What the Paris Agreement teaches us about governing AI’s carbon footprint

Most AI governance frameworks overlook the energy systems and emissions burden powering AI’s ascent. That blind spot risks destabilising both AI’s sustainable rise, and global decarbonisation goals. Equally overlooked is AI’s social license to operate - the same force that phased out coal in the U.K., home of the first industrial revolution. Electrons are to the AI revolution what coal was to the industrial one. It is no surprise that data centre and compute power entities are investing in geothermal, nuclear, and long-duration storage - not primarily for climate goals, but to maintain social viability, ensure uninterrupted compute, and protect long-term investments. 

As AI accelerates, energy and emissions policy must move from the periphery to the centre of AI governance, especially as its primary inputs remain overwhelmingly fossil-based. As a result, net-zero ambitions are increasingly on a collision course with AI - one that could veer the Paris Agreement off track in ways both foreseen and under-explored. What remains overlooked is the physical infrastructure driving AI’s ascent: energy inputs - without a corresponding carbon management plan, such as Carbon Capture Storage (CCS), or an enforceable emissions governance framework. 

AI’s Paris Agreement moment: From carbon capture to global code

AI demand is already reshaping national grid plans, data centre policy, and energy strategies, each a direct touchpoint for national security. That’s why Equinor and Exxon Mobil - which deployed the world’s first CCS project in 1996 (non-enhanced oil recovery) - is now positioning itself as a power supplier to AI data centres by coupling natural gas for electrons with CCS. Microsoft’s recent push toward carbon-aware scheduling, and Amazon’s acquisition of renewable assets for AI workloads, underscore that compute is not just technical capacity - it is carbon accountability waiting to be formalised.

Each country has the right to define how it powers its future. Yet, climate change has already shown us that sovereign decisions taken in the absence of a holistic or global framework can lead to planetary-scale consequences. AI governance will fail if it assumes universal policy convergence. It will succeed if it respects national agency while enabling shared enforcement structures at the top. 

From my lived experience in shaping South America’s first carbon capture law via Brazil in 2024, I am formally proposing the creation of a United Nations Framework Convention on Climate Change (UNFCCC) - style mechanism for AI infrastructure - modelled after the governing body that negotiated and adopted the Paris Agreement in 2015. In practice, this would take the form of a Compute Emissions Transparency Authority (CETA) - a global mechanism to track, verify, and disclose the emissions associated with large-scale AI infrastructure. Just as the International Energy Agency (IEA) monitors energy flows, CETA would anchor AI governance with carbon  management (including CCS implementation and reporting, as ExxonMobil is trending toward this already with their AI move, since CCS storage sites provide such reporting information), providing the pillars for future restraint.

The Global South and the burden of invisible compute 

Moreover, the growing tension - between AI’s infrastructure demands and global emissions reduction goals - can be expected to be felt more acutely in the Global South, where compute-driven emissions may compromise climate goals not by local excess, but by hosting outsourced data centres or AI-intensive supply chains, as well as facing limited internet infrastructure compared to the developed world. Pakistan exemplifies how this dynamic could materialise in the Global South: while it does not yet possess Small Modular Reactors (SMRs), it does have nuclear capabilities. I forecasted a trajectory linking SMRs to AI infrastructure in a Pakistani videocast recorded after renewing the nation’s first low-carbon agreement—just as Amazon and Google were beginning to explore SMRs to power AI infrastructure in the U.S.

More broadly, the carbon pressure of models designed outside the Global South will increasingly constrain the futures of nations already grappling with historic emission and smog issues, such as Lahore in northern Pakistan - particularly if carbon management is not implemented alongside AI expansion. For this reason, AI policy is carbon policy. And carbon policy is energy policy. Energy builds infrastructure - and infrastructure defines sovereignty. To govern AI effectively, we must apply the intentionality of the 2015 Paris Agreement. This raises questions we can no longer ignore: Should non-essential AI usage - like viral content platforms - be deprioritised in favour of net-zero objectives? Should nations in transition to a low-carbon economy be permitted to regulate AI demand when emissions from data centres threaten their national targets? These are now urgent governance decisions. 

Call to optimism - a call to competence

With AI’s rise, we are at the edge of another technological precipice. However, we are not entering this frontier unprepared, or without precedent - the Paris Agreement demonstrated what can be achieved when nations band together, with CCS laws that build on this. We are not without tools, as emissions and carbon management technologies exist, along with conventional and emerging systems to power data centres. This is not a call to optimism - it is a call to competence. To remember what we have achieved, and to build, in the age of AI, with the same intentionality that can preserve our biosphere and shared future.

Discover on illuminem's Data Hub™ the sustainability performance of the companies mentioned in this article like Amazon, Microsoft, Exxon Mobil or Equinor.

illuminem Voices is a democratic space presenting the thoughts and opinions of leading Sustainability & Energy writers, their opinions do not necessarily represent those of illuminem.