A New Work New Culture response to the current AI energy alarm: resource consumption is real. But fear is not an operating system.
Excerpt from the permanent publication Flowbook AI by NWNC: an ongoing trail on AI, New Work, freedom substrates, and public capacity to act. To Flowbook AI.
Thesis
AI’s hunger for energy is not background noise. It is infrastructure policy. But public debate often behaves as if there were only two dignified roles: alarmist or priest of growth. Between them lies a third path: building AI smaller, nearer, more local, and more useful. Not as a romantic tinkering solution, but as mature technical practice.
The real question, then, is not: Is AI good or evil?
The better question is: Why do we accept an AI future in which almost every small mental routine must first pass through a global data center?
That is roughly like sending a freight company to the bakery around the corner because one wants to “strategically scale” the path to the front door.
1. The alarm is right. The shortcut is not always.
The starting point is serious. The current AI boom draws electricity, water, chips, grids, capital, and political attention. Data centers are not clouds. They are locations. Lines. Cooling. Land. Permits. Transformers. Gas contracts. Waste heat. And somewhere there is a mayor who actually wanted to renovate a school.
In its report “Energy and AI”, the IEA expects data centers to more than double their electricity consumption from around 415 TWh in 2024 to about 945 TWh in 2030. That is not a small matter. At the same time, the same IEA says: in 2024, data centers accounted for about 1.5 percent of global electricity consumption. By 2030 they account for roughly one tenth of global demand growth, less than industrial motors, air conditioning, or electric vehicles.
Both are true.
And that is precisely where adult politics begins.
Not: “AI is killing the planet.”
Not: “It is all just panic, carry on.”
Rather: the global share is still limited, but the local force is enormous. Especially in regions where new data centers, semiconductor factories, and grid expansion arrive at the same time. Those who live there do not experience “1.5 percent of global electricity”. They experience connection queues, new power-plant debates, water stress, land conflicts, and an industry that wants very much very quickly.
The sentence “killing the planet” is understandable as a moral alarm. As analysis, it is too crude. It turns a concrete infrastructure conflict into an end-times backdrop. End-times backdrops are practical when one needs attention. They are less practical when one wants to build something.
The design question is: How do we talk about AI energy in a way that creates capacity to act, not just another moral smoke detector?
2. Nvidia is a symptom, not the whole body.
The LinkedIn post points to Nvidia’s sharply rising emissions. That is justified, but the number needs clean edges.
In the NVIDIA Sustainability Report Fiscal Year 2026, Scope 3 emissions are 10,700,940 t CO₂e. In the previous year they were 6,912,577 t CO₂e. That is an increase of around 55 percent. Anyone speaking of “53 percent in 2025” apparently means the jump from reporting year FY25 to FY26, not cleanly a calendar year 2025. Since FY24, Scope 3 emissions based on this table are about 2.94 times as high. So almost tripled, not cleanly “more than tripled” if these reporting years are the basis.
This is not an all-clear. On the contrary.
It is the more precise form of the problem.
Nvidia is fabless. Production sits in supply chains, especially with highly specialized semiconductor manufacturers such as TSMC and memory manufacturers such as SK hynix or Samsung. In the West, the AI boom often appears as a software event. In East Asia it is also a question of fabs, water, chemicals, electricity mix, grid stability, and industrial policy.
That is the blind spot of the beautiful cloud story: when intelligence is sold as a service, the material weight disappears from view. One sees the chat window. One does not see the metabolism behind it.
And that metabolism is considerable. TSMC reports rising requirements for energy, water, and emissions management in its sustainability and annual reports. The pointed warning from Greenpeace East Asia / The Diplomat about particularly energy- and water-intensive new fabs is plausible as a warning signal, but should not be treated as a clean substitute for primary data. What is officially well documented is this: semiconductor production is not an abstract background process. It is its own energy and resource complex.
This should change the European AI debate. Those who speak only about data centers see the server room. Those who speak about chips see the factory. Those who speak about freedom must see both.
The design question is: Which AI infrastructure may grow if its supply chain permanently concentrates ecological and political burdens in other regions?
3. Green Horror is a bad operating manual.
Green Horror has a use. It interrupts the story of frictionless progress. It says: look closely. There is electricity. There is water. There is extraction. There are regions paying for our convenience.
But horror is a bad architect.
It usually knows only three tools: shock, prohibition, guilt.
Shock wakes people up. Prohibition can be necessary. Guilt can establish a relationship to reality. But this still does not create a mode of operation. One cannot run a local AI lab with a bad conscience. Nor can one maintain a municipal data space with outrage. At some point someone needs a model, a device, a socket, a responsibility, training, a procurement decision, and a person who says on Tuesday morning: “Right, now we test this.”
That is not sexy. It is useful.
And precisely for that reason it appears less often in the media.
Prohibition jargon is easier than tool knowledge. Doom jargon is easier than maintenance. A data center with gigantic electricity demand is a clear image. A small local model server in a city library is an appointment with coffee, cables, and three people who want to know whether this also works without the cloud. That is not dramatic. But culture rarely emerges from drama. It emerges from repeatable practice.
There is a peculiar hostility to pleasure in the digital sustainability debate. As if a solution may only be taken seriously once it sounds unpleasant. As if every “We can try this today” were suspicious. Sometimes it feels as if public debate has an allergy to workshops.
And yet pleasure would be appropriate precisely here.
Not consumer pleasure. Not tech-bros-do-another-demo pleasure.
But the pleasure of building things smaller, smarter, and nearer.
The design question is: How does ecological critique become a workshop again, rather than only a warning tone?
4. The missing third sentence: much more can be local than we are told.
Most people experience AI today as a remote control for a data center. One types. Somewhere else something hums. Then text comes back. That feels modern, but it is not the only technical form of AI.
Many tasks do not need a frontier model in a hyperscale data center.
Many tasks need:
- a local summary,
- pre-structuring,
- a classification,
- a draft translation,
- a search in one’s own documents,
- assistance with writing,
- a code explanation,
- administrative help,
- a small RAG application,
- a model close enough not to carry data out of the house.
Mature building blocks for this have long existed. Ollama runs on Mac, Windows, and Linux and makes open models accessible to ordinary users and developers. llama.cpp enables local inference with GGUF models, quantization, and support for a broad hardware base. MLX is an Apple Silicon path for efficient machine learning on devices that already sit on many desks. Google explicitly describes Gemma 3n as a model family for everyday devices such as phones, laptops, and tablets.
That does not mean every laptop replaces a data center.
Of course not.
A small local model does not train a frontier model. It does not solve a global chip supply chain. It does not magically make electricity green. Nor is it automatically democratic just because it sits on a desk. Local hardware also has manufacturing costs. Local inference also needs electricity. Open models also need review, governance, and competence.
But the counter-model is real.
And it is too rarely part of the story.
We could sort AI architecture by proximity:
What really has to be in the cloud?
What can run on one’s own computer?
What belongs on a local server in the organization?
What belongs in public, auditable infrastructure?
What should not be automated at all, because human slowness here is not an error?
These questions are unspectacular. They fit poorly into a culture war. But they are exactly the kind of questions from which a new digital maturity could emerge.
The design question is: Why is “local first, cloud when necessary” not already a standard question in AI strategies?
5. New Work: freedom needs substrates, not just attitude.
From a New Work New Culture perspective, the energy question is not only a climate question. It is a freedom question.
Because infrastructure decides who can act.
If AI exists only as a subscription in a few global platforms, then freedom is always borrowed. The surface is friendly. The dependency lies underneath. Prices can rise. Access can disappear. Data protection can become a contractual question. Models can be politically, culturally, or economically filtered without local users really understanding where the boundary lies.
Local and edge models are therefore not just technical toys. They can be freedom substrates.
A freedom substrate is not a big word for a small tool. It means the condition under which people can truly act: knowledge, spaces, energy, devices, open models, security, learning paths, local responsibility.
A municipal AI Maker Space would be such a substrate.
A school that runs a small local language model for writing training, research critique, and data-protection education.
An administration that does not reflexively pull internal documents into an external cloud, but first tests local assistance systems.
A city library that shows citizens how to use a small model on their own computer.
A local energy cooperative that does not describe AI as a power hog, but uses it as a tool for load management, building efficiency, and grid-data competence.
An editorial office that asks not only whether AI must be prohibited, but which AI practice makes people more independent.
That sounds smaller than “AGI”. Good.
Most useful things in life sound smaller than AGI.
No one stands in front of the kettle in the morning and says: “I now expect a transformative hot-beverage platform.” One wants tea. It should work. And, if possible, not personally insult half the power plant.
New Work does not ask here: How do we make AI more pleasant?
New Work asks: What work do people really, really want to do when machines can take over the machine-like? And what infrastructure do they need so that this does not merely create new dependency?
The design question is: Which local AI capabilities belong to the basic education of a free digital culture?
6. Why media write so rarely about these solutions.
It would be too cheap to say: media only want doom.
That is not true.
But the media system rewards conflict, clear opponents, and big numbers. Local solutions have three disadvantages:
First: they are distributed. There is not one local AI model that saves the world. There are many small deployments, many configurations, many limits of use. That is journalistically less convenient than one corporation, one number, one scandal.
Second: they require competence. Anyone writing about local models has to talk about memory, quantization, data protection, model quality, latency, energy profiles, maintenance, and use cases. That is work. A doom sentence is written faster. A prohibition sentence too.
Third: they make the reader responsible. Green Horror allows a clear emotional role: I am against the bad thing. Local practice asks more uncomfortably: What runs on your computer? What does your organization procure? Which data do you send where? Who learns this? Who maintains it? Who decides when the cloud is really necessary?
That is less comfortable. But it is more adult.
Perhaps what the debate lacks is therefore not only optimism. Perhaps it lacks craft.
The design question is: Which editorial offices, universities, administrations, and civil-society places are building a language for AI workshop practice, rather than only for AI warning?
7. A green AI policy that would have desire
A pleasurable green AI policy would not be naive. It would not gloss over supply chains. It would not release data centers from responsibility. It would regulate water, electricity mix, grid connections, waste heat, location decisions, and semiconductor production hard.
But it would also build.
It would promote local AI competence:
- local model workshops in libraries, schools, adult education centers, and makerspaces,
- public procurement with “local first, justify cloud”,
- small public compute nodes for education, administration, and civil society,
- funding programs for low-energy, auditable specialist models,
- open benchmarks for quality, electricity consumption, and data protection,
- data-center rules that take load flexibility, waste-heat use, and transparent energy sources seriously,
- municipal experiments with AI for buildings, grids, mobility, and administration,
- a European public-AI practice that does not only regulate platforms, but builds its own capabilities.
That would not be anti-AI policy.
It would be adult AI policy.
It would say: yes, Green Horror shows a real problem. But fear is not an operating system. If we only warn, the future remains with those who build. If we only prohibit, practice remains with those who bypass. If we only scale, freedom remains with those who own the infrastructure.
So we build differently.
Smaller, where smaller is enough.
Local, where local makes sense.
Open, where public control is necessary.
Cloud, where cloud is truly needed.
And slow, where human judgment is more important than the next automatic reflex.
That might be the real green point: not less intelligence. Less waste of intelligence.
The design question is: Which AI use would we build differently tomorrow if electricity, data, and freedom were not invisible?
Conclusion
The AI energy boom is real. It deserves critique. It deserves regulation. It deserves hard questions for Nvidia, TSMC, hyperscalers, data-center operators, politics, and investors.
But if the debate stops there, it becomes strangely passive.
Then AI is either monster or messiah.
And we sit beside it like people in a very long panel discussion where no one can find the plug.
New Work New Culture should ask differently:
What can people understand, operate, and shape locally?
Which AI helps them do work they really, really want?
Which infrastructure makes them more sovereign rather than more dependent?
Which models are small enough to be near, and good enough to be useful?
Green Horror shows what is going wrong.
The local AI workshop shows what could begin.
Between the two lies the politics of the next years.
And perhaps also a little desire.
Not the desire for the larger data center.
The desire for a technology that finally fits on the table again.
Fact Check on the Starting Post
- Nvidia emissions: the often-cited increase of around 53 percent is, based on the NVIDIA Sustainability Report FY2026, rather around 55 percent for Scope 3 from FY25 to FY26. It is not cleanly formulated as calendar year “2025 alone”. From FY24 to FY26 the factor is about 2.94, so almost tripled.
- “AI kills the planet”: understandable as a warning call, too crude as a factual sentence. The IEA sees data centers as strongly growing, locally very relevant electricity demand, but in 2024 at around 1.5 percent of global electricity consumption and by 2030 at about one tenth of global demand growth.
- TSMC / fabs: the supply-chain critique is plausible and important. Individual figures such as “1 GW” and “100,000 tons of water per day” for the most advanced fabs should be marked in the article as figures from Greenpeace/media research, not as directly confirmed here from a primary source I have read out.
- Solutions: the IEA itself names considerable solution potential through AI in grids, buildings, industry, and energy efficiency. At the same time it warns of rebound effects and says explicitly: AI is not a panacea and does not replace active policy.
- Local models: Ollama, llama.cpp/GGUF, MLX, and Gemma 3n show that local and edge AI are not a distant utopia. They do not replace frontier training or all cloud applications, but they are realistic for many everyday and organizational tasks.
Sources
- Alexandra Geese / Alistair Alexander LinkedIn starting point, seen in the feed on 2026-07-05: https://www.linkedin.com/feed/update/urn:li:share:7478805501108219904/
- NVIDIA Sustainability Report Fiscal Year 2026: https://images.nvidia.com/aem-dam/Solutions/documents/NVIDIA-Sustainability-Report-Fiscal-Year-2026.pdf
- IEA, Energy and AI, Executive Summary: https://www.iea.org/reports/energy-and-ai/executive-summary
- TSMC Annual Reports / Sustainability reporting index: https://investor.tsmc.com/english/annual-reports
- The Diplomat / Greenpeace East Asia debate piece on Nvidia, TSMC, and East Asian supply chains: https://thediplomat.com/2026/07/nvidias-silent-ai-colonialism-is-trapping-east-asia-in-a-fossil-fueled-hell/
- Ollama README / open models on Mac, Windows, Linux: https://github.com/ollama/ollama
- llama.cpp README / local inference, GGUF, quantization: https://github.com/ggml-org/llama.cpp
- Apple MLX README / machine learning on Apple Silicon: https://github.com/ml-explore/mlx
- Google AI for Developers, Gemma 3n model overview: https://ai.google.dev/gemma/docs/gemma-3n
