Design & sustainability · Digital UX briefing
The object of the word has changed.
Pre-AI technology asked how to extend a product’s life, and how people might become attached to a thing instead of throwing it away. We still use the same word. This page is a working diagram of what the word points at now — from a kettle, a chair, a phone, to a warehouse of chips, a river, a grid, a model that is replaced before it can age.
What this page makes visible
- Object-era durability, still in force
- A translation table from kettle to model
- Energy, water, and chips as the new materials
- Pacing as the waste mechanism
- A dual brief: keep the object, pace the system
- Data-center electricity, 2025
- 485 TWh
- World electricity used by AI
- ~0.5%
- Data centers in 2030, base case
- 950 TWh
- U.S. electricity by 2030
- 11.8%
IEA · ~1.5% of global electricity
Our World in Data, from IEA
IEA · ~3% of global electricity
Berkeley Lab · range 9.5–15.3%
The black line is the word “sustainability.” It still travels. What it is attached to has moved — from a thing you could hold to a warehouse of chips, a river, a grid.
01 — The object
In pre-AI technology, sustainability was a thing you could hold.
The brief was tactile. How do we extend the product lifecycle? How do people become attached to an object and not discard it? Durability was discussed as metal, joinery, a surface that could age. Attachment was discussed as memory, ritual, a story the object could carry. Toggle the mechanism and compare the two briefs.
The crisis of unsustainability is a crisis of behaviour, not one simply of energy and materials alone.
Waste streams filled with electronics that still performed their tasks. The defect was not mechanical. It was relational. The object had outlived the feeling that justified keeping it. Chapman’s warning was precise: pairing excessive material durability with fleeting product meaning is a formula for waste.
Question one
Why do we throw away products that still work?
Question two
How do we design products people want to keep longer?
The pre-AI toolkit
Around 80% of a product’s environmental impact is determined at design. That was the right place to have the argument.
Longevity
Form, material, and joinery meant to outlast a season of desire.
Repair
Screws not glue. Parts not black boxes. The right to open the thing.
Patina
Aging as value. A surface that records use instead of looking used-up.
Circularity
Reuse, remanufacture, then recycle. The grave is a last resort.
Attachment
Ritual, personalization, care. Meaning as a retention strategy.
Sufficiency
Not only a greener kettle — fewer kettles, kept longer.
Pre-AI technology
A thing you could hold
The discarder is the user. Durability is metal, joinery, a surface that can age. Attachment is memory, ritual, a story the object can carry.
- 01The product is visible. You can watch it age.
- 02The user is the one who discards.
- 03Attachment is personal, local, and slow.
- 04A lifecycle is cradle-to-grave of one artifact.
- 05Durability is a virtue. Faster replacement is a vice.
- 06The designer’s power is over form, material, and meaning.
AI era
A competence in a text box
The discarder is a capital cycle. Durability is chip life, water, and whether a model is kept. Attachment is to a voice you cannot repair.
- 01The product is a competence in a text box. The warehouse is offstage.
- 02A capital cycle discards. The user only queries.
- 03Attachment is to a voice, a habit, a capability — reset each version.
- 04A lifecycle is mine, fab, train, infer, discard — then the grid.
- 05Speed is the product. Durability can be a competitive defect.
- 06The designer’s unit of work is a system: model, cluster, contract, grid.
Toggle the mechanism. The left column is what pre-AI technology assumed. The right is what those premises become when the product leaves the room.
02 — The vanishing
Then the object left the room.
AI is not a kettle. You do not hold the model. You do not repair the GPU. You do not decide when a chip is “done.” The product is a competence that arrives in a text box. The infrastructure is a building you will never enter, drawing a river you will never see.
Artificial intelligence does not arrive as a glowing brain. It is mines, water, classification, labor, and power — an extractive industry with a user interface.
The sustainability conversation did not end. It lost the object that made it teachable. That is why the narrative feels both continuous and unrecognizable.
02 — Same word, new object
A translation table.
Press a station. The pair stays on the stage — object-era word on the left, what it points at now on the right. Play walks the word across.
01 / 08
same
Pre-AI
Durability of the thing
AI era
Durability of chips, weights, and relationships
The virtue survived. The unit of keeping did not.
03 — New materials
The footprint is still small. The slope is not.
Energy is the new material. Location is a design decision with a 24× carbon range: under 30 gCO₂/kWh in Norway and Sweden, over 600 in parts of Southeast Asia. Allianz estimates the true 2025 data-center footprint at 286 MtCO₂ — 57% above electricity-only accounts, once hardware and construction are included. Press a year.
485 TWh
Data-center electricity
485 TWh
~1.5% of global electricity
What AI is doing
AI ~0.5% of world electricity
Reading
IEA, Energy and AI. AI-focused sites grew ~50% this year.
Green AI, Red AI, Jevons
Efficiency is not the same as sufficiency.
Red AI
Buying accuracy with compute.
Schwartz, Dodge, Smith, and Etzioni named the trend in 2020: linear gains purchased with exponential cost.
Green AI
Treat cost as a first-class metric.
Report the work — floating-point operations, energy, water — beside the score.
Energy per AI task is falling fast. Usage is not. That is the kettle problem in a new grammar: a more efficient object, used without limit, is not a smaller life.
Water
The river was never in the brief.
814 bn L
Data-center water, 2025
Allianz · ~¾ indirect, via power plants
222 bn L
Direct cooling water, 2025
Rystad · could triple by 2030
+34%
Google water use, 2025
10.9 billion gallons · AI buildout
In pre-AI technology, you did not throw the kettle in the river. AI does — invisibly, per query, through cooling towers and through the thermal plants that feed them.
Hardware that dies twice
Functioning GPUs, discarded for pace.
First death · physical
5–7 years
A well-kept data-center GPU still computes. The silicon is often not finished.
Second death · economic
2–4 years
Hopper to Blackwell is the new fashion season. Then chips cascade — or they do not.
Chapman’s diagnosis transfers with almost no editing: waste is born from pairing leftover physical life with a meaning that has already moved on. De Vries-Gao puts 2030 AI-server e-waste in the range of 131–225 kilotons a year — machines that mostly still work.
04 — The race
We do not keep weights the way we hoped to keep chairs.
New state of the art arrives on a quarterly clock. Distillation exists. The cultural default is still to train bigger. Attachment is almost un-designed: no patina, no repair, no narrative of aging — only a version number.
How big is too big? The environmental bill is paid first by people who will not receive the product.
Pacing
The race is the waste mechanism.
$400B+
Five tech firms’ capex, 2025
IEA · another ~75% jump expected in 2026
2–5 yr
To build a data center
Grid expansion often takes 10+ years
45 GW
SMR offtake pipeline
From 25 GW at end-2024
In product design, planned obsolescence was a tactic. In AI it is a coordination failure that looks like strategy. Pacing is not a side topic in AI sustainability. Pacing is the topic.
Who carries the heat
Ethics as sustainability.
Energy, water, and minerals are not distributed like logins. Neither is the power to train a frontier model.
Misaligned bills
Climate and water costs land first on people outside the product’s market.
Concentrated power
Only the deepest pockets set the pace, the stack, and the default.
Extracted labor
Mines, fabs, clickwork, and content — earth, work, data.
Quality as ethics
Bias, hallucination, and churn are also waste: more retraining, more compute.
Attachment in the cloud
We kept the habit. We did not keep the thing.
No patina
Interfaces reset. Nothing records the years you spent with them.
No repair
You cannot open, mend, or inherit a closed model.
No aging
Replacement is the product cycle. Continuity is a changelog.
If pre-AI technology asked how people become more attached to a product, the honest AI-era answer is: they become attached to a service they cannot open, repair, or keep — and they still throw the last version away. A 2026 study found consumers rate identical products as less sustainable when told they were AI-designed. The mechanism is “genuine care.”
05 — What stayed, what changed
Keep the old work. Add a second brief.
The deep structure did not move. The surface of the problem inverted. The pre-AI question, restated: how do we extend the lifecycle of a system we cannot hold?
What stayed
01
Lifecycle thinking
Cradle-to-grave became cradle-to-cloud: mine, fab, train, infer, discard.
02
Design still locks the damage
Architecture, siting, cooling, and “how big” are the new form-giving.
03
Failed relationships are still waste
Working GPUs, abandoned fine-tunes, unread embeddings, unused heat.
04
Who pays vs who benefits
The moral core of the pre-AI critique, now at the scale of a grid.
05
Efficiency without sufficiency fails
Jevons is the efficient kettle, rewritten for tokens.
06
Novelty still beats care
We prefer the new thing. Attachment remains the unsolved design problem.
What changed
- 01The object left the room. Sustainability became infrastructure.
- 02The discarder is no longer the user. It is a capital cycle.
- 03Speed is the product. Durability can be a competitive defect.
- 04Impacts concentrate and hide: a grid, a basin, a mine, a county.
- 05Scope 3 is large and undercounted — buildings and boards, not just FLOPs.
- 06Ethics, energy, and pacing collapsed into one problem.
- 07The designer’s unit of work is a system: model, cluster, contract, grid.
For objects
Make things people can love, repair, and keep.
- Design for longevity and patina.
- Open the product. Spare the parts.
- Prefer sufficiency to a greener disposable.
- Let narrative accumulate. Do not reset the object every season.
For AI
Pace the system as if it had to be inherited.
- Treat efficiency as a primary metric, then cap use.
- Site compute on clean, water-sensible grids.
- Extend chip life; cascade before shredding.
- Keep models: distill, document, maintain — do not only scale.
- Count the whole life. Align who benefits with who pays.
Coda
What we keep is the measure of what we value.
In pre-AI technology, that was a well-made object. In the AI era, it is a well-paced system: energy we can justify, water we can share, hardware we can reuse, models we can maintain, and benefits that do not leave the bill with someone else.
The word did not change. The object did. The work is to notice both.
Sources
Figures in this briefing.
Numbers move quickly. These are the sources used, dated where the dispute matters. Treat 2030 figures as scenarios, not destiny. Not a client case file.
Object-era design
- Jonathan Chapman, Emotionally Durable Design (2005; 2nd ed. 2015); “Design for (Emotional) Durability,” Design Issues 25:4 (2009).
- European Commission, often cited: ~80% of a product’s environmental impact is determined at the design stage.
Energy, water, carbon
- IEA, Energy and AI (2025) and Key Questions on Energy and AI (2026): ~485 TWh in 2025; ~950 TWh by 2030.
- Our World in Data (2026), from IEA: AI ~0.5% of world electricity in 2025.
- Lawrence Berkeley National Laboratory, data-center energy update (2026): 11.8% of U.S. electricity by 2030 (9.5–15.3%).
- Gartner (June 2026): 565 TWh in 2026, +26%; AI-optimized servers 31% of data-center power.
- Allianz, “Code, carbon, kilowatts” (June 2026): 286 MtCO₂; 814 billion liters water.
- Rystad Energy (2026): 222 billion liters direct cooling water in 2025; could triple by 2030.
- UNECE (Sept 2026) on grid-versus-data-center build times.
Hardware, models, ethics
- Epoch AI on frontier-chip lifespan (~3.9 years) and AI energy.
- Alex de Vries-Gao, Resources, Conservation and Recycling (2026), AI e-waste 131–225 kt/year by 2030.
- Schwartz, Dodge, Smith, Etzioni, “Green AI,” CACM (2020).
- Bender, Gebru, McMillan-Major, Mitchell, “On the Dangers of Stochastic Parrots,” FAccT (2021).
- Kate Crawford, Atlas of AI (Yale, 2021).
- Strubell, Ganesh, McCallum (2019) on NLP training energy.
- Duffek et al. / Georgia State (2026): AI-designed products rated less sustainable via a “genuine care” gap.
September 2026