Graphics processing units, or GPUs, have become the backbone of the artificial intelligence boom. From training large language models to running real-time inference in data centers, these chips are being deployed by the hundreds of thousands. Yet beneath the surface of technological progress lies a massive environmental footprint that spans the entire lifecycle of a GPU — from raw material extraction to manufacturing, operation, and eventual disposal.

Nvidia, the world's most valuable chip company, released what it calls the first GPU in 1999, but the origins of graphics hardware date back to the 1970s arcade machines. What began as a tool for rendering video game graphics has evolved into the engine for generative AI. This shift has brought unprecedented demand for GPUs, and with it, a host of ethical and environmental questions that were once confined to gaming are now magnified on a global scale.

The Energy Hunger of AI Data Centers

Data centers housing thousands of GPUs consume enormous amounts of electricity. In the United States alone, the power used by GPU-accelerated AI servers grew from 2 terawatt-hours in 2017 to over 40 terawatt-hours in 2023, according to a Lawrence Berkeley National Laboratory study. Projections suggest that by 2028, AI servers could use between 165 and 326 terawatt-hours annually — the lower end equivalent to the electricity consumption of more than 8.7 million American homes.

This surge in energy demand comes at a time when U.S. electricity grids still rely heavily on fossil fuels. The resulting carbon emissions from AI data centers in 2025 were estimated to be between 32.6 million and 79.7 million tons of CO2, comparable to the annual emissions of New York City. Air pollution from these facilities also has a direct public health cost: a 2024 study from the University of California, Riverside and Caltech estimated that by 2028, AI-related air pollution could cause 1,300 premature deaths per year, with cumulative health costs exceeding $20 billion.

Water consumption is another critical concern. Data centers need water for cooling and for the power plants that generate their electricity. A 2023 study projected that AI could use up to 600 billion liters of water annually by 2027. More recent estimates from 2025 suggest the range could be between 312.5 and 764.6 billion liters — roughly equivalent to the volume of water in all the plastic bottles consumed globally each year. These figures can spike dramatically during hot weather, placing sudden stress on local water districts that are often already facing drought conditions.

In Southern California, where many data centers are being built, the pattern echoes earlier conflicts over warehouse distribution centers. Communities grapple with increased traffic, noise, and now the whir of cooling fans and backup generators. The NAACP has sued xAI (now SpaceXAI) over air pollution from gas generators at its data centers, highlighting the environmental justice dimension of this boom.

From Cradle to Grave: The Full Lifecycle Impact

The environmental toll of GPUs does not start when they are plugged in. Manufacturing a single GPU requires a complex supply chain spanning multiple continents. A typical Nvidia A100 GPU contains about 90% heavy metals, including copper, iron, tin, and nickel, plus silicon. The copper alone amounts to roughly 1.4 kilograms per chip. With models like GPT-4 requiring between 1,174 and 8,800 such GPUs, the total material demand becomes staggering.

Copper mining is associated with acid mine drainage, which can contaminate local water sources for decades. The demand for copper is projected to grow by 24% over the next decade, driven partly by the electrification of transportation and partly by AI infrastructure. New mines would need to open twice as fast as they did a decade ago to keep pace.

Beyond raw materials, semiconductor fabrication uses chemicals that have left a toxic legacy. Santa Clara County in Silicon Valley, the birthplace of the chip industry, has more Superfund sites than any other county in the U.S. due to contamination from manufacturing. Some chemicals linked to miscarriages were phased out in the 1990s, but production has since moved to Asia, where health risks have resurfaced. The current push for domestic chip manufacturing — including TSMC's new plant in Arizona — has raised concerns about the use of per- and polyfluoroalkyl substances (PFAS), known as forever chemicals, which are linked to kidney and testicular cancer. Residents near the Phoenix facility have expressed deep anxiety about potential water contamination.

At the end of their useful life, GPUs become e-waste. In data centers, the average lifespan of a GPU is about three to five years. By 2030, AI servers could generate between 0.131 million and 0.225 million tons of e-waste annually, comparable to the total e-waste of a country like Denmark or Austria. And while around 30% of e-waste in the Americas is formally recycled, the rest often enters an informal sector in lower-income countries, where workers — sometimes children — manually dismantle electronics without protective gear, burning or burying toxic components. The United States has not ratified the Basel Convention, which restricts international trade in hazardous waste, allowing U.S. recyclers to send e-waste abroad.

Gaming vs. AI: A Changing Landscape

While AI dominates headlines, gaming remains a major user of GPUs. A 2019 study found that gaming in the U.S. consumed about 34 terawatt-hours per year, with carbon emissions equivalent to 5 million cars. However, AI has already surpassed gaming in energy consumption and is growing exponentially. The question is whether the benefits of AI justify the environmental costs — and whether the industry can rein in the "bigger is better" mentality that drives ever larger models and ever more powerful chips.

Some researchers argue for a sufficiency approach: rather than endlessly scaling up, we should ask when a model is good enough. Improving efficiency can also have a rebound effect, where cheaper computing encourages even more usage. Nvidia touts that its Blackwell Ultra GPU is 50 times more efficient than the previous Hopper architecture, but the overall energy consumption continues to rise.

Consumers and regulators have a role to play. Gamers have shown a willingness to take collective action on climate change, according to a 2024 study. And as AI becomes more pervasive, public pressure on tech companies to disclose environmental impacts and invest in renewable energy is growing. Greenpeace East Asia is pushing Nvidia to clean up its supply chain, noting that the company's market cap of $4 trillion shields it from the scrutiny faced by more consumer-facing firms.

What Can Be Done?

Several strategies could reduce the environmental footprint of GPUs. Extending the lifespan of chips before replacement can cut e-waste by up to 86%. Designing models that are more energy-efficient and using recycled materials in manufacturing can also help. Data center operators can locate facilities near renewable energy sources and implement water recycling systems. But these measures require transparency and accountability from the tech industry — and a willingness to prioritize long-term sustainability over short-term growth.

For now, the GPU remains a symbol of both technological promise and environmental peril. As communities from Arizona to Taiwan confront the costs of this boom, the central question persists: is the value we get from AI worth the price we pay?

Source: The Verge News