The Moral Panic Over AI's Carbon Footprint

Generative AI has acquired a peculiar moral status: the one technology people feel entitled to condemn on environmental grounds while scrolling Netflix in bed. The charge sheet is familiar. Data centres drinking rivers dry. GPU clusters devouring electricity that could power cities. Each ChatGPT query burning the equivalent of a small forest.

Most of it is wrong. Not harmlessly wrong in the way that myths usually are, but directionally wrong in a way that distorts where genuine concern should be directed. The numbers tell a more nuanced story, and one that looks very different when you account for what AI replaces rather than simply what it consumes.

What Your AI Session Actually Costs

Start with what most people actually do. A ten-minute back-and-forth with a model like Gemini or Claude involves roughly five to eight exchanges. Each exchange carries perhaps 200 tokens of input and 400 of output. That session totals somewhere around 4,000 to 5,000 tokens. At Sam Altman’s confirmed figure of 0.34 Wh per query, a ten-minute session costs approximately 2 Wh of electricity at the data centre. That is roughly equivalent to leaving a phone screen on for the same period, or to the energy consumed by a 55-inch television in under a minute of streaming.

One million tokens is a more abstract unit, but it has a useful anchor. It represents about 750,000 words, the equivalent of fifteen average novels. In usage terms, it is roughly what a fairly active user might accumulate across a full month of daily sessions. For a developer using AI coding tools heavily, it might be a single working day. At current inference costs on modern hardware, one million tokens consumes somewhere between 200 and 500 Wh. On the UK grid, that produces 40 to 100 grams of CO₂. Here’s what that looks like against familiar activities:

Activity	CO₂
10-minute AI session (~5,000 tokens)	~0.5 g
1 million tokens (text LLM)	~40–100 g
1 hour Netflix streaming	~36–55 g
1 hour PS5 gameplay	~40–80 g
1 hour Zoom call	~17 g
1 beef burger	~2,500 g
London to New York return (per passenger)	~1,700,000 g

The framing that AI is an environmental crisis, while ignoring the beef on the plate and the flight booked for Christmas, is not analysis. It is aesthetic prejudice dressed as ethics.

The imbalance is in modality. Text inference is cheap. Image generation consumes roughly sixty times more energy per query. Video generation exceeds that by another order of magnitude. The environmental footprint of generative AI is not evenly distributed, and the industry conflates a text query with a photorealistic video clip as if they belong in the same conversation.

A Month in the Life: Putting LLMs in the Carbon Ledger

Abstract per-token figures matter less than what they mean alongside everything else a person does. Consider a fairly typical UK adult in a month: they drive a petrol car, eat meat, watch streaming video, and use an AI assistant daily.

Typical UK adult, per month, approximate CO₂:

Activity	Monthly CO₂
Petrol car (UK average mileage)	~100–120 kg
Meat consumption	~90–110 kg
Home heating (gas)	~80–120 kg
Flights (annualised from typical 1–2 per year)	~100–200 kg
Video streaming (2 hrs/day)	~2–4 kg
AI assistant use (8 queries/day, typical)	~0.02 kg (20 g)

The AI figure is not a rounding error. It is below the threshold of detection against everything else. Giving up AI use entirely saves the same carbon as not driving for about eight minutes.

Now consider what shifting the other variables achieves:

Monthly CO₂ savings from lifestyle changes:

Change	Monthly saving
Go vegan	~70–90 kg
Switch to an EV (UK grid)	~80–100 kg
Eliminate flights	~100–200 kg
Stop streaming (replace with books, radio)	~2–4 kg
Stop using AI entirely	~0.02 kg

The hierarchy is unambiguous. Diet and transport are where personal carbon budgets live. The decision to use or not use an AI assistant is statistically invisible. Anyone genuinely motivated by their carbon footprint will achieve more by eating one fewer beef meal per week than by abandoning AI tools for a lifetime.

This is not an argument that scale does not matter. At a billion queries a day, small per-query costs aggregate into real infrastructure demand. But individual guilt is being misdirected, and misdirection has a cost of its own: it consumes the limited attention people have for making choices that actually move the needle.

Inference, Scale, and the Economics of Efficiency

Here the standard narrative inverts. Training is commonly assumed to be the great villain. Training does matter. But multiple analyses from Meta, AWS, and Google now show that sixty to ninety percent of a model’s total lifecycle emissions come from inference, not training. As model deployment scales, that proportion will only grow.

The energy cost of AI is not fixed. It is an ongoing function of queries served, efficiency achieved, and what powers the grid.

Recent testing on H100 hardware showed ten times better energy efficiency per token than the A100 generation two years prior. The Stanford AI Index found the cost to run a GPT-3.5-class model dropped 280-fold between late 2022 and late 2024¹. Computing delivers roughly a hundred-fold improvement in performance per unit of energy every decade.

If that trend continues, the energy cost of a token in 2030 may be a fraction of today’s figure, even as total token consumption grows enormously. Jensen Huang’s claim that AI’s energy demands per person will become “utterly minuscule” in ten years is optimistic on timing but defensible in direction.

The cost pressure driving AI providers to maximise tokens per watt is existential. Inference now represents the dominant operating cost for Anthropic and OpenAI. Every joule saved per token is margin recovered. The commercial incentive to make inference more efficient is, unusually for a technology sector, well-aligned with the environmental one.

Major hyperscalers have committed to matching electricity consumption with renewables. If AI inference shifts substantially to low-carbon grids through siting decisions or carbon-aware scheduling, the per-token CO₂ figure could fall dramatically independent of hardware improvements. These outcomes are plausible and largely invisible from the current debate.

The Substitution Question: What Energy Is Not Being Used?

This question is almost entirely absent from public discourse on AI and energy. Energy accounting for any technology is incomplete unless it asks what the technology displaces. When we evaluate the energy cost of electric vehicles, we compare them to internal combustion engines, not to walking. When we assess the carbon cost of teleconferencing, we subtract the flights not taken. Generative AI deserves the same treatment.

Consider a concrete case: a 100-million-token software project. This is a plausible scope for an AI-assisted web application built with extended reasoning and iterative development. At current inference costs, including thinking tokens, that volume might consume somewhere between 30 and 150 kWh of electricity. A reasonable central estimate is around 100 kWh.

Now consider the energy a three-person development team would consume building the equivalent project over three months. Each person in the US carries a total primary energy load of around 88,200 kWh per year. Even taking just the residential electricity share of approximately 4,400 kWh per person annually, three developers working for a quarter still consume roughly 3,300 kWh between them. The AI-assisted approach, even where the team is merely reduced or accelerated, represents an energy saving that dwarfs the inference cost many times over.

This is apples-to-oranges. Human biological and residential energy baselines exist whether someone is coding or sleeping; silicon compute draws power only when running. The two categories are not equivalent in kind, and the ratio overstates the direct substitution. But even accounting for that, the direction of the effect is clear, and the scale difference is large enough that no reasonable adjustment eliminates it.

A 2024 paper in Scientific Reports found that the human-to-LLM ratio for energy on equivalent written output is around forty to one in favour of the model for a typical large LLM, rising to over a thousand to one for a lightweight model.

There is a related substitution that is harder to quantify but potentially larger in scope: the work that simply would not have been done otherwise. The small business owner who builds a functional website with AI assistance rather than going without one. The researcher who synthesises a literature review in two hours rather than two weeks. The person in a lower-income country who accesses expert-quality medical explanation without access to a doctor. These do not displace a well-resourced human alternative. The counterfactual is not a human expert but the absence of one. The comparison is not AI versus human. It is AI versus nothing, and the cost of nothing is zero utility at any energy price.

The Genuine Concerns

None of the above excuses the industry from scrutiny. Several specific concerns are legitimate and underserved by current discourse.

The transparency deficit is the most urgent. Car manufacturers are required to publish fuel economy figures. AI companies are not. Anthropic, Google, and OpenAI operate at scale without publishing model-specific energy or carbon data. This prevents meaningful comparison, inhibits informed choice, and makes it impossible to hold companies accountable. Standardised inference efficiency reporting, in watt-hours or grams of CO₂ per million tokens, is an overdue requirement, not an ask.

The water question is often overstated at the individual query level but raises real concerns at the systems level. The issue is not global volume but geographical concentration. A data centre in Scotland drawing on abundant rainfall is not the same as one in Arizona drawing on the Colorado River, even if both carry nominally similar water use figures. Building AI infrastructure in water-stressed regions transfers environmental cost to communities that bear it directly. That deserves scrutiny independent of the headline figures.

The reasoning model exception belongs separately. The energy costs here apply to standard text inference. Reasoning models such as OpenAI’s o3 and DeepSeek R1, which generate extensive internal chains before producing output, consume thirty to fifty times more energy per query. A single task on the ARC-AGI benchmark for GPT-o3 reportedly consumed approximately 1,785 kWh, equivalent to two months of an average US household’s electricity. These are different products operating under a different regime. Treating them as equivalent to a chat query misrepresents both.

Finally, the aggregate growth curve. Even if per-token efficiency continues improving, total token consumption is growing faster. The trajectory of AI compute looks less like displacement of an existing technology and more like a new category of energy demand layered on top of everything else. The net outcome depends entirely on whether renewable energy build-out keeps pace, and whether AI use substitutes for energy-intensive alternatives rather than simply layering on top of them.

Proportion

The appropriate response to AI’s energy cost is not absolution and not alarm. It is proportion.

Text inference at current scale is a manageable fraction of digital energy demand, smaller per session than an hour of streaming, dwarfed by aviation, invisible against food systems. The trajectory of efficiency is positive, driven by commercial incentive as much as environmental concern. The substitution effects, while difficult to measure precisely, are directionally significant and absent from public accounting.

At the individual level, unless you are a heavy user of video generation or compute-intensive reasoning tasks, your AI use is not a meaningful lever on your carbon footprint. The burger costs more than the prompt. The flight costs more than the essay. The monthly ledger is settled by what you eat and how you travel, not by how often you ask a language model a question.

What demands sustained attention is institutional: the transparency that would allow AI energy use to be measured and compared; the siting decisions that determine whether data centre growth lands on stressed or abundant water and power systems; the efficiency incentives that will determine whether the token economy of 2035 runs on a fraction of today’s energy or a multiple of it. And most importantly, the counterfactual accounting that asks not just what AI costs, but what it replaces, and whether those substitutions reduce the net burden or merely shift it.

The environmental conversation about AI is worth having. What it absolutely does not need is another round of thoughtless moral panic. The self-righteousness of condemning AI while streaming Netflix for two hours is not a moral position. It is performance.

---

Footnotes

1. Stanford Human-Centered Artificial Intelligence, “The 2025 AI Index Report.” https://hai.stanford.edu/ai-index/2025-ai-index-report ↩