9.4 Green computing and the sustainability pillar

What this is and why it exists

Sustainability in cloud work is mostly the same engineering as cost, pointed at a different meter — and that is good news, because it means the levers are ones you already have. This lesson is about where the energy actually goes, which single decision dominates all the others, what region choice does and does not change, and how to read a provider's sustainability report without either dismissing it or believing all of it.

The vocabulary

• PUE (power usage effectiveness) — total facility energy divided by the energy reaching the computing equipment; a perfect building would be 1.0 and everything above that is overhead, mostly cooling.
• Utilisation — the fraction of a machine's capacity actually doing work.
• Idle draw — the energy a machine uses while doing nothing, which is a large share of its maximum.
• Consolidation — running the same work on fewer, busier machines.
• Carbon intensity — how much carbon dioxide is emitted per unit of electricity on a particular grid, which varies by place and by hour.
• Carbon-aware scheduling — running interruptible work when or where the grid is cleaner.
• Embodied carbon — the emissions from manufacturing the hardware, before it runs anything.

The mental model

Follow the energy in order and the priorities sort themselves. Electricity enters a building; some of it becomes computation and the rest becomes cooling, power conversion and lighting. Power usage effectiveness measures that ratio, and it is the number providers quote because large operators have driven it close to 1. Which is exactly why it is no longer where your leverage is: the building overhead has largely been optimised by people who are not you.

Your leverage is utilisation, and the reason is a fact about hardware rather than about cloud. A server drawing power while idle consumes a substantial fraction of what it draws while busy, so a machine at five percent utilisation is not using five percent of the energy — it is using most of the energy of a busy machine to do almost nothing. Consolidation is therefore the largest single lever available to you: the same work on fewer, busier machines uses less energy, and the machine you turned off uses none at all. This is the sentence to take away, because it means the idle sweep and the right-sizing from the cost lesson are also the sustainability programme. Fewer, busier, off when unused — one action, two meters.

Region choice is the second lever and it is real but bounded. Grids differ substantially in carbon intensity depending on their generation mix, so the same workload emits differently in different places. It is a genuine improvement and it is constrained by everything else: latency to your users, data residency obligations, which services exist where, and cost. Treat it as a tie-breaker rather than an override — when two regions both serve your users well, the cleaner grid is a free improvement, and when they do not, latency and law come first.

Carbon-aware scheduling extends that to time. Grid intensity varies through the day with sun, wind and demand, so interruptible work — batch processing, model training, large data movements, report generation — can be run when the grid is cleaner rather than at the first convenient moment. It applies only to work nobody is waiting for, which is precisely the same set of work that suits spot capacity, so the two decisions are often one.

Then reading a report sceptically, which is a skill rather than a posture. Ask three questions of any such statement. Is it measured or purchased — energy actually consumed, or emissions offset by buying certificates elsewhere? Is it matched hourly or annually — an operator can buy a year's worth of renewable energy and still run on fossil generation at three in the morning, and hourly matching is a much stronger statement than an annual total. And does it include the hardware's manufacture, or only its operation? None of those questions makes a provider's report worthless; they tell you what it is a report of. And the honest closing note is about your own claims: the only number you can stand behind is one about a change you made — machines removed, average utilisation raised, a workload moved to a cleaner region — because those are measurements. "We care about sustainability" is not a measurement, and the point of this topic is being able to say the difference.

What you should now be able to explain or do

Define power usage effectiveness and say why it is not where your leverage is. Explain why a machine at five percent utilisation is not using five percent of the energy, and what follows. Name the single biggest lever you control and connect it to two things from the cost lesson. Say what region choice buys and what constrains it. Describe the work that suits carbon-aware scheduling and why it overlaps with spot capacity. Ask the three questions of a sustainability report, and state one figure about your own work that you could actually defend.

Check yourself