Will AI efficiency actually make the planet hotter by increasing emissions?

The usual story is that AI can optimize power grids and create better batteries, which in theory should reduce emissions. Yet a recent global energy-economy model points in the opposite direction: productivity gains enabled by AI could produce a net increase in CO2 emissions. The reason lies in the "rebound effect." When AI makes a process more efficient, energy savings do not simply remain intact. Instead, we often expand the process so dramatically that total energy consumption surges.

The Productivity Paradox

Why AI may increase emissions

When an LLM agent or an AI workflow makes a business process easier, it reduces production costs. In a capitalist economy, lower costs generally increase demand. If AI makes shipping 20% more efficient, companies do not merely leave 20% of their fuel unused. They ship 30% more goods because doing so has become cheaper. This is a modern application of Jevons Paradox to the compute era.

The model examines the issue from several angles:

Direct energy demand from AI

• Direct Energy Demand: Training and inference require enormous amounts of power. Even as hardware becomes more efficient, deployment is growing exponentially.
• Indirect Economic Stimulus: AI-driven GDP growth increases consumption across industries, not only within tech. Greater wealth generally correlates with more carbon-intensive activity.
• The Transition Lag: Even if AI helps identify a new carbon-capture material, moving that discovery from a laboratory to global industrial deployment takes far longer than launching a new data center.

Breaking Down the Impact

The variables behind this net increase are not encouraging. According to the model, efficiency gains in sectors such as logistics and manufacturing are canceled out by the immense energy demands of the infrastructure that supports AI.

Compute intensity and infrastructure growth

• Compute Intensity: Larger models are raising the minimum energy requirement for "basic" AI tasks.
• Infrastructure Build-out: New power plants and data centers create their carbon debt immediately, whereas the "green" benefits of AI optimization remain theoretical and delayed.
• Sectoral Leakage: Greater efficiency in one area often shifts energy use elsewhere, preventing the global CO2 curve from actually flattening.

Scaling AI without climate assumptions

For anyone building an AI workflow or planning deployment, "AI will fix the climate" is not a sufficient reason to scale without limits. Energy-aware computing requires a practical implementation guide. That involves optimizing prompt engineering to avoid token waste and selecting smaller, distilled models for tasks that do not require a frontier LLM.

The real-world takeaway is that productivity is a double-edged sword. If AI makes us 10x more productive, but we use that productivity to consume 11x more resources, we're moving backward. The only way to break this cycle is to decouple economic growth from carbon output, rather than assuming the software will magically handle the physics of energy consumption.