AI Governance is an Illusion Without Climate Governance

At What Point Does AI Infrastructure Expansion Become a Climate Crisis?

By 2030, AI data centers will consume 945 terawatt-hours annually, generating 220 million tons of CO₂. This is equivalent to Japan’s annual electricity requirement, and to the entire aviation industry’s carbon footprint. So why has AI infrastructure almost completely evaded climate governance. An oversight? It’s a systemic blind spot where efficiency metrics mask exponential growth, and innovation rhetoric crowds out accountability.

The following framework called the Compute-to-Climate Deterrence Ratio (CCDR) was designed to close this gap and answer a question that current policy frameworks aren’t touching: At what point does your AI infrastructure expansion become a climate crisis? The only right answer is NOW.

CCDR proactive risk strategy that supports sustainable innovation by addressing a critical governance vacuum: the absence of policy mechanisms translating AI infrastructure expansion into enforceable climate accountability. CCDR inquires how much non-green compute is permissible under carbon budgets.

The Efficiency Paradox: Context Matters

Today, we are measuring the climate change by measuring post-compute metrics.

• Power Usage Effectiveness (PUE): How efficiently does your data center use power?
• Carbon Usage Effectiveness (CUE): How much CO₂ per kilowatt-hour?
• Water Usage Effectiveness (WUE): Liters of water consumed per kilowatt-hour?

These are good metrics and companies have gotten better at them. Google achieves a PUE of 1.1, Meta’s CUE is down to 0.18. And their absolute emissions have tripled. Current policy frameworks mandate reporting without accountability and consequence:

• EU AI Act (2024), Article 52: Requires ‘systematic environmental impact assessment’ for high-risk systems (>10²⁵ FLOPs). However, this creates only transparency obligations. Once there is data available, there are no caps, taxation, or intervention triggers exist.
• US Executive Order 14110 (2023): Encourages DOE voluntary audits without establishing FLOP thresholds or enforcement mechanisms.
• IEA Net Zero Roadmap (2023): Targets data centre PUE <1.3 by 2025 while ignoring absolute growth trajectory (200→700 TWh).

So how’s that efficient? And that’s the efficiency paradox: optimising performance while simultaneously expanding infrastructure 10x. Great metrics with exploding carbon footprint has consequences that aren’t usually breaking news. This is why data without context is an illusional permission slip to grow more recklessly.

Training GPT-3 evaporated 700,000 liters of freshwater. That’s an equivalent to 14,000 person-days in water-scarce regions. That number sounds quaint next to GPT-4 and extrapolated GPT-5.

Model	Energy (MWh)	Water (ML)	Person-Days
GPT-3	1,287	0.7	14k
GPT-4	~38k	~20	400k
GPT-5	~380k–1.1B	10–30	200k–600k

Current metrics measure how efficiently you’re computing. They don’t ask whether you should be computing this much at all. They don’t consider the fact that resources may be finite, and so is the compute.

What makes Governance an Illusion?

Intergovernmental Panel on Climate Change (IPCC) pathways successfully established 1.5°C emissions targets across sectors, yet AI infrastructure expansion has proceeded without comparable regulatory frameworks. The EU AI Act mandates environmental reporting for high-risk AI systems. The US encourages voluntary audits. The International Energy Agency (IEA) sets PUE targets. None of them establish thresholds to trigger deterrence. The real gap is that compute expansion must be a governable choice rather than a technical inevitability.

Meanwhile, Ireland is dedicating 18% of its national power to data centers, without compute sovereignty, competitive “training races” are multiplying compute demand that overwhelms renewable capacity, and the Global South nations are bearing grid strain and water stress in return for nothing.

The policy gap isn’t efficiency measurement. The targets measured today are absolutely critical and foundational to CCDR. The gap, however, is the lack of behavioral accountability before the damage is done.

How Compute-to-Climate Deterrence Works?

The Compute-to-Climate Deterrence Ratio establishes a threshold-based governance system that triggers progressive interventions when organizations defined compute-climate limits. CCDR is based on the idea that climate, people, and environment cannot be an afterthought. It takes an intersectional approach that fundamentally diverges from current frameworks by creating an assessment framework that incorporates:

• Total compute consumption (measured in FLOPs i.e., Floating Point Operations)
• Environmental impact (CO₂e emissions, water consumption)
• Renewable energy ratio (fossil vs. clean energy)
• Economic subsidy (transparent tax recycling)
• Scarcity-driven escalation patterns (compute growth during competitive races)

That last part is critical. The scarcity mechanism is twofold: artificial scarcity that roots from capitalism, and the green compute capacity which is finite by design. Combined, AI training races create artificial compute scarcity. When organizations compete to build the next GPT/Claude/Gemini, they multiply infrastructure. CCDR treats this as governable behavior and penalizes the excess.

When organizations cross a certain threshold that exploits this scarcity, they cannot write off their Corporate social responsibility (CSR) spending as charity. CCDR is designed to actually hold the corporate socially responsible and accountable. The tax recycling addresses compute sovereignty while creating environmental equity. This revenue flows to designated climate crisis response and community renewable infrastructure funds, not general government revenue. Unlike carbon taxes that fund highways, CCDR mandates directed investment: climate response in the communities bearing infrastructure burden. This isn’t general taxation. It’s reparative allocation.

Why This Matters Now?

We’re at an inflection point. Pre-GenAI data centers generated ~60 million tons of CO₂ annually. Post-GenAI, we’re projecting 220 million tons by 2030. That’s a 700% increase in three years.

The IPCC says we have until 2030 to cut emissions 43% below 2019 levels to stay within 1.5°C warming. AI infrastructure is running in the opposite direction. Canada has committed to 40-45% emissions reduction below 2005 levels by 2030. These goals are incompatible without intervention. And the goal of CCDR is to provide the intervention mechanism that’s currently lacking.

While existing metrics measure efficiency, and academic research quantifies environmental impact, neither framework creates behavioural deterrence against unchecked compute growth. Those measurements are damage control, and the goal of CCDR is to prevent the damage from happening. The through-line: AI advancement without accountability creates harm at every scale–organizational, community, and planetary. The guardrails don’t need to choose between innovation and stewardship but make both possible.

Current AI governance asks: “Can we make this safe?” What it should be asking: “Should we be doing this at all? And if so, under what conditions?” CCDR is a framework for the latter question. It won’t solve AI’s climate crisis alone. But it fills a instrumental gap that addresses: translating compute expansion into binding climate accountability.

The alternative is unchecked infrastructure growth undermining 1.5°C pathways which poses risks far greater than governance friction. The story doesn’t have to end this way, and CCDR is what trying something different looks like.

Actionable Insights

AI infrastructure expansion is on a collision course with climate targets. Current governance mandates reporting without accountability.

• Policymakers & Regulators: It’s time to move beyond transparency obligations to enforceable thresholds. Efficiency metrics (PUE, CUE) are necessary but insufficient. We need to establish FLOP-based intervention triggers that escalate when compute expansion outpaces renewable capacity growth.
• AI Companies & Infrastructure Providers: Recognising CSR only to get a charity right off is not innovation. Adopt proactive compute budgets aligned with carbon budgets, prioritize renewable energy procurement commitments ahead of expansion.
• Climate Organizations & Researchers: Integrate AI infrastructure into emissions reduction strategies. Carbon budget frameworks must account for compute expansion as a discrete sector with enforceable caps, not rely on voluntary corporate commitments.

Only through threshold-based accountability can we align AI advancement with 1.5°C pathways. Accountability isn’t only compute expansion triggering progressive deterrence mechanisms, but also including the Global South for climate equity. The alternative is infrastructure growth that undermines every climate target we’ve set.

---

References

1. Hickel, J. (2020). Less Is More.
2. European Commission. (2024). EU AI Act. Official Journal of the European Union.
3. Goldman Sachs Research. (2024). AI data centers and climate impact projections.
4. IEA. (2024). Data centers and electricity consumption 2024 update. International Energy Agency.
5. IPCC. (2022). Climate Change 2022: Mitigation of Climate Change. AR6 Working Group III.
6. Kaack, L. H., et al. (2024). Machine learning and the environment: Opportunities and challenges. Nature Climate Change.
7. Li, P., Yang, J., Islam, M. A., & Ren, S. (2023). Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv preprint arXiv:2304.03271.
8. Luccioni, A. S., et al. (2023). Estimating the carbon footprint of BLOOM. Journal of Machine Learning Research.
9. Luccioni, A. S., Viguier, S., & Ligozat, A. L. (2019). Estimating the carbon footprint of language models. In AI for Social Good workshop, NeurIPS 2019.
10. United States. (2023). Executive Order 14110 on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. Federal Register.
11. Dodge, J., Prewitt, T., Luccioni, A.S., et al. (2022). Measuring the Carbon Intensity of AI in Cloud Instances. ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22), June 21–24, 2022, Seoul, Republic of Korea. ACM, New York, NY, USA, 25 pages.
12. Bresnihan, Patrick, and Patrick Brodie. 2023. “Data Sinks, Carbon Services: Waste, Storage and Energy Cultures on Ireland’s Peat Bogs.” New Media & Society 25 (2): 361–83.
13. Carbon Footprint of AI. Climate Impact Insights.
14. AI Data Centers Will Double Energy Consumption by 2030. Trax Technologies
15. What We Know About Energy Use At US Data Centers Amid The AI Boom. Pew Research Center.
16. Data Drains: Land and Water Impacts Data Centers
17. AI data centers may consume more power than whole cities
18. Understanding and Calculating Carbon Usage Effectiveness (CUE) in Data Centers
19. 2024 Iron Mountain Data Centers Sustainability Performance Overview