← Latest papers
🤖 machine learning

Position: Neglecting the Sustainability of AI is Fuelling a Global AI Arms Race

This position paper argues that neglecting the economic and social dimensions of AI sustainability, particularly the tension between equitable resource access and environmental impact, fuels a global AI arms race, and proposes a Marxist-informed analysis alongside the Climate and Resource Aware Machine Learning (CARAML) framework to reconcile these factors for truly sustainable AI.

Original authors: Pedram Bakhtiarifard, Pınar Tözün, Christian Igel, Raghavendra Selvan

Published 2026-06-02
📖 4 min read☕ Coffee break read

Original authors: Pedram Bakhtiarifard, Pınar Tözün, Christian Igel, Raghavendra Selvan

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

Imagine the world of Artificial Intelligence (AI) right now is like a massive, high-stakes global race. Countries and giant tech companies are sprinting to build the biggest, fastest, and most powerful AI engines they can. They are pouring billions of dollars into super-computers and hiring thousands of engineers, much like nations in the Cold War rushing to build more nuclear weapons.

This paper argues that this "AI Arms Race" is dangerous because everyone is so focused on winning the race that they are ignoring the cost of the fuel and the exclusion of the runners.

Here is a simple breakdown of the paper's main points:

1. The Three-Legged Stool is Wobbly

The authors say "Sustainability" isn't just about saving the planet (the environment). It's a three-legged stool:

  • The Environment: How much electricity and water does AI use? How much carbon does it spew?
  • The Economy: Who has the money to build these things?
  • Society: Who gets to participate? Is it fair?

Currently, the AI world is obsessed with the Environment leg (trying to make AI "green" or efficient) but is ignoring the Economy and Society legs. They are trying to make the race faster without asking who is allowed to run or if the track is destroying the forest.

2. The Trap of "Efficiency" (The Jevons Paradox)

The paper points out a tricky trap. When engineers make AI models more efficient (using less energy per task), it sounds great. But, it often leads to the opposite result.

  • The Analogy: Imagine you invent a car that uses half as much gas. Great, right? But because it's cheaper to drive, everyone starts driving twice as much, and they buy bigger trucks. Suddenly, you are using more gas than before.
  • In AI: When models like "DeepSeek" became efficient, more people used them, and companies started building even bigger and more complex models (like "reasoning" models) that use massive amounts of energy. The efficiency gains were swallowed up by the sheer volume of usage.

3. The "Rich vs. Poor" Divide

The paper uses a historical lens (inspired by Karl Marx) to explain that the hardware (the "Base") controls the rules and culture (the "Superstructure").

  • The Reality: Only a few countries (mostly in the US, China, and Europe) and a handful of giant companies own the massive data centers and super-computers needed to train AI.
  • The Result: This creates a "Digital Colonialism." Rich nations and companies set the rules, build the models, and capture the profits. Poorer nations and smaller researchers are locked out because they can't afford the "entry fee" (the electricity and hardware).
  • The Irony: Even when people talk about "AI Sovereignty" (countries wanting their own AI), it often just fuels the arms race, making countries build more expensive infrastructure to compete, rather than sharing resources.

4. The Solution: The CARAML Framework

The authors propose a new way to think about AI called CARAML (Climate And Resource Aware Machine Learning). They want to stop the arms race and build a system that is truly sustainable.

They suggest a "Call to Action" for different groups, like a team working together to fix a leaking boat:

  • Individuals (The Researchers): Stop just measuring success by how "smart" the AI is. Start measuring how much energy it took to get there. If an AI is smart but costs a fortune in electricity, it's a failure.
  • The Community (Conferences): Stop wasting time and energy on experiments that are just repeats of old ideas. Require researchers to "pre-register" their plans so we don't waste resources on dead ends.
  • Industry (The Big Tech Companies): They need a "Carbon Cap." Just like a factory can't dump unlimited pollution, AI companies should have a hard limit on how much carbon they can emit.
  • Governments: They need to treat AI like a major construction project. Before building, they should do an "Impact Assessment" to see how it affects jobs, privacy, and the climate.
  • Global Level: We need to fix the "Digital Divide." Rich countries should help poorer countries get access to computing power. Instead of every country building its own expensive engine, we should share the load to solve global problems like climate change together.

The Bottom Line

The paper concludes that we cannot solve climate change with AI if the AI itself is destroying the climate. We cannot have a fair society if only the rich can build the technology.

To win the race for a better future, we have to stop racing against each other to build bigger, dirtier machines. Instead, we need to slow down, share the resources, and make sure that when we build AI, we are building it for everyone and for the planet, not just for profit and power.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →