Ace Business Desk – Training a large AI model is energy-intensive by design. Data centers consume power by the gigawatt, and AI companies now publish carbon reports. But while energy gets measured and stays in the conversation, the metals in AI chips have gone largely unexamined

A new study took a chip apart, analyzed it element by element, and arrived at a number the industry has mostly preferred not to calculate.

Metals in AI chips

To find out what AI hardware actually contains, a team at the University of Bonn pulled apart an Nvidia A100 – the chip that powered the early boom in AI chatbots – and analyzed it in a chemistry lab.

Sophia Falk – a researcher at Bonn’s Sustainable AI Lab and the study’s lead author – worked with colleagues to catalog every element in the device. They found 32.

About 90% of the chip’s mass consists of heavy metals. Copper alone weighs roughly 3 pounds (1.4 kilograms) per unit, with iron, tin, silicon, and nickel rounding out the top five. Gold, silver, platinum, and palladium show up only in trace amounts.

A toxic mix

Of the 32 elements the team cataloged, a striking number are classified as hazardous: arsenic, mercury, lead, cadmium, chromium, zinc, nickel, antimony, cobalt, and beryllium.

By mass, about 93% of a single A100 is made of elements with documented toxic properties. Sealed inside the device, those materials pose no danger to a technician sliding one into a server rack.

The danger is not in the chip sitting in a server. It lies in the ground where those metals were dug up, and in the e-waste pile where old hardware ends up. A separate paper from the same group documents the entire cradle-to-grave cycle.

Chips, metals, and GPT-4

How many chips does a single training run actually require? That depends on two things: how hard the chips are pushed, and how long they last before failing.

Under what the team calls the most plausible baseline – 35% utilization, a two-year lifespan – training one round of GPT-4 consumes the equivalent of about 2,515 A100 chips.

Stretch the lifespan to three years and that drops to roughly 1,676. Push the other way – low utilization, short lifespans – and a single training run can burn through up to 8,800 GPUs.

Either way, that amounts to thousands of devices for one model. The team estimates roughly 4 tons (3.6 metric tons) of extracted material for a single GPT-4 training run.

Diminishing returns in AI

The most striking number in the paper is not about one model. It is the jump between two. OpenAI’s move from GPT-3.5 to GPT-4 required roughly 31 times more GPU resources – a more than 3,000% increase in computing power.

Performance returns were uneven. GPT-4 jumped 61% over its predecessor on a hard math benchmark and 39% on coding. But on commonsense reasoning, it improved by just 14 percent.

“Architectural innovations and training methodologies may offer more effective performance improvements than simply scaling raw resources,” wrote Falk and her co-authors. Bigger, the paper argues, is not the same as smarter.

Where the costs land

At the chip level, the numbers tell one story, while geography tells another. The metals inside an A100 come from mines and refineries far from the data centers that put the chips to work.

Across the nine models the team analyzed, the most plausible scenario adds up to roughly 7 tons (6.4 metric tons) of extracted material, almost all of it classified as hazardous. The worst-case scenario lands close to 22 tons (20 metric tons).

Most of that environmental burden does not fall near the data centers using the chips. It tends to fall near the mines, in regions with thinner environmental oversight than the cities buying the compute power.

The longer-life lever

Two things, the team writes, could meaningfully change those numbers. Run the chips harder while they are in service. Keep them in service longer. The effects compound.

Lift utilization from 20 to 60% and the GPU count for a given training job drops by about two-thirds. Stretch hardware life from one to three years and the result is a similar cut.

Combine both approaches. Run an A100 at 60% utilization for five years rather than one year at 20%, and the chips needed to train GPT-4 fall from 8,800 to 587. That is a 93% reduction.

A separate analysis projects AI workloads will make up nearly 70% of total data center demand by 2030, which sets the stakes if those efficiency gains never arrive.

What this changes

The study’s central contribution is the bridge. That GPUs contain heavy metals was known, and so was AI training’s appetite for those chips. What no one had done was tie those facts together and put a number on the material cost of training a specific model.

A GPT-4 baseline now exists: a few thousand chips and several tons of mined material per training run, most of it toxic. Policymakers, AI developers, and chipmakers have something concrete to work from.

Energy and water are no longer the whole footprint of training a large AI model. The metals – many of them toxic – belong in the same accounting.

Falk’s group is calling on AI labs to disclose training configurations as part of standard sustainability reporting, so the next model’s footprint is not something outsiders have to piece together from leaked spec sheets.

At Sterling Publishing & Media Service Agency, we prioritise transparency and accountability in all our operations. We wish to clarify that we are not responsible for any external content, hyperlinks, or costs associated with our services. Nevertheless, we remain committed to delivering outstanding services and greatly value your continued support. Thank you for your trust in us.