Personal Software and the Solarpunk Bet
Contents
- What the build-out actually costs
- …a single prompt is cheap…
- Personal software is the SolarPunk version of AI
- AI to make computing smaller instead of bigger
- Revive the hardware we already have
- Port and slim down the software
- Right-size the model and the task
- Proof that small can work
- What I can actually do as one person
- So, is this the direction we should go?
- References
In my opinion the solarpunk future is reachable with the AI and the hardware we already have, but only if we point AI at making small, durable, personal things instead of feeding an endless build-out of bigger data centers.
Back in 2024 I wrote Graceful Degradation; One Step towards Solar Punk, where I argued that systems which keep working when parts of them fail (modular laptops, redundant servers, open formats that outlive their apps, you can even take a look at the Djvu format) are a practical step toward a greener, more repairable world. I still believe the core of that post, but a lot has happened since. AI went from a toy to something I use every day, and with it came a new wave of worry about power plants, water and mountains of discarded chips.
So I want to ask a simple question: is AI pulling us away from that solarpunk future, or could it be the thing that finally makes it practical?
I tried to back every claim here with a paper, a report or a primary source, and the links are all in the references at the end so you can check me.
What the build-out actually costs
The issues are real and we should be worried, like data centers are the fastest growing electricity user on the planet, and AI is the main reason why.
The International Energy Agency’s Energy and AI report put global data center use at about 415 TWh in 2024, around 1.5% of the world’s electricity, and projected it to more than double to roughly 945 TWh by 2030, which is a bit more than all of Japan uses today. Their April 2026 follow-up, Key Questions on Energy and AI, says data center demand grew 17% in 2025 while AI-focused data centers grew about 50%, against only about 3% growth in global electricity demand overall. The same report notes that the biggest tech companies spent over 400 billion dollars in capital expenditure in 2025 and are expected to spend about 75% more in 2026.
The United States shows how fast this moves on the ground. Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report found US data centers used 58 TWh in 2014, 76 TWh in 2018 and 176 TWh in 2023, which was about 4.4% of the country’s electricity. By 2028 they expect somewhere between 325 and 580 TWh, or 6.7% to 12% of all US power.
| What | Number | Source |
|---|---|---|
| Global data center electricity, 2024 | ~415 TWh (~1.5% of world use) | IEA, Energy and AI (2025) |
| Global data center electricity, 2030 projection | ~945 TWh | IEA, Energy and AI (2025) |
| Growth in AI-focused data center electricity, 2025 | ~50% in one year | IEA, Key Questions on Energy and AI (2026) |
| US data center share of electricity, 2023 → 2028 | 4.4% → 6.7–12% | LBNL (2024) |
| Global AI water withdrawal, 2027 projection | 4.2–6.6 billion m³ | Li et al., Making AI Less Thirsty |
| Generative AI e-waste, 2020–2030 total | 1.2–5.0 million tonnes | Nature Computational Science (2024) |
| All e-waste, 2022 | 62 million tonnes, 22.3% formally recycled | Global E-waste Monitor 2024 |
Water
Water is the cost people forget. In Making AI Less “Thirsty”, Pengfei Li, Shaolei Ren and colleagues at UC Riverside estimated that training GPT-3 in Microsoft’s US data centers could directly evaporate about 700,000 liters of clean freshwater.
They project global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027, more than the total yearly withdrawal of four to six Denmarks or about half the United Kingdom. Their code is open on GitHub at Ren-Research/Making-AI-Less-Thirsty.
Then there is the hardware itself. A 2024 study in Nature Computational Science estimated that generative AI alone could produce between 1.2 and 5 million tonnes of e-waste between 2020 and 2030, with yearly AI e-waste growing from about 2,600 tonnes in 2023 to as much as 2.5 million tonnes a year by 2030 in the aggressive scenario. The same researchers found that prolonging, reusing and recycling the hardware could cut that by 16% to 86%. That lands on top of a pile that is already enormous: the UN’s Global E-waste Monitor 2024 counted a record 62 million tonnes of e-waste in 2022, with only 22.3% documented as properly collected and recycled, and less than 1% in African countries.
…a single prompt is cheap…
Something to note is that not “every prompt boils a lake”. A single prompt costs very little energy and that is exactly the reason why the total keeps climbing.
In August 2025 Google published Measuring the environmental impact of delivering AI at Google Scale, which metered the median Gemini Apps text prompt at 0.24 Wh of energy, 0.03 g of CO₂e and 0.26 mL of water (about five drops). They compare it to watching TV for less than nine seconds. They also report that over one year their software and clean energy work cut the energy per median prompt by 33 times and its carbon footprint by 44 times. It is worth saying this is the company measuring itself and it was not peer reviewed at release, but it is the most complete measurement we have from a real production system. The IEA makes a similar point at a larger scale: if every ordinary web search were done as a simple AI text query, it would use less than 4 TWh a year, under 1% of what data centers use today.
Small jobs, small machines
What a prompt costs depends a lot on what you ask for. The Power Hungry Processing study by Sasha Luccioni, Yacine Jernite and Emma Strubell measured 88 open models on 10 tasks. Classifying text cost about 0.002 kWh per 1,000 queries, while generating images cost about 2.9 kWh per 1,000, a gap of more than 1,450 times.
Text generation sat in between. So a quick text question and a batch of AI images are not the same thing at all, and that difference matters for how I use these tools.
The catch is the Jevons paradox. In the 1860s William Stanley Jevons noticed that as steam engines burned coal more efficiently, Britain ended up burning more coal, not less. Luccioni, Emma Strubell and Kate Crawford argue in From Efficiency Gains to Rebound Effects (FAccT 2025) that AI is following the same pattern: cheaper and more efficient AI invites more uses, and those extra uses can wipe out the savings. They conclude that technical efficiency alone will not bring net reductions in environmental harm, and that business incentives and policy decide the outcome as much as the chips do. The IEA’s 2025 numbers look like that warning coming true: per-prompt costs fell sharply in the same year total AI data center electricity jumped by half.
So here is what I think: efficiency is necessary but it is not a plan. The question is what we spend the efficiency on. We can spend it on more video slop and more always-on agents running in giant facilities, or we can spend it on making computing smaller, longer-lived and closer to the people who use it.
Personal software is the SolarPunk version of AI
By personal software I mean small tools you make for yourself, your family or your community, which run on hardware you already own and keep working when the internet or a company disappears.
Home-cooked software
The idea is older than the current AI wave. In 2020 Robin Sloan wrote An app can be a home-cooked meal about a tiny messaging app he built only for his family. It never needed to scale, never had a business model, and it did exactly what four people needed.
What has changed since then is who gets to cook. With an AI coding assistant, someone with a bit of programming background (or none) can now build that kind of tool in an afternoon instead of a month.
The second piece is where that software runs. In 2019 Martin Kleppmann and the researchers at Ink & Switch published Local-first software: You own your data, in spite of the cloud, which lays out seven ideals: software that is fast, works across your devices, works offline, supports collaboration, lasts a long time, keeps your data private and leaves you in control. That list reads to me like graceful degradation written for apps. If the server goes down, a local-first app degrades to “works on my machine” instead of “does not work at all”. The open source library behind a lot of that work is Automerge.
The third piece is the model. You do not need a frontier model in a giant data center for most everyday tasks. NVIDIA researchers argue in Small Language Models are the Future of Agentic AI (2025) that models under about 10 billion parameters are already capable enough for most of the narrow, repetitive jobs inside AI agents, and are 10 to 30 times cheaper to run. That is an interesting thing to hear from the company that sells the big chips. Microsoft’s BitNet work pushes the same direction from the hardware side: its bitnet.cpp paper reports 1.58-bit models running on ordinary CPUs with 2.37x to 6.17x speedups and 71.9% to 82.2% less energy on x86 chips, and a 100-billion-parameter model reaching human reading speed (5 to 7 tokens per second) on a single CPU.
Put those three pieces together and you get the picture I care about: home-cooked tools, local-first data, and small models that run on the machine in front of you. That is a computing culture that fits inside solarpunk instead of fighting it. The permacomputing community has been building toward this for years with principles like “care for the chips” and “keep it small”, and AI can make their ideas practical for far more people.
AI to make computing smaller instead of bigger
Perhaps the most useful thing an AI can do for a SolarPunk world is help us keep old machines alive and make old code lean, because most of a computer’s carbon is spent before you run it.
Revive the hardware we already have
Keep the old machines working
For a typical laptop, most of the climate cost is paid at the factory. TCO Certified’s lifecycle study puts 79.8% of a notebook’s greenhouse gas emissions in manufacturing, against 13.2% for actually using it, and found that using a notebook for six years instead of four cuts its average yearly emissions by about 29%.
Every year a working laptop stays in use is carbon that never has to be spent again.
The opposite is happening at scale right now. When Microsoft ended free Windows 10 support on 14 October 2025, the analyst firm Canalys estimated that around 240 million PCs that cannot officially run Windows 11 could become e-waste, about 480 million kilograms of it, even though many of them work perfectly well. This is where AI is genuinely useful for regular people. A coding assistant can walk you through installing a lightweight Linux system on one of those machines, fixing a driver, or writing the small script that makes an old printer work again. A few years ago that took a patient friend who knew Linux. Now it takes a chat window. Hell, you can even reverse engineer your modern software so that it works with your old notebook, it will take a whole lot of tokens but it is definitely possible.
Port and slim down the software
Code migration used to be the boring, endless work that kept old systems stuck. Google’s engineers describe in How is Google using AI for internal code migrations? (2025) how AI-assisted tooling wrote most of the code changes in several large migrations, and the developers estimated about 50% less total time than doing it by hand. Amazon reported using its Q Developer agent to move tens of thousands of internal applications from Java 8 and 11 to Java 17, which it estimated saved over 4,500 developer-years and about 260 million dollars a year, mostly from needing fewer servers once the apps ran on the faster runtime. Both are companies grading their own work, so I take the exact numbers with some salt, but the direction is clear: the cost of rewriting software has collapsed.
That matters for energy because the same work can cost very different amounts of power depending on how it is written. The well known benchmark study by Pereira and colleagues (2017) measured Python using about 75 times the energy of C on the same tasks. A later paper, It’s Not Easy Being Green, pushes back and shows most of that gap comes from how long the program runs and how the language is implemented, not the language name itself. The practical lesson is the same either way: faster, leaner code uses less energy, and AI makes it cheap to rewrite the hot path of a slow tool in something efficient, or to drop a heavy framework for a few hundred lines that do the one job you need.
Right-size the model and the task
When I use AI, I can choose the smallest model that does the job, run it locally where I can, and avoid image or video generation when text will do, given the more than 1,000-fold gap between text classification and image generation measured in Power Hungry Processing. Hugging Face’s AI Energy Score gives star ratings that compare models on the same tasks, which is a good place to check before picking one. For running models on my own machine, llama.cpp and Ollama are the common open source tools, and CodeCarbon can estimate what my own scripts consume.
Proof that small can work
Off-grid, on purpose
Low-tech Magazine shows what this looks like in practice. Since 2018 its solar-powered website has run from a small server on a balcony in Barcelona. Over its first year it used about 0.021 Wh per unique visitor across 865,000 visitors, the server drew around 1 to 2 watts, and it stayed online about 95% of the time.
It goes offline when the weather is bad for long stretches, and they accept that on purpose. That is graceful degradation as a design choice, and it is the clearest small model I know of the world I want (the one we deserve too).
What I can actually do as one person
Start at home
I cannot stop a hyperscaler from building a gigawatt campus, but I can change what my own computing looks like, and I think that is where a solarpunk culture starts.
Most of it is boring on purpose: keep what works, fix what breaks, build only what I need, and run it close to home.
- Keep my devices longer. The cheapest carbon is the laptop I do not buy. Before replacing anything, I ask an AI assistant to help me diagnose it, swap a battery or SSD, or move it to a lighter operating system. Going from four to six years of use cuts a laptop’s yearly footprint by close to a third.
- Give old machines a second job. An old laptop or PC can become a home server, a media box, a backup drive, or a machine for running a small local model. AI makes the setup steps far less scary than they used to be.
- Cook my own tools. When I need a tracker, a converter or a little dashboard, I build the small version with AI help instead of signing up for another subscription. I keep the data in plain files I own, in the spirit of local-first software and the open formats I praised in my 2024 post.
- Use the smallest model that works. Local small models for drafts, summaries and simple code; a big cloud model only when the task really needs it. I save image and video generation for when it actually matters, since generating an image costs around ten times a typical text answer or more, and video far more than that.
- Write lean software. When AI writes code for me, I ask for fewer dependencies, static pages instead of heavy apps where possible, and a rewrite of the slow parts. A static blog like mine is already a small win.
- Design for graceful degradation. Anything I build should still be useful offline, on slow connections and on older hardware. If the cloud part fails, the tool should fall back to something simpler rather than break.
- Measure, then share. Tools like CodeCarbon and the AI Energy Score make the costs visible. Writing about what I learn, and helping friends fix and reuse their own machines, spreads the habit further than any single choice I make.
None of this cancels out a data center. The solarpunk idea was never that one person saves the planet (I am not superman after all), it is that a lot of people decide to choose repairable, local, human-scale technology until it becomes the new normal and we don’t have to rely on greedy corps, and AI just made that choice much easier to make.
So, is this the direction we should go?
Yes, in my opinion, but not by default. The evidence says AI’s footprint is growing fast at the level of the grid, while the cost of a single useful task keeps shrinking. That gap is the opportunity. If the savings go into more scale for its own sake, we get the Jevons version of the future. If they go into keeping hardware alive, rewriting bloated code, running small models locally and building tools that degrade gracefully, we get something much closer to the solarpunk world I was sketching back in 2024. The capacity to do that already exists today, and that is the part that makes me hopeful.
If you are curious about solarpunk and what kind of world it might be, here is a good place to start: Solarpunk - A Solarpunk Manifesto.
References
Energy and water
- IEA (2025). Energy and AI.
- IEA (2026). Key Questions on Energy and AI.
- Shehabi et al., Lawrence Berkeley National Laboratory (2024). 2024 United States Data Center Energy Usage Report.
- Li, Yang, Islam and Ren (2023, CACM 2025). Making AI Less “Thirsty”. Code: Ren-Research/Making-AI-Less-Thirsty.
- Elsworth et al., Google (2025). Measuring the environmental impact of delivering AI at Google Scale.
- Luccioni, Jernite and Strubell (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment?
- Luccioni, Strubell and Crawford (2025). From Efficiency Gains to Rebound Effects: The Problem of Jevons’ Paradox in AI’s Polarized Environmental Debate.
E-waste and hardware
- Nature Computational Science (2024), generative AI e-waste study, as reported by MIT Technology Review.
- ITU and UNITAR (2024). Global E-waste Monitor 2024.
- Canalys (2023). The end of Windows 10 support could turn 240 million PCs into e-waste.
- TCO Certified. Circularity in practice: how to manage notebook computers responsibly.
Software, models and practice
- Nikolov et al., Google (2025). How is Google using AI for internal code migrations?
- AWS (2024). Amazon Q Developer just reached a $260 million dollar milestone.
- Pereira et al. (2017). Energy efficiency across programming languages.
- van Kempen, Kwon, Nguyen and Berger (2024). It’s Not Easy Being Green: On the Energy Efficiency of Programming Languages.
- Belcak et al., NVIDIA (2025). Small Language Models are the Future of Agentic AI.
- Wang et al., Microsoft Research (2024). 1-bit AI Infra: Fast and Lossless BitNet b1.58 Inference on CPUs.
- Kleppmann, Wiggins, van Hardenberg and McGranaghan (2019). Local-first software.
- Robin Sloan (2020). An app can be a home-cooked meal.
- Low-tech Magazine (2020). How sustainable is a solar powered website?
- Permacomputing wiki.
- My earlier post (2024). Graceful Degradation; One Step towards Solar Punk.
Tools mentioned (GitHub)
- ggml-org/llama.cpp, run language models locally on CPUs and modest GPUs.
- ollama/ollama, a simpler way to download and run local models.
- microsoft/BitNet, 1-bit model inference on CPUs.
- huggingface/AIEnergyScore, energy ratings for AI models.
- mlco2/codecarbon, estimate the emissions of your own code.
- automerge/automerge, a library for local-first apps.

