The cloud metaphor was a lie
When marketing invented the word "cloud" to describe data centers, they did a brilliant job of abstraction. Clouds are ethereal, light, invisible. You don't think of clouds consuming electricity. You don't imagine clouds sucking millions of liters of water per day. You don't associate clouds with blackouts.
But data centers don't float. They weigh. And the more we use AI, the harder it becomes to ignore that weight.
The numbers nobody wants to show
While Microsoft, Amazon, and Google announce investments of tens of billions of dollars in AI infrastructure in India and Brazil, press releases talk about "progress," "innovation," and "economic development." What they don't mention:
A single large-scale data center consumes the equivalent of a city of 50,000 inhabitants. Training GPT-4 consumed enough energy to power 1,000 American homes for a year. Every query you make to ChatGPT consumes 10x more energy than a Google search.
Data centers use water for cooling — a lot of water. An average facility consumes between 3 and 5 million liters per day. In 2022, Microsoft's data centers consumed 6.4 billion liters of water globally. The number increased 34% from the previous year, driven by the AI explosion.
Despite "net zero" promises, most data centers still run on power grids dependent on fossil fuels. In India, where Microsoft plans to invest $3 billion, 75% of electricity comes from coal.
Brazil as a case study
Brazil has become a preferred target for data center expansion for obvious reasons: cheap energy (though unstable), available land, and flexible regulation. But the bill is starting to come due.
In 2024, the data center sector already represents an identified risk to domestic energy supply. This isn't conspiracy theory — it's technical analysis from ONS (National Electric System Operator). When a single data center can consume the equivalent of a small city, and dozens are being built simultaneously, the math doesn't add up.
The problem is structural: Brazil's power grid wasn't designed for this type of concentrated load. Unlike traditional industries that can reduce consumption during peak hours, data centers run 24/7 at maximum capacity. They are inflexible consumers in a grid that depends on flexibility to avoid collapse.
The silent resistance
While governors celebrate investment announcements, local communities are beginning to resist. In India, protests against data centers have become common in water-scarce regions. In Chile, projects were canceled after public pressure over water consumption in drought zones. In Ireland, the government imposed a moratorium on new constructions after data centers came to consume 18% of all the country's electricity.
In Brazil, resistance is still nascent, but the signs are there. In Fortaleza, residents question the construction of data centers in a region facing periodic rationing. In São Paulo, the industrial sector already competes with data centers for energy during peak hours.
The dilemma nobody wants to admit
Here's the inconvenient truth: there is no "clean" AI at current scale. Every advance in language models, every new ChatGPT feature, every generative AI startup — all of this translates to more servers, more energy, more water, more pressure on infrastructures already operating at their limits.
Big Tech knows this. That's why they invest in energy efficiency research and buy carbon credits. But these efforts are incremental, while demand growth is exponential. It's like trying to empty the ocean with a bucket while the tide rises.
When Microsoft builds a data center in Ceará, the benefits (technical jobs, taxes, "modernization") go to a specific group. But the cost — more expensive energy, blackout risk, water diverted from agriculture — is distributed to the entire population. It's the old story of socializing losses and privatizing profits, now with a new technological veneer.
What they're not telling you about "renewable energy"
Big Tech loves to announce that their data centers run on "100% renewable energy." This is technically true and practically a lie.
The trick works like this: the company buys renewable energy certificates (RECs) equivalent to what it consumes. But the energy that actually feeds the data center comes from the local grid — which in Brazil is 60% hydroelectric, 10% wind and solar, and 30% thermal (gas and coal). The certificate is a piece of paper; the electrons reaching the server are the same ones reaching your home.
In other words: when you use ChatGPT at night, part of that energy comes from gas plants. Microsoft may have a certificate saying it "offset" that emission by buying credits from a wind farm in Northeast Brazil, but physics doesn't work that way. The CO2 was emitted. The water from the thermal plant was heated and returned to the river.
The future we're building
There's an optimistic scenario: advances in energy efficiency, more robust power grids, waterless cooling, lighter AI models. All of this is possible and being researched.
But there's the realistic scenario: AI demand will grow faster than our ability to make it sustainable. Over the next 5 years, data center energy consumption is expected to triple globally. No amount of solar panels compensates for growth of this magnitude.
What does this mean for Brazil? Probably more blackouts during peak hours. More expensive energy for residential consumers while data centers negotiate special rates. Water conflicts in semi-arid regions. And growing dependence on critical infrastructure controlled by foreign companies.
The choice we didn't make
Nobody asked Brazilians if they wanted their power grid overloaded to train AI models that primarily benefit American companies. Nobody held a referendum on diverting water from agriculture to cool servers. These decisions were made in boardrooms in Seattle and Redmond, ratified by governors eager for photos with Big Tech logos.
This doesn't mean data centers are inherently bad. It means we're making long-term decisions without public debate, without transparent cost-benefit analysis, without asking who wins and who loses.
The "cloud" continues to be sold as something light and abstract. But somewhere in Ceará, a transformer is overheating. Somewhere in Minas Gerais, a reservoir is dropping. And somewhere in São Paulo, a family will pay more on their electricity bill so someone in New York can ask ChatGPT to write an email.
AI promises to solve the world's problems.
It would be ironic if, in the process, it created some new ones.
Technological progress has always had environmental costs. The Industrial Revolution polluted rivers. Cars polluted the air. Now, the AI revolution is pressuring power grids and water resources.
The difference is that this time we know in advance. We can't claim ignorance. The question is whether we'll do things differently — or whether we'll repeat the same old pattern, leaving the cost to those who can least afford to pay.