When we replace a cellphone, we know that the old device has to go somewhere. It might end up in a drawer, be donated, taken to a collection point, or, unfortunately, thrown into the regular trash. Either way, we see the object and feel the weight of deciding what to do with it. But what about the GPU [1] that runs the Artificial Intelligence (AI) we use every day? And what about the drive where the photos we upload to the “cloud” are stored? Where do they go when they stop working?

The word “cloud” is one of technology’s most successful metaphors. It suggests something light, ethereal, weightless, and without an address. But the cloud does have an address. It is made up of enormous warehouses filled with servers, cables, circuit boards, drives, batteries, and cooling systems. And, like all electronic equipment, this hardware breaks down, ages, and is eventually discarded. We need to pause and think about this, especially now that Brazil has decided to fully enter the race for hyperscale data centers [2].

A Problem That Is Already Big

To start the conversation, we draw on the article E-Waste, Artificial Intelligence, and the Law: Toward a Circular and Sustainable Digital Economy, by Ibrahim Badawi, a researcher at Seattle University. The author begins with a stark diagnosis: electronic waste is the fastest-growing waste stream in the world. It contains valuable metals, but also toxic substances such as lead, mercury, and cadmium. The laws that are supposed to address it remain fragmented and timid.

The numbers confirm the scale of the problem. According to the Global E-waste Monitor 2024, humanity generated 62 million tonnes of electronic waste in 2022. That is enough to fill 1.5 million trucks, which, if lined up, could circle the Equator. Of that total, only 22.3% was documented as collected and recycled. In other words, for every four kilograms of e-waste, less than one has a known destination. And the amount is expected to reach 82 million tonnes by 2030. Brazil is no bystander in this story: it generates around 2.4 million tonnes per year and ranks second in the Americas, behind only the United States.

One of the most interesting points in the article is that it does not treat AI merely as a solution. Badawi shows that AI can help track products, predict failures, and automate the dismantling and sorting of waste. He also recognizes, however, that AI itself is part of the problem because it depends on GPUs, TPUs [3], servers, and data centers that are being replaced at an increasingly rapid pace. He invokes the Jevons paradox: when a technology becomes more efficient, the tendency is not necessarily to use less of it, but to use more. Efficiency gains are ultimately absorbed by increased consumption, while the underlying infrastructure continues to expand.

How Data Center Hardware Retires

This is where a technology perspective comes in. There is a common assumption that data center equipment is discarded only when it breaks down. The evidence, however, points to something more concerning: equipment leaves service through two paths, and neither favors a second life.

The first path is failure, and it is frequent. When Meta published the technical report on the training of its Llama 3 model, carried out on a cluster of 16,000 GPUs, the numbers drew attention. Over a 54-day period, there were 466 interruptions, 419 of them unexpected. About 78% of these unexpected interruptions were attributed to hardware problems, with GPUs representing the largest category, accounting for 58.7% of them. A data center GPU operates under heavy loads, generates substantial heat, and is subjected to constant thermal stress. It is not equipment designed to last for decades.

There is even a widely cited estimate that data center GPUs may last between one and three years, depending on utilization. The exact figure is debatable. What is not is that this hardware has a short service life compared with that of a household appliance.

The second path is disposal without any failure at all. This may be the more serious one because it is a choice. A study published in Resources, Conservation and Recycling found that, because of data security concerns, most end-of-life hard drives are shredded even when they are still functional. The authors also point out that reuse is hindered by users’ lack of trust in data-erasure methods other than physical destruction.

In other words, a perfectly functional drive becomes scrap because destruction is simpler, cheaper, and legally more comfortable than securely erasing the data. Technical standards for data sanitization, such as NIST 800-88, already exist and allow data to be securely erased so that equipment can be reused. What is missing is an incentive to use them.

Economics also pushes organizations toward replacement. A study in Nature Computational Science on e-waste from generative AI estimates that this waste stream could accumulate between 1.2 and 5 million tonnes between 2020 and 2030, with the situation potentially worsening as servers are rapidly replaced to reduce operating costs.

What is most concerning is the speed of growth. In 2023, the figure was around 2,600 tonnes. Even under the most optimistic scenario, by 2030 there would be 400,000 tonnes per year—a roughly 150-fold increase. Under the worst-case scenario, the figure would reach 2.5 million tonnes per year, nearly a thousand times more. To put that into perspective, that is almost equivalent to the total amount of e-waste Brazil discards today in a year—around 2.4 million tonnes, including cell phones, computers, TVs, and refrigerators.

The same study brings some good news: circular economy strategies throughout the AI supply chain could reduce e-waste generation by 16% to 86%. The problem, then, is ultimately one of a lack of legal requirements.

Brazil Wants to Become Latin America’s Cloud Hub

The country is attracting data centers at an accelerated pace, and the justification is always the same: abundant renewable energy, submarine cables, and a strategic location. Brasscom projects that the sector’s capacity will grow 2.5-fold, from 1,063 MW in 2025 to 2,629 MW in 2030, with most of the investment—R$299.5 billion—going toward equipment. The megawatt (MW) is how the sector measures the size of a data center: the greater the available electrical capacity, the more equipment it can keep running.

According to Brasscom’s own calculations, each new MW installed requires US$ 10 million in infrastructure and US$30 million in ICT equipment.Three-quarters of the money goes toward servers, GPUs, and storage drives—the very equipment that eventually becomes electronic waste. And this is not a one-time replacement: according to PwC, servers, GPUs, storage, and networking equipment will generally need to be replaced every four to six years, generating recurring investment. Put simply: every megawatt we celebrate today is a shipment of electronic waste scheduled for a few years from now.

The most emblematic example is here in Brazil’s Northeast. At the Pecém Complex in Ceará, ByteDance, the owner of TikTok, is building its first data center in Latin America. The initial investment is R$50 billion, making it the largest project of its kind ever announced in the country, with construction beginning in January 2026 and operations expected to start in the third quarter of 2027. The initial capacity will be 300 MW, comparable to the electricity consumption of a city of around 2.4 million people.

The data center is focused on exporting its services and will be located in the Pecém Export Processing Zone. The equipment arrives imported and with tax incentives, processes data largely from outside the country, and, when it fails or becomes obsolete, will be discarded here. The service leaves. The waste stays.

Badawi’s article devotes an entire section to environmental justice. He describes how e-waste from wealthy countries ends up in countries in Africa and Asia, where vulnerable communities—particularly women and children—bear the burden of contamination. In Brazil’s case, the waste does not need to arrive by ship or through illegal channels. It is generated here, legally, and with tax incentives. The logic is the same: some reap the benefits of the technology while others are left with the waste. Only the packaging has changed.

What Our Laws Say (and What They Don’t)

It is useful to look at laws the way we look at a system: what they cover, what they leave out, and where the gaps are.

Brazil’s National Solid Waste Policy (PNRS), Law 12,305/2010, contains sound principles. Article 9 establishes a hierarchy of priorities that begins with avoiding waste generation and proceeds through reduction and reuse before reaching recycling. Article 31 states that manufacturers and importers should invest in products designed for reuse and that generate the smallest possible amount of waste. On paper, this is exactly what Badawi’s article advocates. The problem is that the law stops at the level of principle. It says nothing about minimum durability, nothing about a right to repair, nothing about spare-parts availability, and creates no mechanism for measuring whether a product was actually designed to last.

Decree 10,240/2020, which regulates the reverse logistics of electronic products, is even more revealing. Right in its title, it states that it applies to household consumer products. And Article 5 makes clear what is excluded: products for corporate use, products used in production processes, and large volumes from large waste generators. A data center falls under all of these exclusions. For such cases, Article 6 says that reverse logistics “may be governed contractually” between companies. In other words, it is left to them. The targets established by the decree, which reach 17% collection, apply to the cell phone forgotten in a drawer—not to the thousands of servers in a 300 MW facility.

So, who handles a data center’s waste? In practice, it falls under the general rules governing corporate waste management plans, provided for in Article 20 of the PNRS. These plans address disposal, not reuse, durability, or traceability. There is no public target, no mandatory reporting on hardware end-of-life, and nothing that says, “before shredding it, try to reuse it.”

In Congress, there are proposals addressing planned obsolescence, such as Bill 7,875/2017 in the Chamber of Deputies, which still needs to go through the Consumer Protection and Constitution and Justice committees before proceeding to the Senate. There is also Bill 805/2024 in the Senate, which seeks to curb planned obsolescence and establish a right to repair by amending the Consumer Protection Code. These are important developments, but notice the limitation: both focus on consumers; neither addresses infrastructure.

Redata and the Opportunity That Is Still Open

Law 15,504, recently enacted, creates the Special Tax Regime for Data Center Services (Redata) and suspends federal taxes on equipment intended for the installation, modernization, and expansion of data centers in Brazil. The Brazilian state is, quite literally, facilitating the large-scale entry of hardware.

The requirements include meeting sustainability criteria that are yet to be defined by regulation, using “clean” or renewable energy sources, reporting a water-efficiency index for cooling systems, and publishing a sustainability report containing the efficiency index, energy sources, and other indicators. Water and energy are there, but the fate of equipment at the end of its useful life is not. Experts had already considered the bill vague in its criteria and pointed to several neglected environmental and social issues, including e-waste disposal.

Looking at the glass as half full, there is still a window of opportunity. The regulations will define the sustainability criteria, and this is where hardware end-of-life can—and should—be included.

Possible Paths Forward

Badawi’s article puts forward proposals that, with some adaptations, could fit the Brazilian context very well. They can be organized into six points:

1. End-to-end traceability: Badawi discusses the Digital Product Passport developed by the European Union, a machine-readable record containing information about a product’s composition, repair instructions, spare-parts availability, and end-of-life guidance. Applied to equipment imported under Redata, this would make it possible to track how many servers entered the country, how many were taken out of service, and where they went. Technically, this is nothing exotic: it is essentially an inventory with serial numbers, something any serious data center already maintains internally. What is missing is making this information public and auditable;
2. Reuse targets, not just recycling: Recycling is better than landfilling, but it is the last step before disposal. The PNRS itself places reuse before recycling. A meaningful regulation could require part of the hardware removed from service to be tested, sanitized, and reused—whether for other purposes, less demanding workloads, or through donations to schools and public institutions;
3. Certified sanitization instead of shredding: If the reason for destroying functioning drives is data security, the answer is to regulate and recognize certified data-erasure methods. This would allow data protection and the circular economy to stop being treated as competing goals;
4. Incentives for extending equipment lifespan: Badawi proposes rewarding, within the framework of extended producer responsibility, companies that use predictive maintenance, repairs, and upgrades to extend the life of their equipment. In the context of Redata, this could become a criterion for obtaining or maintaining the tax benefit;
5. A right to repair beyond consumers: Bills currently under consideration need to address corporate use and infrastructure as well. It makes little sense to protect the right to repair an individual’s cellphone while ignoring thousands of servers discarded as a result of contractual decisions;
6. International coordination: Finally, Badawi advocates updating the Basel Convention to incorporate digital traceability tools. Brazil, which wants to become a data center hub for Latin America, could lead this discussion in the region rather than simply receiving the infrastructure—and, along with it, the waste and environmental impacts.

Between the Cloud and the Ground

We are discussing how much energy and water these facilities will consume, which is already a step forward, but we remain silent about what happens to tonnes of equipment when the business cycle decides they are no longer useful.

The cloud, after all, does not float. It is on the ground, in Caucaia, São Paulo, Fortaleza—and what it discards also stays on the ground. Attracting data centers without thinking about the waste they will generate means repeating, under the veneer of innovation, the old divide in which some reap the benefits while others are left with the waste. If Brazil wants to become a reference in “sustainable data centers,” as the official discourse claims, sustainability needs to include the end of hardware’s life, not just its beginning, because the resources required to manufacture this equipment are finite.

_____

[1] What is a Graphics Processing Unit (GPU)? Available at: https://www.ibm.com/br-pt/think/topics/gpu

[2] What is a Hyperscale Data Center? Available at: https://www.ibm.com/br-pt/think/topics/hyperscale-data-center

[3] Understanding TPUs and GPUs in AI: A Comprehensive Guide. Available at: https://www.datacamp.com/pt/blog/tpu-vs-gpu-ai

Featured image: Nadia Piet & Archival Images of AI + AIxDESIGN / betterimagesofai.org / creativecommons.org/licenses/by/4.0/

Anicely Santos

Master’s candidate in the Computer Engineering Program at the University of Pernambuco (UPE). Holds a degree in Systems Analysis and Development, with a specialization in Data Science and Analytics. President of the Python Brasil Association (2024–2025).