Generative AI is often discussed in terms of computing power, energy demand and new business models. A newer concern is more physical: what happens to the servers, GPUs, CPUs, memory modules and storage devices when the hardware behind AI systems is replaced?
A study published in Nature Computational Science estimates that generative AI could add 1.2 million to 5 million metric tons of e-waste in total by 2030, depending on how quickly the technology is adopted. That would be a relatively small share of the global total of over 60 million metric tons of e-waste each year, but experts warn it would still deepen an already serious waste problem.
Why AI hardware becomes e-waste
E-waste refers to discarded products such as air conditioners, televisions, cell phones, laptops, washing machines and high-performance computers. These devices can contain materials that are hazardous or toxic when handled poorly, including lead, mercury and chromium. They can also contain valuable metals such as copper, gold, silver, aluminum and rare earth elements.
The concern is not only that discarded electronics can harm human health or the environment. It is also that useful materials are removed from the supply chain when equipment is thrown away instead of recovered. In that sense, e-waste is both a pollution issue and a resource issue.
For generative AI, the largest source of expected waste is the high-performance computing equipment used in data centers and server farms. That includes servers, GPUs, CPUs, memory modules and storage devices. These systems are central to training and running generative AI models, and they are also the equipment most likely to be replaced as performance demands rise.
One reason the waste can build quickly is the pace of hardware improvement. Computing devices typically have lifespans of two to five years, and companies often replace them with newer versions. In a fast-moving AI market, that upgrade cycle can turn working equipment into discarded equipment sooner than many people outside the industry might expect.
The scale is smaller than global e-waste, but still significant
The study’s estimate ranges widely because the future adoption rate of generative AI is uncertain. In a lower-growth path, the technology could add 1.2 million metric tons of e-waste in total by 2030. In a higher-growth path, that figure could reach 5 million metric tons.
Compared with the current global total of over 60 million metric tons of e-waste each year, AI’s projected contribution is not the whole problem. But that comparison can be misleading if it makes the issue seem minor. Adding millions of metric tons to a system that already struggles to collect, process and safely recycle electronics is still consequential.
Asaf Tzachor, a researcher at Reichman University in Israel and a co-author of the study, said the increase would worsen the existing e-waste challenge. Kees Baldé, a senior scientific specialist at the United Nations Institute for Training and Research and an author of the latest Global E-Waste Monitor, described the study as novel because it attempts to quantify AI’s e-waste impact.
The finding also matters because generative AI is still developing. That makes this a moment for companies, manufacturers and policymakers to decide whether AI infrastructure will follow the same wasteful patterns as other electronics or be designed around longer use and better recovery from the start.
Reuse and better design could change the outcome
The study points to several ways to reduce the expected waste. The most direct is extending the lifespan of equipment by using it for longer. If hardware remains useful for more time, fewer devices have to be manufactured, shipped, installed and eventually discarded.
Refurbishing and reusing components can also reduce waste. Instead of destroying or discarding hardware after its first use, companies can keep parts in circulation where performance, safety and operational needs allow. Designing hardware so it is easier to upgrade and recycle would also help.
In the best-case scenario projected by the study, these strategies could reduce e-waste generation by up to 86%. The practical steps include:
- Using equipment longer before replacing it with newer versions.
- Refurbishing components so hardware can remain in service.
- Reusing parts where they still meet operational needs.
- Designing hardware for recycling and upgrades so valuable materials are easier to recover.
These measures do not eliminate the broader e-waste problem, but they can lower AI’s contribution to it. They also create a clearer path for recovering metals that would otherwise be lost when equipment is discarded.
Recycling is still not keeping up
Only about 22% of e-waste is being formally collected and recycled today, according to the 2024 Global E-Waste Monitor. More is collected and recovered through informal systems, including in low- and lower-middle-income countries that lack established e-waste management infrastructure.
Informal recovery can extract valuable metals, but Baldé warned that it often does not include safe disposal of hazardous materials. That creates a difficult split: the economic value of metals can make recovery attractive, while the toxic parts of discarded equipment still require careful and costly handling.
This is why policy is likely to matter. More rules may be needed to make sure e-waste, including waste linked to AI, is recycled or disposed of properly. Recovering materials such as iron, gold and silver can strengthen the economic case, but recycling still comes with costs because hazardous materials must be managed safely.
Data security is a practical barrier
For AI companies and other businesses, recycling hardware is not just an environmental decision. It is also a data security decision. Destroying equipment can help prevent information from leaking, while reuse or recycling requires other methods to protect sensitive data.
Tzachor said ensuring that sensitive information is erased from hardware before recycling is critical, especially for companies handling confidential data. That means e-waste planning has to be connected to security procedures, not treated as a separate facilities issue at the end of a device’s life.
The study’s broader message is that AI’s physical footprint should be managed early. Generative AI depends on powerful hardware, and that hardware contains both valuable and hazardous materials. If companies extend equipment life, reuse components, design for recycling and handle data securely, AI-related e-waste can be reduced before it becomes much harder to manage.