WEBVTT

1
00:00:03.850 --> 00:00:09.978
AI is the fastest-growing consumer of energy in the ICT sector, and its demand is still climbing.

2
00:00:09.978 --> 00:00:18.269
So any project proposing to use AI to reduce software energy has an obvious obligation: prove the cure costs less than the disease.

3
00:00:18.269 --> 00:00:19.711
GreenCode takes that literally.

4
00:00:19.711 --> 00:00:26.560
The carbon of the pipeline's own training and inference is modelled, measured and reported alongside the saving it produces.

5
00:00:26.560 --> 00:00:34.851
The accounting is simple enough to audit, the reduction repeats every year the optimised version is deployed, and the carbon is spent once.

6
00:00:36.451 --> 00:00:41.954
The ratio only holds because the pipeline is engineered to spend as little inference as possible.

7
00:00:41.954 --> 00:00:50.209
That is what energy-efficient AI means in practice, and it is as much about the agentic framework around the model as the model itself.

8
00:00:50.209 --> 00:00:59.152
The models are tuned for the narrow task rather than the general one, so a request costs a fraction of what a general-purpose one would spend.

9
00:00:59.152 --> 00:01:03.967
A step reads only the fragment of the codebase it needs, not whole files.

10
00:01:03.967 --> 00:01:07.750
Work is routed to the lightest step that can do it.

11
00:01:07.750 --> 00:01:10.158
And successive runs process only what changed.

12
00:01:10.158 --> 00:01:17.037
An agentic framework has to be efficient as well as effective, or the saving it finds is spent finding it.

13
00:01:18.637 --> 00:01:22.560
The second half of energy-aware computing is when the work happens.

14
00:01:22.560 --> 00:01:29.692
The same computation on a grid full of wind costs a fraction of the carbon it emits on peaking gas.

15
00:01:29.692 --> 00:01:36.468
Because GreenCode analyses a codebase asynchronously rather than interactively, it can hold intensive jobs until clean power is available.

16
00:01:36.468 --> 00:01:41.818
Nothing is waiting on the result in the next second, so the latency is free.

17
00:01:41.818 --> 00:01:47.881
Copilot-style assistants run inference on every keystroke, wherever the developer happens to be, and optimise the typing.

18
00:01:47.881 --> 00:01:53.230
GreenCode optimises the system, in one scheduled pass, with the cost of that run counted.

19
00:01:54.830 --> 00:01:57.735
Scale decides the answer, and it varies enormously.

20
00:01:57.735 --> 00:02:06.085
A compact content system installed on hundreds of millions of sites carries a potential saving vastly larger than the run that finds it.

21
00:02:06.085 --> 00:02:16.615
So GreenCode runs a pre-assessment before it commits: the lines to be processed, the likely improvement, the size of the deployed base, against the modelled carbon of the run.

22
00:02:16.615 --> 00:02:22.424
Where the return does not justify the inference, the honest answer is not to run it.

23
00:02:22.424 --> 00:02:27.507
That check applies to GreenCode's own operation as much as to anything it processes.

24
00:02:29.107 --> 00:02:32.274
The same methods matter well beyond a code optimisation pipeline.

25
00:02:32.274 --> 00:02:45.891
Energy-efficient AI, an agentic framework that is efficient as well as effective, and scheduling against the grid are the techniques that make AI viable at the edge, on battery power, and on constrained hardware, where there is no option to spend more energy.

26
00:02:45.891 --> 00:02:50.325
That work transfers directly to anyone trying to run useful AI on limited power.

27
00:02:50.325 --> 00:02:59.825
Reach out to us with the AI workload you are trying to fit inside a fixed energy budget, and we will help you understand how GreenCode can get it there.
