• Pascal's Chatbot Q&As
  • Posts
  • GPT-4o: The EU AI Act only requires energy disclosure for training, not inference, even though inference is more power-intensive. Data centers will demand gigawatts, threatening to divert power...

GPT-4o: The EU AI Act only requires energy disclosure for training, not inference, even though inference is more power-intensive. Data centers will demand gigawatts, threatening to divert power...

...from other sectors. If power-intensive inference is left unregulated, there’s a risk of massive environmental degradation going unnoticed.

Author

Pascal Hetzscholdt
May 26, 2025


Essay: Artificial Intelligence – Supply Chain Constraints and Energy Implications

By ChatGPT-4o, based on “Artificial Intelligence: Supply Chain Constraints and Energy Implications” by Alex de Vries-Gao, Joule (2025)

Introduction

The explosive growth of generative AI since the introduction of ChatGPT has not only transformed digital services but has also imposed unprecedented strain on global infrastructure. In his 2025 article, Alex de Vries-Gao investigates the hidden environmental and supply chain consequences of this boom. With big tech firms like Google and Microsoft reporting rising emissions and energy usage while offering minimal transparency, the study aims to estimate AI’s energy footprint by analyzing production capacity within the supply chain—specifically, the CoWoS (chip-on-wafer-on-substrate) packaging capacity provided by TSMC.

Core Concerns and Their Universality

1. Opaque Energy Reporting by Big Tech

  • Concern: Companies have ceased disclosing AI-specific energy metrics.

  • Universality: High. Affects all countries depending on cloud infrastructure.

  • Solvability: Difficult without regulation; voluntary measures have failed.

2. AI Hardware Bottlenecks: CoWoS Capacity

  • Concern: A single packaging technology (CoWoS) provided mainly by TSMC has become the chokepoint for AI hardware production.

  • Universality: High. The entire AI industry relies on this capacity.

  • Solvability: Moderate. Capacity can be scaled, but geopolitical risks (e.g., tensions around Taiwan) limit resilience.

3. Power Demand Surging Beyond National Scales

  • Concern: Devices produced in just two years (2023–2024) could consume more power than entire nations like Ireland or Switzerland.

  • Universality: Critical. Energy markets, carbon emissions, and grid capacity are affected globally.

  • Solvability: Limited. Efficiency gains exist, but rebound effects (more use) often negate them.

4. Inference Costs Ignored in Regulation

  • Concern: The EU AI Act only requires energy disclosure for training, not inference, even though inference is more power-intensive.

  • Universality: Legal/regulatory. Applies mainly to EU but has global implications via corporate compliance.

  • Solvability: Solvable via regulatory amendment, but politically challenging.

5. Geopolitical Disruptions

  • Concern: Export controls and trade tensions (e.g., U.S.-China AI chip bans) threaten stability.

  • Universality: High. Affects global supply chains, with countries like China developing their own less energy-efficient systems.

  • Solvability: Uncertain. Diplomacy and industrial diversification may help, but nationalism is increasing.

Issues That Are Hard to Solve

  • Lack of transparency and accountability from private actors.

  • Systemic rebound effects that cancel out efficiency improvements.

  • Concentration risk around single providers like TSMC.

  • Dependence on fossil fuels in new data center developments (e.g., Crusoe’s 4.5 GW natural gas supply).

  • Geopolitical instability affecting supply chains.

Issues That Can Be Addressed

  • Regulation mandating AI-specific energy use disclosure, especially for inference.

  • Encouraging open data-sharing frameworks for supply chain tracking.

  • Diversifying hardware production beyond TSMC via subsidies or trade agreements.

  • Developing and promoting energy-efficient AI models (e.g., DeepSeek’s approach in China).

  • Enforcing sustainability standards like PUE (power usage effectiveness) targets, as seen in Germany’s Energy Efficiency Act.

Societal Consequences

  1. Energy Infrastructure Stress

    • Data centers will demand gigawatts, threatening to divert power from other sectors.

  2. Environmental Setbacks

    • AI may undermine global climate goals if fossil-based infrastructure grows unchecked.

  3. Inequality in AI Access

    • Only wealthy firms or nations can afford high-end, power-hungry models, excluding others.

  4. Public Misinformation and Lack of Oversight

    • If power-intensive inference is left unregulated, there’s a risk of massive environmental degradation going unnoticed.

  5. Resource Nationalism

    • Nations might rush to secure AI hardware or power sources, increasing geopolitical tension.

  6. Policy Paralysis

    • Without hard data, regulators may fail to create effective policy, delaying crucial mitigation.

Conclusion

De Vries-Gao’s analysis reveals the staggering, yet under-acknowledged, energy implications of modern AI development. Despite the promise of AI, its supply chain and energy demands already rival or exceed those of entire nations. Unless immediate transparency measures and sustainability safeguards are implemented, AI risks becoming one of the most environmentally and geopolitically destabilizing technologies of our time. Addressing these challenges requires a mix of regulation, infrastructure reform, global cooperation, and a cultural shift toward energy-conscious digital innovation.

·

20 APR