There is a photovoltaic installation on the roof of the building our office sits in. Since this year, the electricity for laptops, monitors, lights and the coffee machine comes from up there. It is a good change, and a modest one. It also puts a less comfortable question in front of us: where the electricity comes from is the easy half. The interesting half is how much of it a piece of software needs in the first place.
The easy half: where the electricity comes from
The building runs on self-consumption: what the roof produces powers the building itself instead of going to the grid. The electricity travels a few metres instead of a few kilometres, and it gets used where it is made.
On a sunny day the roof carries the whole building. In the evening it does not, at five in the afternoon in December it does not, and through a week of high fog it does not either. Across a year it is a mix, and that is simply how photovoltaics work north of the Alps. We are on it, and we rather like that.
The expensive half runs in a data centre
The thing is, a software company does not spend its electricity in the office. The monitors run eight hours a day. The systems we build run around the clock, for years, on machines somebody else operates. There are orders of magnitude between those two numbers.
And the second number is not settled at runtime. It is settled at design time. How large the instances are, how many replicas ride along, whether a request calls a model or reads a table: those are architecture decisions. They get made once and take effect every day after that.
Optimisation means reducing the workload of your cloud infrastructure
Four levers show up in almost every project:
- Size for the real load. A cluster provisioned for a peak that never arrives draws power in every single hour of the year. Right-sizing is the largest and most boring lever there is.
- Pick the method that fits. An LLM call for a job that OCR, a rule, or a classical model handles costs many times the compute. Our own code analysis moved back off a model and onto a linter and a rule set, with the model kept only for the cases that earn it. And where a model is still the right tool, there is still a right size of model. Generative AI is one shape among many, and usually not the frugal one.
- Switch off what nobody uses. Staging environments overnight, replicas with no readers, jobs whose output nobody has looked at in two years.
- Do not compute twice. A cache is rarely a question of elegance: it is work that does not get repeated. Tuning queue processing so race conditions stop happening is the same logic, and that is where we have seen the clearest gains.
None of this is new, and none of it is a sacrifice. It is the same care we mean by right-sized engineering in the first place: a prototype is allowed to be lean, a product in operation needs depth, and neither should be built larger than it has to be.
The power bill and the cloud bill are the same bill
These decisions save electricity, and they save the same line on the cloud bill. Bring your infrastructure down to the load you actually have and you do not need to choose between a cost argument and an energy argument. It is the same architecture review.
That is why this subject reaches us through operations rather than through a sustainability department. Operations is where the sizing gets decided anyway.
One clarification, because the question comes up naturally: the roof covers our office, not the platforms we operate for clients. Those run at Infomaniak in Switzerland and Hetzner in Germany, for reasons of jurisdiction, and the two operators report on their data centres’ energy themselves.
Right-sizing is the actual contribution
We did not build the roof. We benefit from it, and we are glad it is there.
What we do build is the sizing: how large a cluster really has to be, which method solves a task, what belongs switched off overnight. That is what decides, for years, how much electricity a system draws.
The pleasant part is that the efficient architecture and the cheap one are the same architecture. Nobody has to choose between the power bill and the cloud bill.
