My Take on the Ethics of LLMs

I studied computer science from 2009-2013, received my Bachelors in Science with Honors, then soon after I did my Ph.D. in a niche form of artificial intelligence from 2015-2019. Then I did a couple of postdocs, applying AI to various problems (genomics, satellite imagery, evolution, video games), then I started as a data scientist at the University of Michigan, right at the start of the GenAI boom. I helped bring GenAI to disparate fields of research, I warned researchers of the myriad dangers of LLMs: from use in medicine, to environmental impact, to intellectual property and plagiarism concerns, to training data often stolen from artists without consent, to educational concerns both as teachers and for students using LLMs.

No one, however, is an expert in all of AI. My Ph.D. is in a niche field of AI, but I never studied generative AI during my studies - how could I? I finished my degrees long before ever hearing of ChatGPT.

I also work with AI, and still write research in the field, but ultimately I'm just another fish in a massive pond. So with my background out of the way, you know I have some, but not all, context.

So, say someone posts a screenshot of ChatGPT and a stranger appears in the replies to inform them, with great moral certainty, that every query they send is boiling another bottle of water and killing another tree. The message is usually delivered from an iPhone built on cobalt mined by children in the DRC, while a 4K Netflix stream plays in another tab. I find this ignorant and hypocritical, but I'm mostly irked that the argument is being made against the consumer of a product, rather than against the systems that produced it.

The "AI is destroying the planet" line is not exactly wrong, but it is badly aimed. The real story is more complicated, and the conclusion is different from what the standard outrage suggests. I want to lay out where I actually stand on this, with numbers attached.

Conclusion up front, so you can decide whether to bother with the rest: your guilt about the environmental and ethical cost of LLMs is justified, but it is being aimed at the wrong target. The industry has deliberately structured things so that your anxiety lands on your own fingertip hitting the query button, while the actual decision-makers (hyperscalers choosing fossil fuels, regulators choosing inaction, venture capital choosing scale over sustainability) stay invisible. BP ran the same playbook when it popularised the personal "carbon footprint" in the early 2000s: a PR campaign designed to shift blame for climate change away from corporate polluters and onto the everyday choices of consumers.^[1]

Some Numbers

Training costs in absolute dollars have grown. GPT-3 in 2020 cost around $4.6 million to train.^[3] GPT-4 in 2023 cost somewhere between $63 million and $100 million.^[2] The headline reads like a runaway curve.

Per-parameter cost is a crude proxy, but it's still a useful model. GPT-3 cost roughly $0.000026 per parameter. GPT-4, despite the eye-watering total, comes in at around $0.00005 per parameter for a model roughly ten times the size, multimodal across text, images, audio and video, with a 128k context window instead of GPT-3's 2k.^[3] Parameters are not units of capability, especially under MoE (more on that below), so this is not a clean efficiency metric. Per benchmark point on standard evaluations, training has gotten cheaper, not more expensive.^[3]

Two things drove that. First, the Chinchilla scaling laws (DeepMind, 2022) showed that almost every model in the field was undertrained, and that compute-optimal training wants about 20 tokens per parameter.^[4] The field stopped bolting on more parameters and started feeding the models properly. Second, Mixture-of-Experts (see jargon section above) broke the link between total size and per-query cost. DeepSeek's V3/R1 reported around $5–6 million for the final training run on a 671B-parameter MoE model that holds its own against frontier reasoning models.^[4] That figure is the marginal compute cost of the final run only; total programme cost (R&D, prior runs, salaries, infrastructure) is much higher. Even allowing for that, a 671B-parameter dense GPT-3-style model trained the old way would have cost hundreds of millions.

Capability per training dollar is going up, and the published literature is consistent on that.

You can run near-cutting-edge AI on a gaming PC

The other thing the doom narrative misses is what has happened at the small end. Meta's Llama 3 8B was trained on 15 trillion tokens, 1,875 tokens per parameter, almost 100x the Chinchilla-optimal ratio. Alibaba's Qwen3-0.6B pushed it to 60,000:1.^[4] The bet is to spend much more on training a small model so that inference, the part you pay for forever, is almost free.

The result is that I can sit at my desk with an Nvidia RTX 4090 (24 GB VRAM, around $1,600 new) and run a quantized Llama 3.1 70B locally. Offline. Private. No API key. No subscription. With a more modest 16 GB card like a 4070 Ti Super, you can comfortably run Phi-3 14B or Mixtral 8x7B. Even an 8 GB RTX 4060 will run Mistral 7B or Llama 3.1 8B at usable speeds.^[5]

If you have a recent gaming PC, you already own infrastructure that would have been classed as a research-lab supercomputer in 2020. I have written about how I actually use this in my post on running a local LLM as a professional second brain. The point worth landing here is that the centralised, cloud-only, beg-OpenAI-for-access model of AI is not the only option, and it is increasingly not even the best one (security, costs, ethics, requires internet, etc.).

Energy, Water, AI and Netflix

To be clear, I'm very much on team "let's not destroy our planet in exchange for Studio Ghibli Facebook profile pictures", but I'm also on team "let's not destroy our planet in exchange for doom-scrolling TikTok for 5 hours while bedrotting".

A single ChatGPT query uses somewhere between 0.001 and 0.01 kWh, emits roughly 0.7 to 4.3 grams of CO2, and consumes around 10 to 30 millilitres of water in datacenter cooling.^[6][7][8] Those numbers are probably smaller than you thought - "a whole glass of water per query" is what I saw a lot on social media: exaggeration and lying cause short term gains to a movement but long term losses, so let's look at the real numbers.

For context, here is what one hour of various normal digital activities looks like:

Activity	Energy	CO2	Water
1 ChatGPT query	~0.003 kWh	~3 g	~10–30 ml
1 hour Netflix in HD/4K	~0.077 kWh	~34 g	2–12 L
1 hour Zoom call	N/A	N/A	~1.7 L
1 hour social media scrolling	N/A	N/A	~430 ml
15-mile car commute	N/A	~6,000 g	N/A

Sources: Bowdoin College energy brief (2025),^[6] JISC National Centre for AI (May 2025),^[7] Online Learning Consortium (2025).^[8] Ranges are wide because they depend on hardware, location and grid mix. One caveat on the comparison: the ChatGPT figure covers data center cooling only, while the Zoom and Netflix figures include user device power and network overhead. The boundaries are not identical, and a heavy LLM session (agentic loops, long reasoning chains, image generation) can run hundreds of queries per hour. The point is that a casual ChatGPT exchange is not the carbon villain it is made out to be, not that all AI use is small.

An hour of HD streaming uses the energy of roughly 26 casual ChatGPT exchanges. A one-hour Zoom call uses more cooling water than around 50 of them.^[7] Nobody is being scolded online for joining a Zoom or finishing a Netflix series. The outrage is highly specific to one technology, and it does not track the actual numbers.

Training a frontier model is expensive. GPT-3's training emitted about 500 metric tons of CO2, roughly the annual emissions of 110 average US passenger cars.^[9] That is a real cost, but a one-time cost, amortised over hundreds of millions of users and billions of queries, and it is dwarfed within weeks by the operational footprint of, say, global TikTok consumption.

The cobalt in your hand

The hidden hypocrisy is the hardware in your hand. The phone you are holding while you tweet about boycotting ChatGPT contains lithium, rare earths, and cobalt. A lot of that cobalt comes from the Democratic Republic of the Congo, much of it mined under conditions Siddharth Kara documented in Cobalt Red (artisanal mining, child labour, almost no safety equipment, rivers running orange, slave labor).^[11] Those mines feed the batteries in your phone, your laptop, your earbuds, and the GPU clusters AI runs on.

Embodied carbon is a much bigger part of the story than most coverage admits. Gupta et al. (Chasing Carbon, IEEE Micro 2021) showed that as facilities run on cleaner electricity, manufacturing the servers approaches and can exceed operational emissions over the device lifetime: as you decarbonise the grid, the ratio swings further toward embodied. On a dirty grid (40% gas, 15% coal) operational emissions still dominate, which is why grid composition (covered later in this post) still matters. The same logic applies to consumer hardware: estimates put the renewal of AI processors and consumer electronics at up to 2.5 million tonnes of e-waste a year by 2030, roughly the mass of 13 billion iPhone 15 Pros. Currently we recycle about 12.5% of it.^[10]

I am not arguing that two wrongs make a right. I am arguing that the application of a moral argument used to criticise AI should also condemn the device you used to type the criticism. Further, I am arguing that the focus should be on systemic issues (why is your phone made from slave labor in the first place?) rather than blaming the individual for using a phone. The same goes for AI.

The Jevons trap

The counter to all of the above is the Jevons paradox, and I want to take it seriously because it is the strongest argument the doom side has.

William Stanley Jevons, in 1865, noticed that as steam engines became more efficient at burning coal, England burned more coal, not less.^[12] Cheaper energy unlocked new uses, more industries adopted steam, and total consumption shot up despite per-engine efficiency. The pattern repeats: more efficient cars led to more driving. Cheaper LED lighting led to more lighting. Cheaper compute led to vastly more compute.

It applies to AI. As models get cheaper to run, demand expands to fill the new capacity.^[13] Medical imaging is an instructive case. In 2016 Geoffrey Hinton famously suggested radiologists would soon be obsolete; the opposite happened. The number of radiologists has grown, and CT scan rates in US emergency rooms have roughly doubled over the last 15 years. AI is one of several factors making imaging cheaper and faster to interpret rather than the sole cause, but the broader pattern, that capacity-cheapening tech tends to expand demand rather than reduce it, holds.

The same pattern is showing up in software. When DeepSeek announced R1 in early 2025, the market wiped about $600 billion off Nvidia's valuation on the assumption that more efficient models would mean fewer chips needed. The valuation recovered. Cheaper inference unlocks agentic systems making thousands of recursive calls in the background. It unlocks LLMs embedded in operating systems, IDEs, search bars, browsers, every form field. Total inference is moving from billions of queries a day toward trillions.

So no, software efficiency alone will not save us. The 4-bit quantized 8B model on my desk is not a get-out-of-jail-free card. It just means more people will run more models more often. That is genuinely a problem the field has to confront.

But notice this does not land on "the individual user is the villain" -- it lands on the underlying physical infrastructure: the grid, the cooling, the silicon supply chain. We have to raise a stink about that infrastructure.

The access argument

Telling people not to use LLMs on environmental grounds is, very often, a position only available to people who don't need them.

About 92% of low-income Americans, roughly 36 million people, cannot afford a lawyer for civil matters.^[14] They are routinely facing landlords, debt collectors and employers who can. An LLM that helps a pro se litigant fill in housing-defence paperwork correctly is not a luxury. The Suffolk LIT Lab and Legal Aid of Eastern Missouri are already using LLMs to streamline legal-aid intake.^[15] The rAInbow chatbot helps domestic-abuse survivors understand the legal protections available to them.^[14] The early standalone "AI lawyer" products (DoNotPay being the obvious example, which paid an FTC settlement in 2024 over deceptive marketing) are not the case I am making here. The case is for LLMs as scaffolding inside properly-supervised legal-aid workflows.

The medical case is at least as strong. The WHO projects a global shortage of around 11 million health workers by 2030, mostly in low- and middle-income countries, where roughly 8 million people die every year from treatable conditions due to under-resourced triage.^[16] Penda Health in Nairobi has integrated an LLM-based decision-support tool into its records system to reduce diagnostic errors. PROMPTS, also in Kenya, uses AI over SMS to coach expectant mothers on birth preparedness.^[16] In the US, AI text-message interventions are being used in Diabetes Self-Management Education programs that were previously too expensive to deliver at scale.^[17] And translation alone, the simple ability for someone with limited English to get a clear plain-language explanation of a diagnosis, makes a real difference for people who would otherwise nod along without understanding.^[17]

Then there's education. Trainee doctors at well-funded medical schools are using premium LLMs to map out diagnostic reasoning. If we tell students at underfunded institutions they shouldn't use the same tools, we have built what some researchers (correctly, I think) call epistemic exclusion: the rich keep the cognitive support, the poor are told it would be unethical of them to take it.^[18][19]

The pattern repeats everywhere. Wealthy law firms will burn through API tokens and will not feel one ounce of guilt about it. Tech executives will continue to use these tools to widen their advantage. If a single mum studying nursing part-time, or an immigrant trying to write a tenancy complaint, or a teenager on free school meals trying to understand calculus is told using AI is unethical, they pay for the sins of others. It's like blaming the family on a tight budget for grabbing a plastic carrier bag at the checkout, or for buying the cheap supermarket burger because that's what fits the weekly shop, instead of blaming the supermarkets churning out billions of single-use bags and the industrial meat suppliers behind the cheap mince. Reusable bags and vegan groceries are great if you can afford the time, money and nearby shops to make them work; for a lot of people they aren't a real choice, and the moral weight belongs upstream with the corporations that designed the system, not with the person at the till.

I am not suggesting using AI for medical or legal purposes. I'm pointing out that people do every single day because they have no alternatives. Saying they shouldn't doesn't change that they do.

If you care about marginalised communities, the move is to push them toward locally-run open-source models like Llama, Mistral and Phi, not to talk them out of the technology entirely. They get the benefits, and the centralised cloud bill goes elsewhere. Or even a "library" type solution where communities can access local LMs on more powerful machines remotely (from phone apps, or even library computers) without the need to afford a modern gaming PC. We also need smaller, more powerful models, to open accessibility, so these models CAN run on people's phones and 5 year old 3rd hand laptops.

So where is the actual problem?

The problem is the infrastructure layer, and it is mostly a regulation problem.

Data centers used 1–2% of global electricity in 2024. The IEA expects this to more than double by 2030, exceeding 1,000 TWh.^[20] In the US specifically, data center consumption is projected to roughly double from 183 TWh in 2024 to around 426 TWh by 2030.^[25] That is real load, on a real grid, and it is being added faster than zero-carbon supply.

Three things are going wrong simultaneously, and they are all fixable by policy rather than by yelling at end users.

1. Tax breaks without strings. States are still bidding against each other to attract data centers with sweeping tax incentives, and almost none of those deals require "renewable additionality": a clause that says the data center must fund new clean energy capacity, not just plug into the existing grid and let residential rates rise to cover the upgrades.^[21] Singapore, Amsterdam and Dublin have moratoria and stricter performance rules. Germany has legislated efficiency requirements. New York, as of 2025, requires registration for data centers over 500 kW.^[22] Most of the US still doesn't.

2. Cooling is stuck in the 1990s. Cooling is up to 50% of data center energy. Most of it is air cooling, where air handlers chill huge rooms and evaporative towers throw freshwater into the sky (around 80% of the freshwater drawn evaporates).^[22] Direct-to-chip liquid cooling, where coolant runs in a closed loop over the actual hot silicon, is a much better answer. Microsoft published a Nature life-cycle assessment showing that switching from air to cold-plate cooling cuts greenhouse emissions by 15–21%, total energy demand by 15–20%, and "blue water" use by 30–52% over a facility's lifetime.^[23] The technology exists, we just don't make companies use it, while incentivizing them to continue full steam (pun intended) ahead.

3. The grid is dirty. Around 40% of data center electricity globally comes from natural gas, with another 15% from coal.^[20] Goldman Sachs estimates 60% of new data center demand over the next decade will be met by burning fossil fuels, adding roughly 220 million tonnes of CO2.^[25] Renewables alone won't fix this, because data centers run 24/7 and grid-scale battery storage isn't there yet. The answer for baseload is nuclear, and right now we are watching it happen the wrong way. Microsoft is restarting Three Mile Island under a private 20-year power purchase agreement. Amazon is co-locating Small Modular Reactors with Dominion Energy next to its Virginia campuses. Michigan's Palisades plant is restarting in 2026 backed by a $1.52 billion DOE Loan Programs Office guarantee to a private operator, Holtec.^[24] Hyperscalers get the megawatts; the public gets the construction risk and the long-term liability. The US target of quadrupling nuclear capacity from 100 GW to 400 GW by 2050 is the right scale, but ownership matters. Nuclear built by public utilities, sold at a regulated rate to whoever needs it, would serve the people paying for it, while nuclear built behind a hyperscaler's fence does not.

I know "support nuclear" is awkward in environmental circles, so I'll be plain. The problem with renewables as a sole solution is intermittency: when the wind drops or the sky clouds over, something has to fill in, and at present that something is almost always gas. Nuclear runs at full capacity around the clock regardless of weather. Its lifecycle carbon footprint, including mining, construction, and decommissioning, is comparable to wind and to utility-scale solar (IPCC AR6 medians: nuclear ~12 gCO2/kWh, onshore wind ~11, utility solar PV ~48; rooftop solar in sunny grids can come in lower).^[26] Measured per unit of energy produced, nuclear kills fewer people than any fossil fuel by orders of magnitude, mostly because it produces no air pollution.^[27] One uranium fuel pellet the size of a fingertip carries the energy of a tonne of coal.^[28] The waste is real, has to be managed, and is a legitimate concern, but it is small in volume and containable. Modern Gen III+/IV designs (AP1000, NuScale and similar SMRs) use passive safety: gravity, natural circulation, convection. They shut down on their own, without active intervention. Many of the reactors being restarted right now are older Gen II PWRs, so the "passively safe" claim applies most strongly to the new builds rather than the restarts. Chernobyl and Fukushima were older designs operating under conditions modern reactors are built to survive.^[27] Opposing nuclear while accepting a grid that burns coal is not a coherent environmental position. Opposing private hyperscaler nuclear, and demanding public nuclear instead, is.

All of this needs legislators to do their jobs and to keep the resulting capacity in public hands. Your guilt shouldn't be about hitting a button, it should be about what you're doing about the people who decided to build this infrastructure, on this grid, with these incentives, and who is profiting while you carry the moral weight.

Where I actually land

I've spent enough of my career around the actual science, and around the scientists, to see a bit deeper into the systemic issues and propaganda blaming individuals, so I see the cartoon version of this debate just keeps eyes away from the increasing profits of oligarchs while real problems persist and worsen.

The cartoon "AI is going to save us" pitch is rubbish. There are real costs, real bias, real labour issues, real water and grid concerns, and a real Jevons trap that means efficiency alone won't bail us out.

The cartoon "using ChatGPT is killing the planet" pitch is also rubbish, and it is the more harmful of the two, because it is aimed at the people with the least leverage. It lets the actual decision-makers (utilities, regulators, hyperscaler procurement teams, planning boards) off the hook, and it tells the people who would benefit most from these tools that they are bad people for using them.

If you genuinely care about the environmental cost of AI, here is the move:

• Run small models locally where you can. Llama 3.1 8B on your own machine beats a cloud round-trip on almost every axis (privacy, latency, cost, footprint).
• Pressure your representatives on data-center siting, renewable additionality, and mandatory liquid cooling. The decisions that actually move the numbers happen at planning boards and utility commissions, not on your keyboard.
• Support public nuclear. The grid needs zero-carbon baseload, and there is no other technology that does it at scale today. Build it through public utilities at regulated rates, not behind hyperscaler fences where the public pays the construction risk and a single tenant gets the electricity.
• Cut a Netflix hour before you cut a query that helps you do your job. The numbers are not close. Or even start your own plex server or run physical media (DVDs, Blu-rays, often free to borrow from libraries).
• Stop telling people in worse situations than yours that one of the most useful tools available right now is morally off-limits to them. The wealthy law firm down the road is not going to feel guilty on their behalf.
• Keep some perspective on scale. A single short LLM prompt uses around 0.4 Wh of electricity and 10 to 50 ml of cooling water; a heavy 70B-class prompt can hit 29 to 33 Wh.^[29][22] A 100 g beef patty embodies roughly 1,540 litres of total water (about 77 litres even if you only count "blue" freshwater) and thousands of Wh of lifecycle energy.^[30][31] One burger is in the same ballpark as 1,500 to 7,700 short AI queries on freshwater, and thousands of queries on energy. The aggregate AI footprint still matters because the per-query number is multiplied by hundreds of millions of users a day,^[29] but "AI is destroying the planet" while you eat a burger and stream Netflix is not a coherent position.
• Buy circular, modular, and transparent hardware where you can. No modern high-performance device is genuinely "conflict-free" or cobalt-free; the supply chains are too tangled for that. Fairphone is the clearest example for smartphones: they invest directly in mining conditions through the Fair Cobalt Alliance and build modular devices designed to last. For everything else, refurbished and certified pre-owned (Apple Refurbished, dedicated circular-economy retailers) extends the life of devices that have already been built. For new purchases, look for brands that audit their smelters through the Responsible Minerals Assurance Process, currently the best available check against funding armed conflict in the DRC.

References

1. The fossil fuel industry invented the carbon footprint to shift blame onto consumers. BMJ, 383, p2553 (2023). doi:10.1136/bmj.p2553
2. Cottier, B., et al. (2024). The rising costs of training frontier AI models. Epoch AI. Standard reference for GPT-4 training-cost estimates ($63M–$100M range). epoch.ai
3. Maslej, N., et al. (2024). Artificial Intelligence Index Report 2024. Stanford Institute for Human-Centered AI (HAI). Authoritative tracking of foundation-model training costs and compute usage. aiindex.stanford.edu
4. Hoffmann, J., et al. (2022). Training Compute-Optimal Large Language Models. arXiv:2203.15556. arxiv.org/abs/2203.15556. DeepSeek-AI (2024). DeepSeek-V3 Technical Report. arxiv.org/abs/2412.19437. DeepSeek-AI (2025). DeepSeek-R1: Incentivising Reasoning Capability in LLMs via Reinforcement Learning. arxiv.org/abs/2501.12948. Qwen Team (2024). Qwen2 Technical Report. arxiv.org/abs/2407.10671. Primary sources for compute-optimal training, MoE, and frontier-model training cost claims.
5. AI VRAM Requirements: How Much GPU Memory for Every Model. LocalAIMaster. localaimaster.com
6. Energy and Water Consumption of Technologies. Bowdoin College. bowdoin.edu
7. Artificial intelligence and the environment: Putting the numbers into perspective. JISC National Centre for AI, May 2025. nationalcentreforai.jiscinvolve.org
8. The Real Environmental Footprint of Generative AI: What 2025 Data Tell Us. Online Learning Consortium, 2025. onlinelearningconsortium.org
9. The Carbon Footprint of AI (GPT-3 training emissions ~500 t CO2). Climate Impact Partners. climateimpact.com
10. Gupta, U., et al. (2021). Chasing Carbon: The Elusive Environmental Footprint of Computing. IEEE Micro, 41(4), 44–53. Primary source for hardware embodied-carbon and the operational-vs-manufacturing split in hyperscale facilities. doi:10.1109/MM.2021.3061394
11. Kara, S. (2023). Cobalt Red: How the Blood of the Congo Powers Our Lives. St. Martin's Press. Backed by primary reporting: Amnesty International (2016). "This is what we die for": Human rights abuses in the Democratic Republic of the Congo power the global trade in cobalt. amnesty.org
12. Jevons, W. S. (1865). The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of our Coal-Mines. Macmillan, London. Original formulation of the rebound effect. oll.libertyfund.org. Modern survey: Sorrell, S. (2009). Jevons' paradox revisited: The evidence for backfire from improved energy efficiency. Energy Policy, 37(4), 1456–1469. doi:10.1016/j.enpol.2008.12.003
13. Freire-González, J. (2020). Energy efficiency and rebound effects in energy transitions. Energy Policy. Theoretical framework for rebound effects. The CT-scan rebound and the Nvidia market reaction discussed in-text are widely-reported industry events rather than findings from this paper.
14. The Promise and Peril of AI Legal Services to Equalize Justice (92% civil-legal-needs gap; rAInbow). Harvard JOLT. jolt.law.harvard.edu
15. Can LLMs help streamline legal aid intake? (Suffolk LIT Lab; Legal Aid of Eastern Missouri). Stanford Justice Innovation. justiceinnovation.law.stanford.edu
16. AI for social good: Making AI work for health systems (WHO 11M shortage; Penda Health; PROMPTS). J-PAL / Abdul Latif Jameel Poverty Action Lab. povertyactionlab.org
17. Leveraging large language models to foster equity in healthcare (DSME/S text interventions; translation/plain-language access). PMC / NIH. pmc.ncbi.nlm.nih.gov
18. Not all AI is created equal: considerations for equity in medical education. PMC. pmc.ncbi.nlm.nih.gov
19. New language-learning algorithms risk reinforcing inequalities. Ford School, University of Michigan. fordschool.umich.edu
20. Energy supply for AI (data-center share of global electricity; gas/coal mix). International Energy Agency (IEA). iea.org
21. What Happens When Data Centers Come to Town? (tax incentives, renewable additionality). Ford School STPP. fordschool.umich.edu
22. Li, P., et al. (2023). Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv:2304.03271. arxiv.org/abs/2304.03271. Jurisdictional and evaporative-loss figures corroborated by Global Action Plan (2026), Not a drop to drink. globalactionplan.org.uk
23. Microsoft quantifies environmental impacts of datacenter cooling from cradle to grave in new Nature study (cold-plate vs air cooling LCA). Microsoft. news.microsoft.com
24. From Atoms to Algorithms: Nuclear Energy's Comeback in the Age of AI (Three Mile Island PPA; Amazon SMRs; Palisades restart; 100→400 GW target). Georgetown Environmental Law Review. law.georgetown.edu
25. Goldman Sachs (April 2024). Generational growth: AI, data centers and the coming US power demand surge. Global Investment Research. Primary source for the US 183→426 TWh demand projection and the 60% fossil-fuel grid ratio. goldmansachs.com
26. Five reasons the clean energy transition needs nuclear power (lifecycle carbon footprint comparable to wind). International Atomic Energy Agency (IAEA). iaea.org. Lifecycle medians cross-checked against IPCC AR6 WG3 Annex III.
27. Safest sources of energy (deaths per unit of energy; nuclear vs fossil fuels). Our World in Data. ourworldindata.org
28. Nuclear power is the most reliable energy source and it's not even close (energy density; capacity factors). U.S. Department of Energy, Office of Nuclear Energy. energy.gov
29. Jegham, N., Abdelatti, M., Elmoubarki, L., & Hendawi, A. (2025). How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference. arXiv:2505.09598. arxiv.org/abs/2505.09598. Per-prompt energy/water for short and 70B-class long prompts; aggregate-scale framing. Husom, E. J., Goknil, A., Shar, L. K., & Sen, S. (2024). The Price of Prompting: Profiling Energy Use in Large Language Models Inference. arXiv:2407.16893. arxiv.org/abs/2407.16893.
30. Aboagye, I. A., Valappil, G., Dutta, B., Imbeault-Tétreault, H., Ominski, K. H., Cordeiro, M. R. C., Kröbel, R., Pogue, S. J., & McAllister, T. A. (2024). An assessment of the environmental sustainability of beef production in Canada. Canadian Journal of Animal Science, 104, 221–240. doi:10.1139/cjas-2023-0077. Lifecycle energy intensity of beef production.
31. Trujillo Nava, A., Valdivia Alcalá, R., Hernandez Ortíz, J., & López Santiago, M. A. (2023). Estimation of the water footprint in the production of beef from European cattle in Mexico. Agro Productividad. doi:10.32854/agrop.v16i7.2453. Global-average ~15,400 L/kg beef figure. Spore, T., Mekonnen, M., Neale, C., Watson, A. K., MacDonald, J. C., & Erickson, G. E. (2020). Evaluation of the Water Footprint of Beef Cattle Production in Nebraska. Journal of Animal Science, 98, 142. doi:10.1093/jas/skaa054.248. Green vs blue vs grey water split.

---

Back to all posts