Interactive version of a paper · ≈15 min read
AI Data Centers in Brazil: opportunities and risks
Artificial intelligence has turned data centers into a new kind of large consumer of electricity and water, concentrated in a few places and growing fast. This page uses charts and a simulator to explain the literature review published in the original paper: what is known about the impact of these facilities and what it means for Brazil.
The paper in one minute
The paper's conclusion: the sustainability of data centers in Brazil depends on where they are built, how they are cooled, the energy consumption model, whether indirect water is counted and whether data is transparent, as well as on clarity about the long-term commercial and strategic benefits.
Why the topic took off now
Until the mid-2010s, data centers used only slightly more energy each year: efficiency gains offset rising demand. From 2017 on, servers with GPUs for AI changed the curve. In the US, consumption stayed flat at ≈60 TWh between 2014 and 2016 and reached 176 TWh in 2023 (4.4% of the country's electricity). Worldwide, it was ≈415 TWh in 2024, growing ≈12% a year since 2017.
The light bar for 2028 is the projected range (325 to 580 TWh, or 6.7% to 12% of projected consumption). Source: LBNL [3].
2030 is a projection. Source: IEA [8].
Share of the total: the IEA projects data centers will account for less than 10% of global electricity demand growth through 2030. The problem is not the aggregate, it is the concentration: almost half of US capacity sits in five regional hubs. In Ireland, data centers already used 18% of national electricity in 2022.
Not every data center is the same
Enterprise
Run by a company for its own use, usually mid-sized.
Colocation
Rents space, power, cooling and connectivity to clients who bring their own servers. Modules of ≈1,000 m².
Hyperscale
Owned by those running internet services at massive scale. Modules of ≈2,800 m²; Amazon, Microsoft and Google hold more than half.
What changes in an AI data center, with GPU servers to train and run models? The paper points to three practical differences:
Racks with 72 GPUs (GB200) have already been announced. This density exceeds what air cooling can handle and pushes the industry toward liquid cooling. Source: LBNL [3].
How a data center is cooled
At heart, a data center is a building that turns electricity into computation and has to remove the resulting heat. It uses 5 times more energy per m² than an office, and cooling can reach ≈40% of total consumption. It is the main engineering lever over water and energy. Choose a technology:
- How it works
- Air pre-cooled by evaporating water.
- Water
- High, especially in hot climates.
- Energy
- Efficient in dry climates.
- Risks and limits
- Evaporation and blowdown require makeup water.
- How it works
- Heat exchange with outside air (economizer or dry cooler).
- Water
- Low.
- Energy
- Higher power use in hot climates.
- Risks and limits
- Performance depends on the local climate.
- How it works
- Liquid (e.g. 75% water, 25% propylene glycol) flows over the chips.
- Water
- Blue water consumption 31% (grid) to 50% (renewable) lower than air.
- Energy
- −15%.
- Risks and limits
- Copper components replaced with every server cycle.
- How it works
- Servers submerged in a dielectric fluid, usually a hydrocarbon.
- Water
- −45% (grid) to −80% (renewable).
- Energy
- −15%.
- Risks and limits
- Flammable fluid; oily waste; risk of soil contamination from leaks.
- How it works
- A fluid that boils at 30-50 °C and condenses back.
- Water
- −48% (grid) to −82% (renewable).
- Energy
- −20%.
- Risks and limits
- Fluids are usually PFAS, under regulatory restriction in the EU and the US: risk of a stranded asset.
Reductions relative to an air-cooled data center, per virtual core per year. Table 1 of the paper; LCA by Alissa et al. (Microsoft) [5].
Closed loop ≠ zero consumption
In evaporative systems, water evaporates and leaves salts behind, which require discharge and makeup. The metric that matters is blue water consumption: what leaves the river and does not come back.
Climate matters as much as technology
The same system uses more water and energy in the heat. The average US PUE fell from 1.6 (2014) to 1.4 (2023).
There is always a trade-off
Two-phase immersion saves water and energy but brings chemical risk (PFAS). Dry cooling saves water but uses more electricity, and with it more water and emissions at the power plants.
Obsolescence and e-waste
The building lasts ≈15 years; the servers, ≈6 (average service life at LBNL: 4.4 years through 2019 and 5.7 years in 2023). Within the same building, the server fleet is replaced two to three times. Wang et al. estimate 1.2 to 5.0 million tonnes of e-waste linked to large language models between 2020 and 2030; a circular economy could cut that by 16% to 86%.
Water: what exactly is measured
Before reading on, take a guess: of the water linked to US data centers, how much is used for on-site cooling?
Less than 8%. In 2023, US data centers consumed 66 billion liters directly, but almost 800 billion liters at the power plants that generated their electricity (4.52 L per kWh). More than 92% of the water footprint occurred outside the data center.
Source: LBNL [3]; share calculated in the paper.
≈550 MW of IT; equivalent to the consumption of more than 100,000 households. Source: Liang [12].
Much of the confusion comes from mixing metrics that measure different things. Tap each term:
Withdrawal
Volume taken from rivers, aquifers or the public network. It affects short-term availability and quality, but part of it returns.
Blue water consumption
Withdrawal minus what is returned in usable form (what evaporates or is embedded). It is the metric that can be compared across sectors.
Direct (on-site) water
Consumption for cooling the facility itself (site WUE). It affects the local basin.
Indirect (source) water
Consumption in generating the electricity used: evaporation at thermal plants and reservoirs. It is usually the largest share and affects other basins.
Drinking water vs. reclaimed water
Where the withdrawn water comes from. It changes competition with human supply, not necessarily physical consumption.
Green water
Rainwater consumed by crops and pastures. Relevant for agriculture; not comparable with withdrawn water.
Reuse helps, but does not solve everything. Replacing drinking water with treated effluent is already done at scale (Quincy, US, with ≈5% of makeup water still drinkable). But reuse changes the source, not the consumption: evaporated water does not return to the local cycle. And it does not touch indirect water, the largest share. In the US, a fifth of servers' direct water comes from basins with moderate or high water stress, and fewer than a third of operators worldwide measure their water consumption.
Simulator: a 100 MW data center
To give the numbers scale, the paper uses a hypothetical case: 100 MW consumed by IT equipment during all 8,760 hours of the year. Choose the cooling system and the efficiency (PUE) and compare with a Brazilian city.
The higher the PUE, the more energy goes to cooling and infrastructure beyond IT. The 1.2-1.4 range is the one used in the paper. Indirect water is held at 1.35 million m³: the change in electricity use is not reflected in that estimate.
In 2,500 m³ swimming pools per day (3.6)
at the data center in power generation
The water used on site equals ≈18% of the water that, by assumption, would not return to the basin in a city of 1 million people.
Bars: central values; whiskers: approximate ranges. The dashed line is the Redata limit (0.044). Source: LBNL [3].
The "Your scenario" bar follows the chosen PUE. Source: EPE (Empresa de Pesquisa Energética, Brazil's Energy Research Office) and IBGE (Instituto Brasileiro de Geografia e Estatística, Brazilian Institute of Geography and Statistics) [78, 79].
The power generation share is held constant only to compare the systems. Source: LBNL [3] and Liang [12].
See the simulator data as a table (Tables 3, 5 and 6 of the paper)
| System | Site WUE (L/kWh of IT) | Million m³/year | Pools/year | Pools/day | Equivalent people |
|---|---|---|---|---|---|
| Evaporative tower + water-side economizer | ≈2.2 (1.7-2.8) | 1.93 | 771 | 2.1 | ≈29 thousand (23-37 thousand) |
| Air-side economizer + adiabatic (water-cooled chiller) | ≈0.5 (0-1.5) | 0.44 | 175 | 0.5 | ≈7 thousand (0-20 thousand) |
| Direct-to-chip liquid + evaporative tower | ≈2.0 (1.6-2.5) | 1.75 | 701 | 1.9 | ≈27 thousand (21-33 thousand) |
| Direct-to-chip liquid + dry cooler with adiabatic assist | ≈0.1 (0-0.3) | 0.09 | 35 | 0.1 | ≈1.3 thousand (0-4 thousand) |
| Direct-to-chip liquid + waterless dry cooler (closed loop) | ≈0 | 0.00 | 0 | 0.0 | ≈0 in operation |
| Redata legal limit | 0.05 | 0.044 | - | - | ≈0.7 thousand |
| Water associated with electricity generation (Liang [12]) | - | 1.35 | 542 | 1.5 | ≈21 thousand |
Equivalent people: annual volume only, 180 L per person per day (≈65.7 m³/year, SINISA (Sistema Nacional de Informações em Saneamento Básico, Brazil's National Sanitation Information System) 2025). It does not represent supply capacity or the same basin. At a continuous 100 MW, 1 L/kWh corresponds to 0.876 million m³/year. Electricity: 876 GWh of IT × PUE. City of 1 million: 2,656 GWh/year (all consumer classes) and 10.9 million m³/year of water that, by assumption, would not return to the basin. These are LBNL scenarios for US climates, not forecasts for a Brazilian facility.
Does comparing with other sectors help?
Globally, AI is small next to food production. Food production emits ≈17.3 Gt of CO₂eq a year, 57% of it attributed to animal-based food. AI would have emitted between 32.6 and 79.7 Mt in 2025, in the range of New York City. Choose the AI estimate and count the squares:
AI in 2025 1 square
Animal-based food (≈9.9 Gt of CO₂eq, 57% of the 17.3 Gt from food production) 120 squares
The grid has 300 squares: with the high AI estimate, the empty ones show how far the ratio would go with the low one. The ratio of 120 to 300 times should be read as an order of magnitude. The food figure includes methane, nitrous oxide and land-use change and refers to ≈2010; the AI figure excludes the carbon from manufacturing the servers. Sources: Xu et al. [21] and de Vries-Gao [7].
But the risk is local
Livestock demand is diffuse; a data center's is concentrated and continuous, and the sector's global consumption has grown ≈12% a year since 2017. What matters is the basin and the grid where it is built.
Different reach
Livestock supplies 16% of the world's calories and 36% of its protein. Generative AI was used by 18.8% of people aged 15 to 64 (≈1 billion) in June 2026. These are different measures, and neither measures social benefit per m³ of water.
Beware of conversions
Even peer-reviewed papers use fragile factors: one study compares the São Paulo hub with beef using ≈1,600 m³/t, inconsistent with the reference value of 15,400 m³/t.
Energy: the IEA projects that, by the end of the decade, the US will spend more electricity on data centers than on producing aluminum, steel, cement, chemicals and all other electricity-intensive goods combined. But that is electricity only: heavy industry burns more fuel directly. Food and computing are not interchangeable, and weighing them is a normative choice, not an empirical one.
Where Brazil stands
Relevant in Latin America, small worldwide. Source: World Bank [4].
A renewable power mix, but dependent on rain
Emissions from fossil thermal plants in the interconnected system were 75% higher than in 2020. Source: IEMA (Instituto de Energia e Meio Ambiente, Energy and Environment Institute) [33].
Some Amazon lowland dams can exceed fossil plants. Reservoirs also consume water through evaporation. Source: Almeida et al. [13].
Rare earths: the trump card that does not become a chip on its own
Reserves in million t of oxides. Source: USGS [24].
The USGS rounds Brazil's reserve to 11 Mt, behind China (44 Mt); the ANM (Agência Nacional de Mineração, National Mining Agency) records 11.4 Mt. There are 2,777 mining claims in 23 states. Permanent magnets account for ≈35% of the volume but 83% of the value of rare earths.
In Minaçu (Goiás), Serra Verde has produced since 2024 a carbonate with Nd, Pr, Dy and Tb, sold for 15 years to a vehicle backed by the US government. In Lagoa Santa (Minas Gerais), the MagBras project is testing the "mine to magnet" chain at pilot scale.
Reserves and carbonate do not imply making GPUs. Their strategic importance can be used to negotiate refining, research, training and industry in the country, but those gains depend on negotiated commitments.
The risks for Brazil
The 2014-2015 water crisis in the Southeast put 40 million people at risk; in São Paulo, network pressure was reduced and people went without water for hours or days. Attribution to climate is disputed, but both sides agree: vulnerability comes from variable supply + growing demand, and large continuous loads add to demand.
The São Paulo hub (≈550 MW of IT) lies in the Alto Tietê basin, with high to critical water stress, which supplies more than 20 million people. In a drought with the 2021 power mix, indirect consumption rises 27.2%.
The result depends on the method for attributing hydropower water; electricity consumption was estimated, not measured. Source: Liang [12].
The 2015 drought cut hydropower generation by 110 TWh. Brazilian generation follows a 10- to 15-year cycle, and dry periods tend to coincide with supply crises; in 2015, the extra cost of thermal plants was passed on to consumers.
A new continuous load connected during a drought tends to be served by the marginal source, often a thermal plant. Renewable energy contracts improve the accounting balance, but do not decide which plants are dispatched. Nuclear is an option to examine, but it only displaces thermal plants if it is new generation.
Components highlighted by the authors within the total loss. Source: Hunt et al. [17].
In February 2026, the ONS validated 43 requests for access to the Rede Básica (the high-voltage transmission grid), 38 of them from data centers, totaling ≈7,040 MW; São Paulo accounted for 20 requests and 3,914 MW.
Not every request becomes consumption. In the ONS projection, data centers go from 304 average MW in 2026 to 3,457 average MW in 2030, ≈3.5% of the system, close to the 4.4% in the US in 2023. Electric cars add further demand: from 627 GWh (2025) to 7.8 TWh (2035).
Decree No. 12.772/2025 provides for a competitive process when requests at the same connection point exceed capacity. And concentrating more than half of the new load in a single state calls for geographic redundancy, not just physical protection.
Load requested by data centers (Feb 2026)
ONS projection: average consumption of data centers on the Rede Básica
Sources: ONS [83, 23].
A data center's life cycle emits regulated pollutants (PM2.5, SO₂, NOₓ) through three routes: server manufacturing, fossil generation and backup diesel generators. The evidence is from the US:
No Brazilian study was found on backup diesel generators at data centers or on the health effects of the thermal plants that serve this load: it is a research gap. Source: Han et al. [9].
The Redata (Law No. 15.504/2026) is a five-year regime that suspends federal taxes on equipment, but some of these benefits only apply until December 31, 2026. Estimated tax expenditure of R$ 5.2 billion in 2026, in exchange for commitments:
- 100% renewable or clean energy
- Water efficiency: WUE ≤ 0.05 L/kWh, measured annually
- 2% investment in research, development and innovation
- At least 10% of capacity offered to the domestic market, which can be replaced by additional investment in research, development and innovation (it is not a guaranteed reserve)
- Domestic supply and research commitments 20% lower in the regions specified by the law
The gaps pointed out by the paper:
- It does not count indirect water. Site WUE ignores the water used in generation: more than 90% of the total in the US and 46% in the São Paulo hub.
- "Renewable" without criteria. If contracts and certificates count, the physical load can be served by thermal plants during droughts.
- It trades water for energy. Reaching 0.05 L/kWh in a tropical climate requires dry cooling, which uses more electricity on hot days and can shift the impact to indirect water and emissions.
To justify an AI-oriented expansion, one has to look at what it delivers:
This does not show that AI is useless; it shows that computing capacity does not, by itself, guarantee benefit. Water and energy costs stay local; the benefits may go far away. No source calculates the net benefit of Brazil hosting infrastructure aimed at foreign demand. Each project should weigh jobs, tax revenue, research and data autonomy against the local costs of water, energy, grid and incentives.
The opportunities
Land and location
Location is the single variable with the largest effect on water and carbon: in the US, the water footprint of the same servers varies within this range depending on where they are. Brazil can choose.
Energy
Emissions when running on renewables instead of the average US grid. Wind and solar are already almost a quarter of generation, and one study suggests that wind in the Northeast complements the hydropower basins during droughts.
Water
Blue water consumption with liquid cooling compared with air; reusing effluent replaces drinking water almost one for one. The country can require this as a condition for incentives.
Investment and grid
Reinforcements recommended by the EPE would open ≈4 GW of connection margin in São Paulo. The sector already attracts international capital, provided the costs of the reinforcements are defined.
Jobs and training
The greatest potential lies in construction and technical training: electricians, HVAC, infrastructure engineering. The R&D commitment can fund training, if well targeted.
Capacity for the region
Accessible computing, with GPUs, viable prices and training, for Brazil and for neighbors that today cannot benefit from AI for lack of infrastructure. A possibility, not an automatic effect.
Typical facility of ≈20,000 m²; half of the 50 permanent jobs are outsourced. In a typical US county, Brookings estimates 100 to 200 jobs, and wages barely change. Sources: JLARC [29] and Brookings [28].
Quality of life: the PBIA plans applications in health, education and public services, but no evidence was found that hosting a data center improves the lives of the local population. The benefit depends on use, not on hosting.
What the paper proposes to measure and require
Brazil has favorable conditions: 86.6% renewables in electricity supply in 2025, land to choose sites and planning institutions. For this to turn into development, licenses and incentives for large data centers should require:
- Full water accounting: disclose consumption at the facility and in electricity generation.
- Physical carbon: emissions based on the energy actually delivered to the region, not only on certificates.
- New energy: contract additional clean generation, with production and consumption hours close to each other.
- Basins in drought: assess the water stress of the local basin and of the basins that supply the energy, under drought scenarios.
- Crisis plan: provide for temporary consumption cuts during crises, as in Ireland.
- Diesel and thermal plants: restrict and monitor diesel generators and the use of thermal plants.
- Servers: targets for reuse and responsible disposal.
- Real access: disclose the GPUs available to the domestic market, prices, users served and results.
- Open data: a public repository of anonymized operational data.
Quick glossary
PUE
Power Usage Effectiveness: total facility energy ÷ energy delivered to IT. It measures the infrastructure, not server efficiency: an excellent PUE can coexist with enormous consumption.
WUE
Water Usage Effectiveness: liters of water per kWh of IT. "Site" WUE does not include the water consumed at power plants.
TWh, GWh, average MW
1 TWh = 1,000 GWh = 1 million MWh. "Average MW" is the mean power over the year: 100 average MW × 8,760 h = 876 GWh.
Rede Básica
The high-voltage transmission lines of Brazil's interconnected system, operated under ONS coordination. Large data centers request direct access to it.
Additionality
When the contracted energy comes from new generation. Reserving the output of an existing plant may just push other consumers toward thermal plants.
Location-based vs. market-based
"Location-based" emissions use the physical mix of the connected grid; "market-based" ones use the certificates purchased. Local pollution cannot be offset with clean energy generated elsewhere.
GPU
Graphics processing unit: the processor used to train and run AI models, with much higher power and heat than an ordinary server.
Redata
Regime Especial de Tributação para Serviços de Datacenter (Special Tax Regime for Data Center Services, Law No. 15.504/2026).
Original paper and how to cite
This page summarizes and illustrates the paper, which is written in Portuguese; the numbers come from its tables and text. In case of any discrepancy, the original prevails: it contains the 109 references, the full methodology and all the caveats.
APA
Damico, J. (2026). Data Centers de IA no Brasil: Oportunidades e Riscos [AI Data Centers in Brazil: Opportunities and Risks] (Version 1). Zenodo. https://doi.org/10.5281/zenodo.23092193
ABNT (Associação Brasileira de Normas Técnicas, Brazilian Association of Technical Standards)
DAMICO, José. Data Centers de IA no Brasil: Oportunidades e Riscos. Zenodo, 2026. DOI: 10.5281/zenodo.23092193. Disponível em: https://zenodo.org/records/23092193.