The AI paradox: from carbon cost to climate dividend?

Core Thesis

The core thesis of the article is that AI drives up energy demand in the short term, but over the long term, it can significantly reduce emissions by optimizing grids and industrial efficiency. This judgment diverges from market consensus — the market generally focuses on the surge in energy demand caused by AI (e.g., data center power consumption), while the article emphasizes AI's potential as an efficiency tool, arguing that its emission reduction effects may outweigh the emissions it generates. The article specifically notes that grids and industrial sectors are the most credible near-term application scenarios, as the business cases are clear and computing power was once a bottleneck.

Chain of Evidence

The author supports the thesis with the following data and reasoning:

1. Data Center Energy Consumption: Current Status and Future:

• In 2024, global data centers consumed approximately 415 TWh of electricity (1.5% of global electricity), with AI accounting for about 15%, corresponding to roughly 27 Mt CO₂ emissions.
• By 2030, data center electricity consumption is expected to reach 1,000 TWh (nearly 3% of global total), with AI as the primary driver.
• Emissions depend on grid carbon intensity: if current trends continue (360 g CO₂/kWh), emissions would be about 360 Mt; under a net-zero pathway (165 g CO₂/kWh), about 165 Mt.

2. Emission Reduction Potential from Grid Efficiency Improvements (quantitative estimates for three application scenarios):

Application Scenario	Description	2030 Emission Reduction (additional clean electricity)	Corresponding CO₂ Reduction
Moderate Application	Comprehensive improvements in renewable integration and demand flexibility	~415 TWh	~150–170 Mt
Widespread Adoption	Combining better forecasting, grid smoothing, and flexible demand	~1,600 TWh	~550–650 Mt
Aggressive Adoption	Fully embedding AI into grid management and market design	~3,000 TWh	Close to 1 Gt

3. Efficiency Gains in the Industrial Sector:

• Heavy industry, materials, and manufacturing account for 30% of global energy use, emitting approximately 12–13 Gt CO₂.
• Through predictive maintenance and process optimization (e.g., sensor modeling, digital twins), AI can reduce electricity consumption by hundreds of TWh, corresponding to hundreds of Mt CO₂ reductions, cumulatively reaching multiple Gt by 2030 (depending on adoption speed).

4. Long-Term Opportunities (higher uncertainty):

• Food systems: AI accelerates the development of plant-based proteins and fermentation technologies, which can replace animal products, offering significant emission reduction potential but limited by consumer acceptance.
• Direct air capture: AI can optimize chemical and material efficiency, but progress remains uncertain.
• Weather and climate modeling: Improves extreme event prediction and renewable energy dispatch, indirectly reducing emissions while enhancing system resilience.

Companies/Themes Involved

• Arm: The article explicitly mentions at the outset that Arm's low-power chip design can enhance AI energy efficiency in data centers and edge devices, supporting large-scale emission reductions. Author's stance: Bullish, viewing it as a key technological foundation for AI energy efficiency improvements.
• Grid and Industrial AI Applications: No specific companies are named, but the theme is clear. Author's stance: Actively following, believing these areas have clear business cases and represent near-term investment opportunities.
• Food Systems, Direct Air Capture: As long-term themes, the author's stance is: Neutral observation, acknowledging potential but emphasizing high uncertainty.

Investment Implications

• Actionable Directions:

1. Focus on energy efficiency in AI infrastructure: Invest in segments such as low-power chip design (e.g., Arm), AI-driven grid optimization software, and industrial automation platforms (e.g., predictive maintenance, digital twins).

2. Position for clean energy and grid upgrades: AI's release of grid capacity will increase demand for renewable energy and transmission infrastructure, involving related operators and equipment manufacturers.

3. Be wary of short-term emission risks: The front-loaded emissions from new data center capacity (2025–2030) may intensify environmental regulatory pressure; watch for carbon market-related investment opportunities.

• Perspective Bias: Baillie Gifford, as a growth-oriented investment firm, is naturally inclined to favor AI's long-term innovation narrative, potentially underestimating the impact of short-term energy demand surges on social acceptance and regulation. Investors need to balance the optimistic outlook for AI efficiency gains with the actual pace of implementation.

---

Core Argument

This chapter explicitly identifies two major risks of AI in the energy transition that could offset its potential emission reduction benefits. The first risk is “brown AI” — AI can also be used to optimize the production efficiency of fossil fuels, thereby extending the lifecycle of high-emission assets. The second risk is the “rebound effect” — efficiency gains reduce energy costs, and without carbon pricing or regulatory constraints, rising total consumption may partially offset the emissions saved. The article suggests that the market generally focuses on AI’s positive impact on clean energy but underestimates these counteracting effects.

Chain of Evidence

This chapter does not provide specific data or quantitative analysis, but merely presents the risk mechanisms through logical deduction. The author constructs the argument via two parallel risk points:

• Brown AI: AI tools that improve efficiency in the clean electricity sector can equally be applied to the exploration, extraction, transportation, and distribution of fossil fuels, lowering their costs and thereby delaying the phase-out of high-emission assets.
• Rebound: Efficiency gains are typically accompanied by cost reductions, which in turn stimulate additional demand. Without carbon pricing or regulatory constraints, overall consumption may increase, partially offsetting the efficiency gains brought by AI.

The conclusion further emphasizes: AI itself is not a climate solution. In the short term, it increases energy demand (especially in carbon-intensive grid regions); in the long term, the emission reduction effect depends on whether companies decarbonize quickly and whether AI is applied to the most inefficient parts of the system.

Companies/Themes Involved

• Brown AI: A theme the author views as a risk. No specific companies are named, but it implies that traditional oil and gas companies or AI service providers may use AI to improve fossil fuel efficiency.
• Rebound: An economic behavior risk, which the author sees as a potential offsetting factor requiring policy intervention. No specific companies are mentioned.

Investment Implications

• Investors should be wary that the “emission reduction promises” implicit in AI investments may be reversed in application or offset in aggregate. Specifically:
• Watch whether fossil fuel companies use AI to reduce operating costs, thereby extending asset life — this could mean the “stranded asset” risk implicit in their valuations is underestimated.
• Watch for the energy demand growth driven by AI data center expansion. Without policy constraints, efficiency gains may be fully offset by demand growth, putting the “green” narrative of related companies at risk of delivery failure.
• The author's firm, Baillie Gifford, is a growth-oriented asset manager. Its long-term investment perspective may naturally lean toward believing in the positive effects of technological progress. However, the chapter’s candid discussion of risks shows a degree of balance, though it does not delve into quantifying the probability or magnitude of the risks.