Weighing the Costs and Benefits of Generative AI

1. Introduction

The sources compiled by Emma Andrea paint a sobering picture of generative AI's environmental footprint, resource consumption, labour exploitation, and inequality. These are real, documented issues that deserve rigorous attention. However, a one-sided catalogue of harms risks obscuring the equally real benefits, the rapid pace of mitigation, and the sheer impracticality of halting a technology already woven into the global economy. This report provides a structured counter-argument, assesses whether the "bad" outweighs the "good," and offers a forward-looking perspective grounded in the reality that generative AI is here to stay.

2. Deconstructing the Core Concerns

2.1 Energy Consumption: Contextualising the Numbers

The Claim: Data centre electricity use will approach 1,050 TWh by 2026, training a single model can power 120 homes for a year, and grid fluctuations require diesel generators.

The Counter-Argument:

• Scale vs. service provided: Training a frontier model like GPT-4 consumes roughly 50–60 GWh, or the annual electricity of ~5,000 US homes. Yet that single model serves billions of queries. Amortised over its lifetime, the energy per query is trivial: Sam Altman's figure of 0.34 Wh per ChatGPT query is less than a typical Google search (~0.3 Wh as far back as 2009). Even at 10×, it stays orders of magnitude below streaming an hour of video (~100 Wh) or driving a petrol car one mile (~1,000 Wh).
• Efficiency gains are already happening: The NVIDIA H100 delivers up to 9× the performance per watt of the A100 for inference. Frontier models are being distilled into smaller, task-specific versions running on a fraction of the energy.
• Renewables are absorbing growth: Hyperscalers are the world's largest corporate buyers of renewable energy. Google has matched 100% of its annual electricity with renewables since 2017 and targets 24/7 carbon-free energy by 2030. Their long-term power purchase agreements directly finance new wind, solar, and geothermal capacity.
• Diesel generators are a grid-protection feature, not a primary source: Milliseconds-to-minutes of frequency regulation is a tiny fraction of emissions, on par with hospital backup systems - a legacy grid issue, not a unique sin of AI.

2.2 Water Consumption: A Solvable Engineering Problem

The Claim: Data centres need ~2 L of water per kWh for cooling. GPT-3 training used 5 people's yearly water; Google's 2023 data centres used the equivalent of 206,906 people's yearly consumption.

The Counter-Argument:

• Location and technology matter enormously: The 2 L/kWh figure is worst-case evaporative cooling in hot, dry climates. Cooler-climate centres (Nordics, Pacific Northwest) often use zero water via air-side economisation; closed-loop systems recycle the same water indefinitely.
• "People's yearly consumption" is a misleading denominator: Total US data-centre water withdrawal in 2023 (~1.7 billion L/day) is under 0.5% of US freshwater withdrawals. Thermoelectric power uses ~133 billion L/day - 78× more. A single golf course in a dry region can use as much as a mid-sized data centre.
• Net water impact can be positive: AI-driven smart grids, precision agriculture, and leak detection save far more water than the servers consume. DeepMind cut Google's cooling energy by 40%, indirectly saving water too.

2.3 Hardware and Material Footprint: The GPU Chip is Just One Piece

The Claim: 3.85M GPUs shipped in 2023 (+44% YoY). Nvidia controls 94% of the market. The GPU chip contributes 81% of climate impact in the A100's cradle-to-gate analysis.

The Counter-Argument:

• The study focuses on a single, now-superseded GPU: The A100 is 7nm. Current H100/H200 and upcoming B200 on 4nm achieve far more computation per gram of material; newer cradle-to-grave assessments will show better compute-per-impact ratios.
• Concentration enables efficiency: A 94% share means one company's sustainability work propagates across the whole installed base. Nvidia's liquid-cooled racks already cut cooling overhead by up to 30%, and competition (AMD, Intel, ASICs) is emerging.
• GPUs are multi-purpose: The same chips accelerate climate modelling, drug discovery, fusion, and materials science - attributing their whole footprint to generative AI is inaccurate.

2.4 Labour Exploitation and Global Inequality

The Claim: 150–430 million data labourers underpin AI, often in poor conditions. AI may widen inequality and make most people poorer while enriching a few.

The Counter-Argument:

• The worst practices are being rooted out: The "ghost work" economy is now under rising regulatory and buyer scrutiny, moving toward formal employment and minimum-wage guarantees. AI itself reduces the most psychologically harmful moderation tasks by automating detection of graphic material.
• AI is a powerful equaliser of access: A farmer in rural Kenya can diagnose crop disease in Swahili for free; a student in Bangladesh gets high-quality physics tutoring without a private teacher. Like electricity and the internet, it starts with elites but trends toward broad development.
• Displacement fears lack historical perspective: Spreadsheets killed bookkeeping jobs but created financial analysis. 2024–2025 studies show the largest gen-AI productivity gains accrue to lower-skilled workers (e.g. +14% for AI-augmented call-centre staff), compressing wage gaps.
• The Global South can leapfrog: As mobile phones bypassed landlines, edge AI brings expertise to areas that never had enough doctors or engineers. Restricting AI would only cement existing advantages.

3. Does the Bad Outweigh the Good?

The "bad" is real but largely transient or manageable: grids are decarbonising, water use is shifting to closed-loop, hardware efficiency is on an exponential curve, and labour standards are rising under pressure. The "good" is a permanent structural uplift in human capability.

Verdict: the bad does not outweigh the good. The harms are front-loaded and visible; the benefits are diffuse and cumulative. Refusing to deploy AI over its footprint would be like refusing to build hospitals because they consume water and concrete.

4. The Future: Acceptance, Not Surrender

The fatalist line - "you can't fight it since it's here to stay" - can be reframed: acceptance is not surrender; it is the prerequisite for effective governance.

• Regulation is accelerating: the EU AI Act, the US Executive Order, and national data-centre moratoriums force transparency on energy and water - turning anecdotes into auditable, improvable metrics.
• Technological solutions are emerging: photonic computing, direct-chip liquid cooling, small modular reactors for data centres, and non-GPU models (spiking neural networks).
• Economic incentives align: energy is one of the largest costs for AI companies - efficiency is directly profitable.
• Public pressure works: reports like Emma Andrea's mean companies can no longer hide their impacts. Transparency → accountability → improvement.

5. Conclusion

The catalogue of consequences - 1,050 TWh of electricity, millions of litres of water, precarious data labour - must not be dismissed. But weighed against lives saved, knowledge democratised, and crises AI can help solve, the balance tips toward a net positive. The task is not to stop generative AI, but to harness it with eyes wide open, rigorous policy, and an unwavering commitment to mitigating its footprint while amplifying its promise. The worst outcome would be to let a fear of the bad prevent the good.

← Back to AI Forum