💻 computer science

The Rising Unsustainability of AI Graphics Cards Production

This study estimates the escalating environmental damages of AI graphics card production from 2013 to 2025, revealing that despite operational efficiency gains, the growing resource depletion and carbon emissions from manufacturing necessitate a shift toward structural policy changes, durable hardware design, and a cultural move away from perpetual growth.

Original authors: Clément Morand, Aurélie Névéol, Anne-Laure Ligozat

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Clément Morand, Aurélie Névéol, Anne-Laure Ligozat

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The Hidden Cost of the "Brain"

Imagine the Artificial Intelligence (AI) boom as a massive construction project building a new, super-smart city. Everyone is talking about how much electricity this city uses to run its lights and computers (the "operational" cost).

However, this paper argues that we are ignoring the cost of building the bricks themselves. The authors focus on graphics cards (the specialized computer chips that power AI). They found that while we are trying to make these chips run more efficiently, the process of manufacturing them is actually getting dirtier, heavier, and more resource-intensive every year.

The Investigation: Tracking the "Bricks"

The researchers acted like detectives, looking at the history of NVIDIA's workstation graphics cards (the high-end "bricks" used for AI) from 2013 to 2025.

They built a dataset of 174 different card models. To understand the environmental damage, they used a method called Life Cycle Assessment (LCA). Think of this as a "cradle-to-gate" audit:

Cradle: Digging up the raw metals and minerals.
Gate: The moment the finished card rolls off the assembly line.
Note: They did not count the electricity used when the card is actually running in a data center; they only counted the cost of making it.

They measured two main things:

Carbon Footprint: How much greenhouse gas is emitted to make the card (like the exhaust from a factory).
Resource Depletion: How much of the Earth's finite metals (like gold, copper, and rare earths) are used up to make the card.

The Findings: Bigger, Faster, and More Expensive

The study found a clear, worrying trend: Making a single graphics card is becoming much more damaging over time.

1. The "Super-Size" Effect

The Analogy: Imagine a smartphone. Ten years ago, it had a small screen and a tiny camera. Today, it has a massive screen, a huge camera, and a super-fast processor.
The Reality: AI cards have followed the same path. The "brain" (GPU) inside the card has gotten physically larger, and the memory (storage) has exploded in size.
The Catch: Even though the technology inside has become more precise (using smaller, finer circuits), the sheer size of the chip and the amount of memory needed have grown so fast that the environmental cost of making one card has quadrupled since 2013.
- 2013: Making a card cost about 50 kg of CO2 emissions.
- 2025: Making a card costs about 200 kg of CO2 emissions.

2. The "Rebound" Trap

The Analogy: If you buy a super-fuel-efficient car, you might feel good about it. But if that efficiency makes you drive 10 times further every day, you end up using more gas than before.
The Reality: Engineers have made chips more energy-efficient. But because they are so efficient, companies buy bigger cards and run them harder. The paper notes that while the chips are better, the total energy needed to run a single card has actually gone up slightly because we are asking them to do so much more work.

3. The Sales Explosion

The Analogy: It's not just that each brick is heavier; it's that we are building a skyscraper instead of a house.
The Reality: The number of these cards being sold has skyrocketed. In 2022, companies bought a few hundred thousand cards. By 2024, they were buying over a million per quarter.
The Result: The total environmental damage from just making these cards has surged. The authors calculated that the pollution from making the cards sold in 2024 is roughly equivalent to the annual emissions of 700,000 people living in a city like Copenhagen, Denmark.

The "Base" Problem

The researchers broke down the damage into three parts:

The GPU (The Brain): Getting worse over time.
The Memory (The Storage): Getting much worse over time (this is the fastest-growing source of damage).
The Base (The Case, Cooling, and Circuit Board): This part stays roughly the same, but because the GPU and Memory are getting so huge, the "Base" now looks like a tiny fraction of the total problem.

The Conclusion: Efficiency Isn't Enough

The paper concludes with a hard truth: We cannot just "green" our way out of this problem.

The Misconception: Many people think if we just power our data centers with solar energy, we solve the AI environmental crisis.
The Reality: Switching to clean energy helps with the running costs, but it does nothing for the making costs. The damage happens before the card is even plugged in.

The Final Metaphor:
Imagine you are trying to save the planet by driving an electric car. That's great! But if you are also buying a new car every month and throwing the old one in a landfill, the environmental damage of manufacturing all those new cars will eventually outweigh the benefits of driving them.

The authors argue that the AI industry needs to stop assuming that "more powerful" always equals "better." They suggest we need to look at sufficiency—asking if we really need to build these massive, resource-hungry cards every single year, or if we can find ways to use what we have for longer.

In short: The paper warns that the "bricks" of the AI revolution are becoming increasingly toxic to produce, and simply making them run on clean electricity won't fix the damage caused by building them in the first place.

Technical Summary: The Rising Unsustainability of AI Graphics Cards Production

Problem Statement
The rapid expansion of Artificial Intelligence (AI) infrastructure has driven massive investments in hardware, particularly graphics cards, which are central to AI training. While significant attention has been paid to the operational energy consumption and carbon emissions of data centers, the environmental damages associated with the production of these hardware components remain understudied. Existing literature often lacks transparency in lifecycle data, and while parametric models exist for servers and user devices, high-quality data on graphics cards—specifically regarding resource depletion and embodied carbon—has been scarce. This gap prevents a full understanding of the unsustainability of the AI sector, especially as large-scale data center projects require hundreds of thousands of graphics cards per installation. Furthermore, current mitigation strategies often focus on greening electricity supplies, which fails to address the escalating impacts of hardware manufacturing.

Methodology
To address this gap, the authors conducted a Cradle-to-Gate Attributional Life Cycle Assessment (LCA) of NVIDIA workstation graphics cards released between 2013 and February 2025.

Data Curation: The authors compiled a dataset of 174 workstation graphics card models. Data sources included the TechPowerUp GPU database, Wikipedia, and official NVIDIA datasheets (used as the reference for resolving discrepancies). Key technical parameters collected included release date, GPU die area, technological node, memory type, memory size, and Thermal Design Power (TDP).
Parametric Modeling: The study adapted the parametric model implemented in MLCA (based on Boavizta data) to estimate production impacts. The model treats the graphics card as a sum of its parts:
- GPU Die: Impact calculated based on die surface area and a fixed impact per $cm^2$ (derived from Intel CPU production data).
- Memory: Impact calculated by estimating the memory die area required for the total memory volume (using a constant memory density factor) and multiplying by the impact per $cm^2$ of memory die.
- Base Components: A constant impact value accounts for the casing, heat sink, Printed Circuit Board (PCB), and upstream transportation.
Impact Categories: The assessment focused on two primary indicators:
1. Carbon Footprint: Measured as Global Warming Potential (GWP) in $kgCO_2$ eq.
2. Metallic Resource Depletion: Measured as Abiotic Depletion Potential for elements (ADPe) in $kgSb$ eq (antimony equivalent).
Global Scale Analysis: The production impact per card was combined with sales data from Epoch AI (2022–2025) to estimate the total global environmental damages incurred by AI graphics card procurement.

Key Contributions

Dataset Release: The authors provide a curated dataset documenting the environmental damages of producing NVIDIA workstation graphics cards released between 2013 and 2025, along with the code used for reproducibility.
Trend Analysis: The study provides evidence that the environmental impact of individual graphics cards has increased consistently over the past decade.
Procurement Trends: The paper analyzes the trajectory of graphics card procurement, linking sales volume growth with escalating cumulative production damages.

Results

Hardware Evolution: Over the 12-year period, the average GPU die area increased linearly, while the technological node size decreased (finer nodes). Memory size (in GB) grew exponentially (approx. 30% CAGR). Despite improvements in compute efficiency, the total energy consumption (TDP) of cards has slightly increased, consistent with a rebound effect.
Production Impacts:
- Carbon Footprint (GWP): The production carbon footprint of a single workstation card has risen from approximately 50 $kgCO_2$ eq in 2013 to around 200 $kgCO_2$ eq for recent models.
- Resource Depletion (ADPe): Metallic resource depletion per unit has remained relatively constant at approximately 0.01 $kgSb$ eq, though the composition of impacts has shifted.
- Component Shifts: While base components (PCB, casing) historically dominated resource depletion, the share of impacts from GPU and memory has grown. In recent years, memory has become the dominant source of GWP impacts, while base components remain dominant for ADPe.
Global Scale: Between Q1 2022 and Q3 2024, the cumulative production damages of sold NVIDIA AI cards showed a sharp increase. The total estimated GWP for this period (approx. 698 $ktCO_2$ eq) is equivalent to the maximum annual emissions of roughly 700,000 people under planetary boundary constraints.
Validation: The study's parametric model yields results consistent with previous independent studies (e.g., Lees-Perasso et al., ADEME) and NVIDIA's own disclosed data for specific models like the H100, with differences generally within acceptable margins (e.g., <33% for GWP).

Significance and Claims
The paper argues that the AI community cannot rely solely on operational efficiency improvements (such as energy-efficient training or carbon-aware computing) to achieve sustainability. The study highlights that production-related impacts are escalating and constitute a significant, often overlooked, portion of the AI sector's environmental footprint.

The authors claim that the current trajectory of AI hardware development—characterized by frequent updates, increasing die sizes, and exponential sales growth—is fundamentally at odds with global sustainability limits and planetary boundaries. They assert that addressing the unsustainability of AI requires moving beyond incremental optimizations toward "sufficiency." This necessitates structural changes, including policy interventions, hardware design focused on longevity, and a cultural shift away from the imperative of perpetual growth and performance increases. The study underscores the urgent need for greater transparency in lifecycle data to fully grasp the material implications of AI infrastructure.