Bridging Product and Engineering for a Greener Future: Embedding Sustainability into Software Teams

Carsten Windler at UXDX EMEA. Video: https://youtu.be/Ro5B9gazErg

Readable transcript: edited from the recording's captions for readability (fillers and false starts removed, punctuation and section headings added). Wording is the speaker's own. Timestamps are positions in the video. Names marked [?] could not be verified against the audio.

Software keeps getting bigger

[00:00:09] Hello everybody. My name is Carsten. I'm a backend developer from Germany, as you can see from my slides and as you can tell from my accent. Sorry for both already. Let's jump right on it.

[00:00:21] You might have noticed that there's a trend of things getting bigger, larger, heavier over time. This cars example, for example. Yaks, they tend to get bigger. My belly tends to get bigger, unfortunately, if I don't do anything about it. And we have software. Some of you might still remember these black things where we saved data on back in the days.

[00:00:48] Why is this a problem, or why am I here today to talk about this? I have to go back a little bit in time, early corona days, let's say. We were at home watching a lot of YouTube videos. We had time back then. And there was this talk. I don't know why YouTube brought this up in the timeline, but it was really for me something which changed my life. I have to say, these are the little moments in life which sometimes can change everything. A talk from Asim Hussain from the Green Software Foundation about sustainable software engineering. I had no clue this actually existed, but I learned so much about this in that half an hour that I decided to pivot my career and want to work in green tech afterwards. So that's a really cool thing.

[00:01:28] Why is this now a problem? Why is software getting bigger a problem for us? You can see it here on the left on the graph. I'm sorry it's in German, but I think you get the idea. This is the actual and the projected demand for electricity in Germany for the ICT, the information and communication technology sector. And you see it's going up, especially the dark blue part here, which are the data centers. So the demand is increasing. This is not only a German thing, this is happening worldwide. UK, US already have 6% of the electricity demand coming from data centers.

[00:02:06] Now, sustainable software engineering is actually about helping on that problem. Just very briefly, we have three main pillars here. We have energy efficiency, which basically means we want to use the least amount of electricity possible for the same operation. We have hardware efficiency, which means that we want to emit the least amount of embodied carbon. And this is every time when you open your laptop, think about that this laptop, the production of this laptop, already has emitted something around 500 kilograms of carbon, like half a ton of carbon, just when you open it, it's already there. So this we also want to reduce. And there's the concept of carbon awareness, which means that we want to use electricity when it's greener, when there's more renewables in the grid.

Raising awareness at Plan A

[00:02:59] This is all cool and you can talk for hours about that. The thing is just, how can we embed sustainability within software teams? Because this is a different thing. I can of course do my own stuff, but when it comes to working within software teams, it gets a different thing. So this is a little bit the story of what we did at Plan A a while ago.

[00:03:25] First, the plan was to raise awareness, which is always good, awareness always sounds cool. So basically what we did first, raise awareness that every operation causes emissions. This is something which you have to have in mind. The second thing which we would like to put on the table is that we can measure and reduce these emissions, I will come to this later. And of course raise awareness that architecture decisions also play a large role in this. You can see it here on the right side, a simple example, just start highlighting that any architectural decisions you do indeed have an impact on the environment. So some architectures use less energy, some use more.

[00:04:13] And this is exactly what we did, highlight the carbon emissions. And actually the feedback was quite positive, which, okay, maybe not so surprising because we work in a company which is about carbon accounting. But still, it went quite well. We started with internal sessions where we talked a little bit about how this all works. We started what we called the Green IT chapter. It's basically just a work group where people were invited to join. We could talk about this topic. We started hosting internal brown bag meetings. We had this at the company at the time, these brown bag meetings where we can just talk about anything. So here on the right side was just one of the slides we used there. And at some point we even hosted meetups. So it was really cool.

Tracking emissions you cannot see

[00:05:03] But of course, we can talk about things, but you really need to get into action. So the second thing which I wanted to do was to track emissions. And this is where it gets interesting, because as you can imagine, when you open your laptop you don't see any fumes coming out of it, which I think is a good thing, because otherwise imagine everybody here sitting in the room and all your laptops would have this little exhaust fume coming out of it. This would be quite not so cool. But the thing is that because of that, we don't see the emissions, we don't think of them. If you do, I don't know, npm install, you feel how the laptop is getting warm. But this is everything. We don't really think about this. And the other thing is we cannot measure these emissions directly, because for obvious reasons you don't see them coming out.

[00:05:58] So how to measure these? Now, there are various approaches. I cannot go through all in detail here, not enough time, I would really love to do so, but let's just look at some tools. Let's look at the first approach. You might know this, this is the Website Carbon Calculator. This is a tool which people use to estimate the carbon emissions of a website. So I used CNN here as an example, and you can see you just type in the name of the domain and it gives you this number here, 1.17 grams. You might think, okay, that's not so much, maybe, per page view. Okay, what is one gram of CO2? Maybe you have to keep in mind that CNN has 480 million page views per month, each one gram of CO2. I think you can imagine how much that is, or you can ask ChatGPT to calculate that for you.

[00:06:59] But this is not the point actually that I wanted to make. The point is, how do they come up with this number, this 1.17? This sounds really quite accurate, right, a very, very accurate number. Truth is, nobody really can. This is an estimation. So there's this Sustainable Web Design model which is used by many of these calculators. And yes, don't worry, I will not go through everything here. This is the equation which is behind that, or the system which is used to calculate that. Basically it's a proxy. It's data transfer which we take, we multiply it with some emission factors, saying some things have more emissions, some have less emissions, and the carbon intensity, which is basically the dirtiness of the grid. And then we come up with emissions. We do some more calculations and whoops, we have this number on the website. So you can see this is not very accurate of course, but it's a good start. We have to start somewhere.

Getting emissions data out of the cloud

[00:08:02] Of course, nowadays applications look a bit more like this. This is just a random screenshot from some AWS documentation page. But I think many of your companies' applications or infrastructures look pretty much the same, or equally or even more complex. So how do we get the emissions from that?

[00:08:25] You could use, for example, their tools, the AWS sustainability module, which is actually relatively new, where they did actually quite some good work this year. I will not go into detail, scope 1, 2, 3 and this MBM and LBM, this is all carbon accounting specific stuff, but the good thing is that there is already something you can use in AWS to calculate your emissions. There's here some graph which shows it over time. But there are some drawbacks with this tool.

[00:08:58] One thing is that actually everything is "other". We use of course not only S3 and EC2, so basically everything else is other, which doesn't really help me because it doesn't show me where to reduce emissions. And the other thing is this screenshot was made in May, so this month, but it only dates back to March. This is okay if I want to do carbon reporting on the yearly level, but it doesn't help me if I want to do short-term optimizations. If, let's say, I tweak my infrastructure and I have to wait two months to get an idea if this worked or not, this is not really cool.

[00:09:35] Luckily, there are a lot of people who spend a lot of time trying to solve exactly that problem. I cannot go through all, they all have different approaches. Kepler for example is for Kubernetes clusters, GreenFrame is perfect for Docker containers, CodeCarbon for your Python code. Just some name dropping here. We at Plan A decided to go with the tool Cloud Carbon Footprint.

[00:10:03] And this is actually a very cool tool because it uses the usage statistics from AWS or any other cloud providers which you like. So basically cloud providers, they do track everything you do, which is okay, this is their business model, but you can get this data from them. They can tell you exactly how much CPU hours you consumed, how much RAM you consumed over time. And this tool takes the values and puts it into carbon emissions.

What the data showed us

[00:10:34] And this we used to build a case, because I finally wanted to know what we can do with that. So you can see here, it's a bit small, I apologize, but you see the emissions going up over time, over two years back. Now, this is when we ran this tool the first time, and there are actually two interesting things happening here. The first, the arrow on the left side, this was when we introduced a new internal data layer, and this was to be expected, that there was some increase in emissions. It was a new tool, very, very computational heavy, so that's okay.

[00:11:13] The second one on the right hand side was a bit surprising, because at that time we did not introduce anything. So I went into the logs and tried to find out what happened at that moment, and it turned out that we did have some performance problems at this time. And what do developers do when they have performance problems? Yeah, they just add more instances to the whole cluster. So just add more power to it and don't try to optimize, just add more. And this is exactly what happened.

[00:11:45] Now, this is actually also not such a bad thing, you can do this, but you have to make sure that it works fine, and in our case it did not. It was just some autoscaling which was misconfigured. So basically the cluster ran on full speed the whole time, although this was actually not necessary, which was for us a great thing because then we could fix that easily. And you can see here on the left hand side, this was from our development environment, how after applying the correct configuration the task count, so the number of instances in the cluster, went down immediately and brought us quite some success.

[00:12:28] We also had some other things. For example, there's another tool at AWS which is called the cost optimizer, and it gives you some indications where you spend too much money, which is actually surprising when you think about that it's AWS, but yeah, they do. Here for example you see that we had some instances which we were using for other things. They were just overprovisioned. This happens when you have no idea how much power you need. You just start with a number and then you forget about it. And this tool tells you, look, you can rightsize the instance to save some money. It's not much, again, this was here in the development environment, but at least everything counts.

[00:13:08] And after a couple of little tweaks, we ended up with a pretty big reduction in carbon emissions: 70% less, which was really cool. And you see three curves here. Again, it's a bit bad to see. One are the CO2 emissions, the other is the kilowatt hours used, and the other one is the cost. And they all are pretty much the same, so it actually doesn't really matter. And this is a cool thing because we not only reduced 70% emissions, but also 70% costs at that point. And this was of course something which is always cool for your manager. So if you need some nice buzzwords for your management, go with ROI of decarbonization. That sounds pretty cool. So it's cost reduction, it's better performance when you apply some, I don't know, autoscaling or something, this usually leads to better performance, and improved stability because green wants us to use simpler systems.

The product management boss fight

[00:14:17] So far we looked at the more infrastructure side of things. But now came the big, let's say the big boss fight: product management. This is of course something where it gets a bit complicated, because let's face it, we are a startup. We have our tight roadmaps, we have plans, and now somebody else comes up and says, hey, let's add some more work here.

[00:14:44] How it went so far. As I showed you already, we started adding sustainability as a non-functional requirement. So thinking about, when you implement something, what is the impact on the environment? And second was that we started introducing some sustainability debt. Think about this like technical debt. So we started gathering ideas of what things we can change, we started writing them down, and as with all tech debt in the world, what happens with tech debt? It ends up in the backlog and there it stays. But we have a Green IT backlog with ideas which we can use to work on that.

[00:15:30] Last thing what we did was talking about KPIs. I already showed you here on the right hand side things we can do. For example, we can measure total carbon emissions with the tools I showed you. We can count the zombies in our infrastructure. Now a zombie is a server which is idling around. So it's not eating brains, but it's consuming energy and it's doing nothing. Maybe you have this probably also in your company, some forgotten servers here and there which are just running around. The thing is, a server which is just idling is also consuming a lot of energy. Just to keep it operating, to keep everything running, already consumes quite a lot of energy. So reducing the zombie count is good.

[00:16:21] You can look at the response times of your application, this is something you hopefully already do. And also, for the designers or the front-end developers, the front-end page weight. So make sure that you stay within a certain budget when you have the size of the application, and all the other things. So these are basically KPIs which many of you probably already cover or collect, but you can use them also for our case here.

Headwinds: the EU omnibus, climate skepticism and AI

[00:16:51] The challenge of course was time and resources. As I said, startup, small team, a lot of things to do. And then came also a couple of other challenges. Speaking about challenges, we had for example in 2025 the EU omnibus, which maybe some of you have heard of. Basically it was a couple of laws which made carbon accounting or sustainability reporting less strict. So of course for a company which is working in this field, you can imagine this changes a lot. So we had to adjust here.

[00:17:28] Also the kind of political schema is a bit changing, I have the feeling. So climate skepticism, or climate change skepticism, is becoming more and more a thing, I have the feeling. Even though this wasn't back then, it's just happening in the last couple of years. And all this led to some reshuffling. We had to rethink our roadmaps and everything.

[00:17:53] And then one thing more, of course, the elephant in the room, I always say, and this is actually the only AI generated image here which is in my slides. AI, of course, I know we already talked a lot about it here today, but there's also some aspect which many people are not looking at, and this is of course the energy consumption of AI. Sam Altman will tell you this is not much, this is just a little bit and half a tablespoon of water and I don't know. I'm not sure about you, but I don't trust that person.

[00:18:31] So it's better to start asking the experts about it. There are tools like for example the AI Impact Tracker. This isn't a website, I'm sorry, it's a browser extension which you just activate and then it tracks the ChatGPT messages or the conversations and gives you in the end some idea of how much energy you consumed. Again, it's an estimation of course. And it puts it into scale, like if all 900 million ChatGPT users would use it the same way as you, how much energy would that be? In this example they come up with one terawatt hour of energy per year, which is already the consumption of a little city, I'm sorry, country.

[00:19:14] Now you can say, okay, chill, all good, we are using Claude, so we're on the safe side and Anthropic are the good guys. In my opinion, no, they just have good marketing. The thing is Anthropic now hosts Claude in the data centers from xAI, and these data centers, Colossus 1 and 2, they have the little specialty that they are powered by illegally installed gas turbines. So you can imagine, when you put a huge data center somewhere, the local grid has to keep up with that. In this case it didn't. So what do tech companies do? They just install illegal gas turbines.

[00:19:54] You can say, okay, this is just a temporary solution. In fact, xAI just installed 19 new ones. Although being already in an ongoing lawsuit about that, they decided to install 19 more. And these gas turbines, they are terrible for the environment and they're terrible for the local surroundings. So people around these data centers really struggle to breathe. But most of us actually don't care, as long as we can keep on lollygagging or dillydallying or creating these images which everybody somehow likes to do, I don't know. So this would be actually enough for a different talk, and then there's another talk I would love to give, but time's running out so I have to come to an end.

How it went, and lessons learned

[00:20:41] How it went. As I already said, we ran into a little bit of trouble on the long run, the bumpy ride in 2025. However, we used the time well I would say, because we worked on an improved data layer. The other one which I already showed you was not very efficient, we have to say, so we reworked that and we had a huge efficiency gain through that. And you can see it here in this graph actually, because you can see from May 2025 the emissions have decreased quite a lot. This was also because of this whole work which we did.

[00:21:21] And if you would ask me for lessons learned in the whole process: driving sustainability from below is really hard. So we tried it like from grassroots, if you like. This is very, very cumbersome because you have to constantly push it, you have to drive it yourself somehow. So next time I would probably go directly to the management and try to get this approved.

[00:21:51] And last thing, observability. This is also something, if you want to enable teams to work on this topic, you should make sure that they have the data at hand. So this is what we didn't have. We had this tool which tracked the emissions, but the data was not available easily for others and it was not really fine grained. So we would have needed to invest more time into making this more fine grained, like saying this is for team A, team B, so that they could work on their own emissions.

[00:22:21] To wrap it up, if I could give you just one little thing to take away: consider sustainable IT, because it reduces emissions and costs, and with costs you can usually go to the management, because it needs management to support it to fully thrive. Otherwise, again, it's more like a small movement. And it's an ongoing effort, you have to just keep going, especially with AI now. I have to read so much, I can spend the whole day just going through all the newest data about AI and sustainability, it's crazy, but we have to do it.

[00:22:58] Okay, at the end of the slides this is just my typical information slide. I have at the very end some GitHub resources here which you can have a look at if you like. I find it very interesting that GitHub itself has a green software directory. So they are also quite a bit into sustainability, which is cool. With that I leave you. Thank you very much for your attention. If you have any questions, feel free to connect or catch up afterwards. Thank you.

Q&A

[00:23:35] Host: Thank you very much. It's a great point about it just being invisible. To most people there isn't, luckily, an exhaust coming out of our laptops, but that doesn't mean it's not happening somewhere, and increasingly with dodgily built data centers. Okay, we'll jump into the questions. Is reducing cloud costs always also reducing CO2 emissions? Reducing cost is easier to prioritize than reducing emissions, and I wonder in which cases these both goals are not connected.

[00:24:10] Carsten: Usually you can say it depends, unfortunately, is the answer. If you reduce cloud costs, sometimes it means that you shift your compute to a different region, and every cloud region has different carbon intensity. So for example, if you have it in Sweden, which is usually a bit more expensive than, I don't know, some other regions like Ireland for example, I think these are quite cheap, this is why startups usually go to Ireland. So reducing the costs would not reduce the carbon emissions. But if it comes to compute, so if you say I want to just reduce the use of my compute, then it usually does, because less compute means less energy consumed and with that the CO2 emissions.

[00:24:58] Host: So it's about looking at the efficiencies rather than just moving it around to a cheaper supplier.

[00:25:04] Carsten: Yeah, exactly.

[00:25:07] Host: Okay. Why is anyone in the western world even worried about CO2 emissions? Wake up. Such a long text. Okay, wow. So you clearly mentioned some people don't believe in climate change. I'm going to summarize this: are we damaging ourselves by kind of shooting ourselves in the foot, making it harder in Europe when other countries don't care about it? So we're making ourselves uncompetitive.

[00:25:38] Carsten: I wouldn't say that other countries don't care about it. Especially China for example, they are spearheading the renewable energy movement because they're actually smart, which we here in the western world are sometimes not, because especially in Germany we love our combustion engines, we love to burn stuff. They are actually more smart than we are in this case. So I would say we should rather keep up with them and not do what we do here, clinging to the past, but rather look into the future.

[00:26:08] Host: They installed solar and wind in China, I think it's the same as...

[00:26:13] Carsten: Yeah, I mean they still have a lot of fossil fuel of course, but they also ramped up the production of everything quite a lot. But I have seen just a number recently that the usage of fossil fuel has stalled right now, so the energy produced through renewables is so much stronger right now that they don't need extra coal plants for example.

[00:26:36] Host: So there is some good news. How is Plan A handling AI usage within the company? Are there any guardrails because of sustainability that other companies could learn from?

[00:26:47] Carsten: Yeah, this is what we're currently figuring out, because we are also exploring AI a lot. I use AI also a lot. Shame. But well, I have to, because as a developer we have to keep up, and if you don't use AI people look at you like, huh. But to the question, we are working on that. What I also mentioned is it's very hard to tell the emissions actually from AI, so this is something which I would really like, I still need to work on this. So far we don't have them. We have our policies of course, but not regarding sustainability, which is something which we need to do definitely.

[00:27:27] Host: Okay, and it becomes difficult with your own cloud. You know at least where it's being hosted, but with AI you don't know which centers it's hitting, and what the carbon footprint of it is.

[00:27:36] Carsten: I personally fancy locally hosted LLMs for example, this would be one thing, but we are not there yet.

[00:27:44] Host: People love convenience. Have you found some good alternatives that create less carbon footprint for tasks? Most people use AI and not a local LLM.

[00:27:55] Carsten: Good alternatives. That's actually a good question. I would rather reframe this. Do we need AI for everything? Do we need to always create any new images or movies, like these little short videos for example? They are super energy consumption, so this is really, really bad stuff. Do we really need that? I would ask that question. Do we, as product for example, do we need to put AI everywhere? I see everywhere now these little buttons where you can say enhance with AI, or it just does it for me without asking me, like the resume which Google for example does, just without asking me just gives me the resume of my question. This is something I would actually rather do. It's really hard to find an alternative for code, I have to say.

[00:28:44] Host: Fighting with that. And people love convenience. I think it's something that the cat's not going to go back in the bag.

[00:28:52] Carsten: AI will be here to stay. We just should really think about using it responsibly and not just burning tokens because token mixing is the new thing.

[00:29:02] Host: Yes. While CO2 emissions from AI grow, hopefully renewable energy percentage increases as well. From your point of view, is the emission problem one that will solve itself? And if not, can you make the scale of the remaining problem tangible for us? There was that study that I saw as well that there's no new coal plants because the amount of energy that we increase is being met by green energy. But what is the scale of the problem?

[00:29:33] Carsten: This is really hard to tell because the companies don't tell us anything. But from the example I just showed you in the talk with the Colossus 1 and 2 data centers, maybe globally there are not more fossil fuel plants installed, but the data center is a local thing and the local grid needs to keep up with that.

[00:29:57] Carsten: And will the problem solve itself? I found this very interesting, this thought of having this kind of carbon debt which we already have. So we emit now a lot of carbon or other greenhouse gases like methane through that, and this is there and will stay in the atmosphere for thousands of years. So the things we do now directly affect our future generations. And if we say, ah, AI will solve climate change in the future somehow, I don't know how, this is what also tech CEOs will tell us, climate change will be solved, nobody knows how, but we give it full speed now to generate AI slop. I'm sorry, I probably didn't answer the question properly.

[00:30:39] Host: I guess one positive thing is every year they have to re-estimate the growth of green energy, because even their best estimates keep getting surpassed by the installed amount. So I guess that's the one positive.

[00:30:50] Carsten: But the increase of energy demand of AI is crazy.

[00:30:56] Host: Okay, and that actually brings us to time. So thank you very much, Carsten.

[00:31:01] Carsten: Thank you. Thanks everyone.