F201 TA5: Environment and Society
What all this computing costs the planet, what green computing can actually do about it, and how big data is changing work and society.
The environmental cost
Where the energy goes
| Source | Detail |
|---|---|
| Servers | Continuous power, whether busy or idle. An idle server still draws a large fraction of its peak |
| Cooling | Historically as much energy as the computing itself; modern designs have cut this sharply |
| Storage | Keeping data costs power continuously, for years, whether or not it is read |
| Networking | Moving data between regions and to users |
| Model training | Weeks of thousands of GPUs, repeated every time a model is retrained |
| Inference | Each individual query is small, and multiplied by billions it exceeds the training cost |
Power usage effectiveness
PUE is total facility power divided by the power delivered to the computing equipment. A PUE of 2.0 means half the electricity is spent on cooling and overheads; the best modern facilities run close to 1.1. It is the standard measure, and it is examinable.
Water and materials
Evaporative cooling consumes fresh water, which matters where data centres sit in water-stressed regions. Hardware manufacture needs mined metals including rare earths, with their own extraction impact, and a three to five year replacement cycle produces e-waste containing both valuable and hazardous materials. Only a minority of global e-waste is formally recycled.
Carbon, and why location matters
The same workload produces very different emissions depending on the grid it runs on. A job run where the grid is coal-heavy can emit several times what it would on a grid dominated by wind, hydro or nuclear, which is why the location and the timing of a workload are now genuine environmental decisions rather than purely technical ones.
A data lake nobody queries still draws power every second of every year. Deleting data that serves no purpose is both a data protection duty under storage limitation and the simplest environmental saving available.
Green computing
What actually reduces the impact
| Measure | How it helps |
|---|---|
| Virtualisation and consolidation | Many workloads on fewer physical machines, so fewer idle servers |
| Elastic scaling | Capacity released when the job finishes rather than running all night |
| Efficient hardware | Newer processors and accelerators do far more work per watt |
| Renewable supply | Buying or generating clean power, ideally matched hour by hour rather than annually |
| Carbon aware scheduling | Running flexible batch jobs when and where the grid is cleanest |
| Better cooling | Free air cooling, liquid cooling, warm-water loops, cold climates |
| Heat reuse | Waste heat piped to district heating or swimming pools |
| Efficient code and queries | A query scanning one column instead of a whole table costs a fraction of the energy |
| Model efficiency | Smaller models, pruning, quantisation, reusing a pretrained model rather than training from scratch |
| Data retention | Deleting what has no purpose, and archiving cold data to lower-power storage |
| Hardware lifecycle | Longer refresh cycles, repair, resale, certified recycling |
The analyst's own contribution
- Do not re-run the query every five minutes if the data updates daily
- Filter and aggregate at the source rather than pulling everything and discarding most of it
- Partition tables so a query touches the partitions it needs
- Cache results that many people request
- Delete intermediate datasets when the pipeline is finished
- Reuse a pretrained model rather than training one from nothing
Rebound effects, honestly
Efficiency does not automatically reduce total consumption, because cheaper computing gets used more. Data centre efficiency has improved dramatically while total data centre energy use has continued to rise, since the work has grown faster than the efficiency. An honest answer acknowledges this rather than presenting efficiency as a solution on its own.
Society, work and culture
Employment
| Effect | Detail |
|---|---|
| Automation of routine work | Data entry, basic reporting, first-line support, routine document review |
| New roles | Data engineers, analysts, scientists, ML engineers, governance and AI ethics roles |
| Changed roles | Most jobs gain a data component rather than disappearing: clinicians, farmers, marketers, mechanics |
| Polarisation | Growth at both ends of the skill range and hollowing in the middle |
| Skills gap | Persistent shortage of people able to do the new work, and a retraining problem for those displaced |
| Monitoring at work | Productivity measured continuously, which changes the experience of work itself |
Society and culture
- Personalisation and filter bubbles: recommendation optimised for engagement narrows what people see, and the effect is strongest on contested topics
- Misinformation: synthetic text, images and video are cheap to produce and hard to detect, which shifts the burden to verification
- The digital divide: benefits accrue to people and regions with connectivity, devices and skills; those without fall further behind, and public services moving online deepens it
- Public health and research: population-scale data has produced genuine advances in outbreak detection, screening and treatment
- Smart cities: transport, energy and waste optimised against real demand, alongside the surveillance that comes with it
- Trust: a visible breach or a visibly unfair algorithm damages trust across the whole sector, not just the organisation responsible
- Attention and wellbeing: systems optimised to maximise time spent have measurable effects on sleep and mental health, particularly for younger users
A balanced position
Big data is not a force with a single direction. The same capability that detects an outbreak early also enables surveillance; the same recommendation engine that helps somebody find work also narrows what they see. The examinable skill is weighing a specific application rather than judging the technology as a whole, and strong answers do exactly that: name the benefit, name the cost, say who carries each, and conclude with what would have to be true for it to be worth doing.
It quantifies something, names a measure such as PUE or a retention period, acknowledges a rebound effect or a trade-off, and reaches a clear conclusion that somebody could act on. Lists of advantages and disadvantages with no judgement stay in the middle bands.
Key terms you must be able to define
- PUE
- Power usage effectiveness: total facility power divided by power used by the computing equipment.
- Green computing
- Designing and operating computing to reduce its environmental impact.
- Carbon aware scheduling
- Running flexible workloads when or where the electricity grid is cleanest.
- E-waste
- Discarded electronic equipment, containing both valuable and hazardous materials.
- Virtualisation
- Running several logical machines on one physical machine to reduce idle hardware.
- Heat reuse
- Capturing waste heat from a data centre for another purpose such as district heating.
- Rebound effect
- The tendency for efficiency gains to be offset by increased use.
- Digital divide
- The gap between those with access to technology, connectivity and skills and those without.
- Filter bubble
- The narrowing of what a person sees caused by personalisation optimised for engagement.
- Model inference
- Using a trained model to make a prediction, whose cumulative energy cost can exceed training.
Flashcards
Say the definition out loud before you flip the card. Retrieving it from memory is what builds recall, reading it again is not.
Reach the top grade on this topic
Quantify, name a measure, and decide. "Their PUE is 1.6, so 60% overhead on top of the computing load; moving the flexible nightly jobs to a region with a cleaner grid and raising the inlet temperature would cut both, though the saving will be partly offset as the freed capacity gets used."
Mistakes that cost marks here
- Treating the environmental cost as only the electricity for the servers, and ignoring cooling, storage, water and manufacture.
- Assuming model training dominates. At scale, inference across billions of queries usually exceeds it.
- Presenting efficiency as a complete solution without acknowledging the rebound effect.
- Saying 'data is stored in the cloud so there is no environmental impact'.
- Writing only about job losses. Most roles change rather than disappear.
- Giving a list of advantages and disadvantages with no judgement at the end.
Training one large AI model can use as much electricity as a few hundred UK homes use in a year, and the data centre housing it may evaporate millions of litres of water doing it. Both figures are now reported, which is itself a recent change.
Knowledge check
Exam-style questions
Five written questions in the style of the real paper. Write a full answer, then let the marker check it against the mark scheme.
- LOne valid measure
- LA second, different valid measure
Using free air cooling rather than refrigeration where the climate allows it, and virtualising workloads so that fewer physical servers run, reducing the number sitting idle.
- LEmissions depend on how the local electricity is generated
- LThe same workload can emit several times as much on a dirtier grid
- LClimate affects the cooling energy required
- LA consequence, such as carbon aware placement or scheduling
The emissions from a computing workload come almost entirely from the electricity it consumes, and the carbon intensity of electricity varies enormously by grid: a kilowatt hour on a coal-dominated grid can carry several times the emissions of the same kilowatt hour on a grid dominated by wind, hydro or nuclear. The identical job therefore has a very different footprint depending only on where it runs. Location also changes how much electricity is needed in the first place, because a facility in a cool climate can use outside air for much of the year while one in a hot climate runs refrigeration continuously, and cooling has historically been a large share of total consumption. The practical consequence is that placement and timing become environmental decisions: flexible batch work and model training can be scheduled into regions and hours when the grid is cleanest, without changing the work itself.
- LEnergy use of servers and cooling is addressed
- LA measure such as PUE is named and used
- LStorage's continuous cost is addressed
- LTraining and inference are distinguished
- LWater, materials or e-waste are addressed
- LRealistic mitigations are proposed
- LThe rebound effect or a limitation is acknowledged
- LA clear evaluative judgement
The impact is larger and broader than the electricity bill suggests. Servers draw power continuously and an idle machine still consumes a substantial share of its peak, while cooling adds an overhead measured by power usage effectiveness: a facility at a PUE of 1.8 spends eighty per cent as much again on cooling and losses as it does on computing, where the best modern designs run near 1.1. Storage is the cost organisations most often overlook, because data kept for years draws power every second whether or not a single query ever reads it, which makes a neglected data lake a continuing environmental liability as well as a compliance one. In machine learning the headline figure is training, which can occupy thousands of accelerators for weeks and is repeated at every retrain, but for any widely used model the cumulative energy of inference across billions of individual queries exceeds it. Beyond energy there is water, since evaporative cooling consumes fresh water in regions that may be water stressed, and materials, since a three to five year hardware cycle depends on mined metals and produces a continuous stream of e-waste of which only a minority is formally recycled. What organisations can realistically do divides into the effective and the cosmetic. The genuinely effective measures are consolidating workloads through virtualisation so fewer machines idle, using elastic capacity that is released when a job finishes rather than running overnight, scheduling flexible batch work and training into regions and hours when the grid is cleanest, enforcing retention policies so unused data is deleted and cold data moved to lower-power storage, writing queries that scan partitions and columns rather than whole tables, and reusing or fine-tuning a pretrained model instead of training from scratch. Procuring renewable supply matters, though matching it hour by hour is far stronger than an annual accounting claim, and extending hardware life with repair and certified recycling addresses the materials side that energy measures ignore. The limitation has to be stated honestly: efficiency alone has not reduced total consumption, because data centre efficiency has improved markedly while total data centre energy use has continued to rise, as cheaper computing simply gets used more. Judged overall, the measures with the largest certain effect are the unglamorous ones, deleting data that serves no purpose, not running what does not need to run, and sizing models and queries to the question, because they reduce the work itself rather than making the same growing workload marginally cheaper to perform.