Sustainability
The cheapest energy is the compute you never spend.
Most of what gets filed under AI sustainability is an offset bought after the fact. We think it belongs earlier, in the architecture, where the decisions that determine consumption are actually made. It is how we build, it is work we do for clients, and it is what we are researching.
Our position
Efficiency is an engineering decision, not a reporting exercise.
The energy a system consumes is settled long before anyone measures it. It is decided by what gets sent to a model, how often, how much infrastructure sits idle waiting for traffic that never comes, and whether the same answer is computed a thousand times because nobody cached it. By the time that reaches a sustainability report, the choices that mattered were made two years earlier by engineers who were not thinking about carbon.
So we treat it as an engineering target. The useful part is that it is not a trade against anything: the design that consumes less is generally the one that costs less, answers faster and exposes less data. We are not asking anyone to pay a premium to be responsible.
How we build
Four habits, applied to our work and to yours.
Smallest sufficient input
A model gets what the question needs and no more. Most analysis does not require a model to read the data at all, only its shape. Restraint here is a privacy decision, a cost decision and an energy decision at the same time.
Compute sized to the question
The smallest model that clears the bar. Cache what repeats. Do in ordinary code what does not need a model. Most of our engineering effort goes into not making the call.
Data minimisation by default
We move summaries rather than copies, and working data is removed on a schedule rather than kept because storage is cheap. Less to move, less to store, less to secure.
Infrastructure sized to the load
Architecture built for the traffic that exists and the next order of magnitude, not for a hypothetical one after that. Over-provisioning is carbon and cost that nobody ever reviews.
What we do for clients
Sustainability work that shows up on the bill.
Every engagement below has a commercial case as well as an environmental one. That is deliberate. Work that depends on goodwill gets cut in the first difficult quarter.
Green cloud and FinOps
Right-size workloads, retire what nobody owns, and schedule what can wait. We make running cost a design input rather than a monthly surprise, and the same changes that cut the bill cut the energy behind it.
Cloud engineering →Efficient AI engineering
Reduce the cost and energy of each answer: model selection, caching, retrieval that fetches less, and evaluation that proves a cheaper approach is no worse. Deployed AI that nobody has measured is usually paying for capability it does not use.
AI engineering →Operations and supply chain optimisation
Routing, allocation, scheduling and cost-to-serve. The decisions that repeat weekly are where waste accumulates quietly, and where a better answer compounds.
Supply chain →ESG data and reporting
Disclosure is a traceability problem before it is a reporting problem. One definition per measure, one place it lives, and every published figure able to name the source it came from and be reproduced on demand.
Data strategy →Research
Where we think the next reduction comes from.
Early work, described as early work. None of it goes in front of a client until it beats the conventional approach on a measured benchmark.
Optimisation for energy-intensive decisions
Quantum and quantum-inspired methods applied to routing, scheduling and allocation. These are the problems where a better answer moves fuel, electricity and idle capacity, not just a number in a report.
Measuring efficiency honestly
Benchmarking energy and cost per answer between a conventional design and ours. Until that measurement exists we describe the architecture and not the saving.
Physical AI in operational settings
Perception and control in warehouses, clinical settings and plants, where efficiency and safety are the same engineering problem.
For the review board
The questions that hold up an AI purchase.
Every AI purchase reaches the same meeting. Here are our answers before you have to ask.
Does our data train your models?
No. Your content is used to answer your request and nothing else.
Can you show how a number was produced?
Yes. Published figures trace to a stored calculation that can be re-run and compared.
What does a model actually see?
The smallest input that answers the question. For analysis that is a statistical profile and a capped sample, not the dataset.
How long is our data kept?
Working data is removed automatically on a schedule. What persists is the summary, not a copy of the source.
Can the AI take an action on its own?
No. It proposes. Execution is ordinary code behind a validated contract.
What we have not done yet
The part most sustainability pages leave out.
Three things we are often asked about and cannot yet answer with a number:
- We do not publish a carbon figure. We have not measured one we would stand behind, and an unmeasured number is worse than none.
- We are not claiming a certification we have not completed. When we hold one, it will be named here with its date.
- We have not benchmarked our approach against a comparable system built the conventional way. The engineering argument is sound; the measurement is still owed.
What we will do is answer any of it directly, in a call, with the engineer who built the thing you are asking about.