI training and inference are electricity businesses and the company wants known cost and known supply. Tying that to renewables is convenient on the climate page and convenient on the procurement page at the same time.
What it leaves open is the question of how the resulting electrons are counted in carbon reports — particularly in hours when the wind is not blowing.
The clean-energy accounting fight just got decided
The third signal is regulatory.
Folha reports that tech companies have won a dispute over the clean-energy accounting rules that apply to gas-powered data centers in Brazil.
The substance matters. Modern data centers that run partly on natural gas can pair that gas with renewable-energy certificates, biogas attribution or other accounting mechanisms to claim a low-carbon profile. The fight in Brazil was over how aggressive those claims can be — for example, whether emissions from on-site gas turbines can be offset by purchased renewable power booked against the same site.
Folha’s coverage says the industry side prevailed on the headline rule. Combined-cycle gas plants near AI data centers can keep using flexible accounting to support clean-energy claims. The decision does not eliminate disclosure. It does tilt it toward the operators.
For AI buyers, the practical result is that a “renewable” Brazilian data center may now be running on a mix of grid power, dedicated wind and solar contracts and on-site gas. Whether that fits a buyer’s own scope-2 commitments is a question they will have to answer themselves, with less help from Brazilian rules than they might have hoped.
The boom is still bottlenecked
None of this means Brazil’s AI build is racing ahead.
Folha’s separate reporting says AI data-center expansion is being held back by bureaucracy, labor shortages, energy access and equipment bottlenecks. Permits move slowly. Specialized electricians and HVAC technicians are scarce. Substation upgrades drag. Imported transformers, switchgear and chillers can sit in queues for months.
That bottleneck is the part of the story that has not changed. It is what kept Brazilian AI capacity below demand for most of 2024 and 2025. It is also what makes the sustainability accounting story now possible: the sites that do get built are large, expensive and politically visible, which is exactly where disclosure pressure lands.
What changed
The earlier Brazil AI infrastructure story was binary. Either the grid would carry the next wave of training clusters, or it would not. Reports focused on substations, transmission lines and the federal data-center incentive plan.
The new story is not binary. It is an accounting problem with several lines:
- Water use per site, especially where evaporation cooling is chosen over closed-loop designs.
- The structure of long-dated electricity contracts that finance both AI sites and renewable build-out.
- The clean-energy claims that can and cannot be booked against on-site gas generation.
- The local-versus-global rollup of all of the above into corporate sustainability reports.
For AI buyers running procurement in Brazil — banks, public-sector tenants, large industrial customers — that means a new column in the vendor questionnaire. Total power-usage effectiveness is no longer enough. Site-level water-usage effectiveness, contracted-versus-attributed renewable supply, and the treatment of any on-site gas now all need named answers, with the contract numbers behind them.
For investors, it means Brazilian AI infrastructure starts to look less like a pure capacity story and more like an ESG-data story. The same names — Microsoft, Amazon, Google, ByteDance, Equinix, Ascenty, Scala — will still dominate the build. What will increasingly separate them is who can produce defensible water and carbon numbers per site, not who can switch on another megawatt first.
Brazil has not solved AI’s sustainability problem. It is, however, becoming one of the clearest places in the world to watch the problem move from press release into contract footnote.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.