What Kind of Resources Does AI Consume Besides Electricity?
During a recent visit to the Computer History Museum in Mountain View, we joined a guided tour led by a Silicon Valley product manager. She walked us through decades of computing evolution, from punch cards to cutting-edge AI hardware and models.
When she reached the topic of AI, someone in the group asked, “How sustainable is all of this?”
It was a simple question that opened up a deeper conversation.
While electricity often dominates the sustainability debate, AI systems consume far more than just power. Beneath every prompt and behind every GPU lies a hidden supply chain of natural resources, rare materials, and even drinking water.
The Physical Cost: Hardware and Raw Materials
AI depends heavily on the physical infrastructure that supports it, including servers, chips, data centers, and cooling systems. All of this requires a range of natural resources, beginning with rare earth minerals like lithium, neodymium, and praseodymium. These materials are extracted through environmentally intensive mining and refined through energy-intensive manufacturing processes.
What complicates this further is the rapid pace of hardware development. As companies push to build faster, more powerful AI models, the demand for high-performance chips and servers grows exponentially. For example, Meta has reportedly acquired more than 350,000 of NVIDIA’s latest-generation GPUs to power its AI ambitions. This needs a huge amount of raw materials, electricity, and manufacturing output.
Meanwhile, the technology itself becomes obsolete quickly. Hardware that was state-of-the-art just a few years ago may already be outdated. With limited systems in place for reuse or recycling, much of that material becomes e-waste, given the high production costs, but they have a shorter lifespan in use and are difficult to recycle.
The Hidden Cost: Water Use for Cooling
The environmental footprint of AI doesn’t stop at mining or electricity. Data centers also rely heavily on water, often treated drinking water, for cooling. This helps prevent problems like mineral buildup or rust that could damage delicate equipment. Besides that, in some regions, it’s the only practical option available.
If you wonder how much water, just during the training of GPT-4, Microsoft used roughly 11.5 million gallons of water to cool its data centers in Iowa (O’Brien, 2025). And even a single user session with ChatGPT, let’s say 10 to 50 prompts, can indirectly consume up to half a liter of water. That adds up quickly, especially as AI usage becomes routine for millions of people every day.
Why This Matters
We often think of AI as just digital, but it’s built on the physical world, from global supply chains, mined materials, industrial cooling systems, and infrastructure systems. Each of these carries its own environmental consequences. If we only track carbon emissions, we miss the broader ecological footprint.
To create sustainable AI, we need to consider its full lifecycle: from material sourcing and manufacturing to deployment and disposal.
Key Takeaways
- AI hardware depends on rare minerals like lithium and neodymium, extracted through environmentally intensive mining processes.
- High-performance GPUs and servers require large amounts of raw materials, yet quickly become obsolete due to rapid hardware innovation.
- Many data centers consume vast amounts of drinking water to keep AI systems cool and operational.
- Sustainability in AI must account for its entire lifecycle, including material sourcing, manufacturing, deployment, and disposal.
Understanding AI’s full resource footprint is the first step toward building technology that’s not just intelligent, but also responsible.
This article is just a glimpse of what’s in our full AI Sustainability White Paper.
