The green software movement is moving from theory to practice. At last week’s Green IO conference in London, major organizations like HSBC, OVO Energy and the British government shared real-world case studies demonstrating how they’ve reduced carbon emissions through sustainable technology practices and cut costs in the process.From HSBC’s open source tool for measuring software carbon footprints, to OVO’s ingenious use of customer home boilers as distributed computing nodes, these aren’t mere feel-good experiments. These are strategies that demonstrate how environmental responsibility and business sense align.As AI adoption accelerates and concerns about its environmental impact grow, these practical approaches offer a roadmap for making technology more sustainable at scale.HSBC: A Holistic Approach to IT SustainabilityHSBC employs 35,000 full-time staff, plus a vast number of contractors. It has a publicly stated group ambition of being net zero for greenhouse gas emissions across its operations, supply chain and finance divisions by 2050.Natalie Pullin, keynote speaker at Green IO, works as head of sustainability at HSBC Technology, which is part of the CTO function at the financial services group. She coordinates a cross-functional team that takes a holistic view of sustainability.“Our focus is to ensure that we purchase and install only what we need, design things in the most sustainable way, optimize the run phase and understand how we can decommission e-waste in the most appropriate way,” Pullin told the Green IO audience.The team has seen some early successes, she said, such as switching the way it ships laptops to its employees from air to sea freight.For software, the company’s starting point is to gather data, Pullin said: “We want a single way to measure energy and carbon, and eventually water and e-waste, across technology.”HSBC uses public cloud, but it also runs a number of its own data centers and a private cloud estate. In 2023, the team started work on a methodology — and corresponding open source Java Spring application — for measuring the real power draw and carbon emissions of software applications.The HSBC team created Measuring Real Power Use of Software Applications (MRPUSA), which allows businesses and consumers to accurately assess emissions at an individual software level and provides granularity in reporting.Having designed the product, the team then had to drive adoption. Given the sheer number of employees at HSBC, the question Pullin has been grappling with is how to make sustainability an implicit decision characteristic for tech.“It is not going to be a new priority in and of itself,” she said. “But I would love to say in the future that sustainability helps HSBC Technology achieve its commercial goals.”FinOps is often used for this, and the HSBC sustainability team does align with the company’s FinOps team. “There are a lot of grassroots cottage industries working toward sustainability,” Pullin said. “As a big organization, we are trying to bring those people together, to create a set of standards across different IT asset classes.”The language the team uses also matters, she suggested. “We call ourselves the ‘net-zero technology’ function because it aligns with HSBC’s narrative around its net-zero transition plans,” she said. “Clearly, this program is bringing massive change within the group.”Such initiatives also need buy-in from the top. “Nothing would be possible without the support of senior leadership at HSBC,” Pullin said. “Support comes in different guises, from investment and involvement in the group steering committee, to opening doors and getting access to different teams. We’ve learned how to utilize our stakeholders to keep nudging this topic forward.”Encouragingly, she has not encountered much opposition. “I haven’t met anyone who is anti this topic,” she said. “I’ve met people who are cynical or neutral, but most people just ask, ‘How can we play a role?’”The Environmental Impact of the AI BoomOf course, no tech conference in 2025 can escape featuring AI as a key topic, and the implications of AI on green software are significant.This is in large part down to the speed of adoption. A paper on ChatGPT published in September by the National Bureau of Economic Research reported a remarkable statistic: By the end of this past July, the chatbot had more than 700 million weekly active users. This is nearly 10% of the world’s adult population (although, of course, not all ChatGPT users are adults).“It was the fastest adopted piece of technology humanity had ever seen,” said Green IO’s opening keynote speaker, Strategically Green Founder Anne Currie. “This suggests that arguing against its use on environmental grounds is not going to be effective, because you’re arguing against something that people really want.”Currie suggested that one of the interesting aspects of transformer models that underpin this generation of chatbots is that we have yet to hit a scaling limit.“Ever the showman, [OpenAI CEO] Sam Altman gave this a name — the ‘AI scaling law,’” Currie said. “It’s not a law of physics; we will at some point hit a limit. It’s more like Moore’s Law — the kind of thing you call a law so that investors keep giving you money. It’s very effective at doing that.”Equally importantly, the big cloud providers realized how important the AI scaling law was for them. “In effect, Altman said to the hyperscalers, ‘There is unprecedented human demand for something that can be delivered using your superpower of hyperscaling,’ which was exciting for them. But that scaling is mostly going to be fueled by fossil fuels.”DeepSeek, the Chinese startup that provided similar functionality to ChatGPT at a fraction of the carbon cost, offered some hope. The company used a number of approaches, such as reducing bit precision for weights and activation values from 64-bit to 8-bit, plus a mixture of expert models.However, there are growing reasons to be skeptical of some of DeepSeek’s accomplishments. “I think if they had really made them all, they would have seen far bigger than their claimed 10x reduction,” Currie argued. “But it almost doesn’t matter, because it demonstrated that there was demand for cheaper, more efficient models.”Equally important is the rapid growth of renewable energy sources, she said. “AI is one of the fastest-growing pieces of technology we’ve ever seen. Another one is green energy production, particularly using solar. If we can align AI with renewables, then we make the problem significantly less without arguing against AI.”DeepSeek’s mixture-of-models approach is being used by NTT Data. Andy Westbrook, who runs that company’s FinOps practice, spoke at Green IO about his company’s internal use of AI.“We are seeing our consumption of tokens grow exponentially month by month,” Westbrook said. “In April, we were consuming 5 billion tokens per month. In July, that was 22 billion, and it’s probably doubled again in the last couple of months.”Seeking to make its AI use more sustainable, the company started looking at carbon and water consumption by large language models (LLMs). Obtaining this data is notoriously difficult, so NTT worked with specialist firms Greenpixie and Ecologic to obtain the necessary data.This process gave NTT a metric per 5,000 tokens for each LLM. Armed with this data, the company’s developers can test the lower-consumption LLMs to make sure that there isn’t an impact on latency and performance, then shift workloads accordingly.The firm tests new LLMs all the time for performance, carbon kilowatt and water consumption, Westbrook said. “We’re currently testing Llama, Claude, GPT-5, GPT-5 mini and nano, in comparison to what we’re seeing with GPT-4.”Armed with this data, and using orchestrators such as LangChain, Semantic Kernel or Prompt flow, NTT’s developers are able to do further optimization. “They’re breaking down a given prompt, sending the complex parts to a complex LLM and the rest to a cheaper LLM,” he said. “We are therefore optimizing that AI use case, and selecting LMMs that have a lower carbon density.”Where you run your workloads also matters, so Westbrook’s team is encouraging developers to run their AI workloads via Microsoft Azure in Sweden, which has largely carbon-free electricity on the grid from a mixture of hydro, nuclear and wind.OVO Energy: Using Waste Heat for CI/CD PipelinesTypical data center Power Usage Effectiveness (PUE) is around 1.6, which means that 60% of its energy use is for cooling. Mark Buss, digital sustainability manager at OVO, a U.K. energy supplier, presented a fascinating proof-of-concept case study at the conference, with an approach to using waste heat in collaboration with Heata and CarbonRunner.Heata, an English startup based in Godalming, Surrey, has created an innovative cloud network where computers are attached to the boilers in people’s homes. It is a wonderfully elegant idea, effectively using electricity twice by providing services to cloud computing and domestic hot water.The British energy provider OVO is combining boilers with servers that run CI/CD jobs, as a way to make use of waste heat.CarbonRunner provides a mechanism for running GitHub Actions in the lowest-CO₂ regions across Amazon Web Services, Azure, Google Cloud Platform and now, also, Heata units.OVO, Heata and CarbonRunner are collaborating on a project to make use of waste heat.“We’ve ring-fenced a few Heata devices that live in OVO customer homes,” Buss said. “And we run some of our CI/CD jobs via CarbonRunner, who farm those jobs out to Heata devices in our customers’ homes.”While the cost for each CI/CD job is small, running all OVO’s CI/CD jobs this way could be hugely beneficial. “We would save 17 tons of CO₂ equivalent and heat enough water for 24,000 showers, saving our customers £3,000 in the process.”Having tested the idea, Buss is looking at a larger-scale deployment, including for some AI workloads. “We’ve proved that CI/CD works, so we’re now looking at potential capacity and what other jobs we can run on those devices,” he said.GOV.UK and One Login: Greening Government ITThe U.K.’s Government Digital Service (GDS) team presented several case studies at Green IO, including one about One Login, the government’s single sign-on program.“Our director at One Login has a vision to make it one of the greenest digital services across government,” Jennifer Marks, sustainable tech leader at GDS, told the conference. “We want One Login to use as little carbon as possible, and demonstrate that technology can be designed sustainably and delivered efficiently.”In order to achieve this, GDS developed its own methodology aligned to the Greenhouse Gas Protocol and used it to get a baseline measurement, first for One Login and then for the whole of GOV.UK.Baseline sustainability metrics compare One Login and GOV.UK.With a baseline established, the team worked to get more granular data so they could measure improvements and identify which areas to target. This was based around three use cases — people logging in, identifying themselves and creating an account.However, the project team encountered a major hurdle. “While AWS could provide us with proxy data, the level of data we needed required AWS resource and account tagging,” said Marks. “Historically, One Login had not been strict with a tagging schema, meaning that a great deal of data was categorized as ‘other.’”Once this process was completed, the One Login team started to look for quick wins. From an architectural point of view, One Login is based on an event-driven, entirely serverless architecture, with AWS Fargate for the containers, DynamoDB for data, SQS for messaging and Amazon Simple Storage Service (S3) for long-term object storage.There weren’t a lot of options available for improvements, but one, which I explored in my ebook on sustainability for The New Stack, is processor choice.“We found that by changing one line in our code and switching to Graviton2, we can experience an average 72% reduction in power consumption,” Marks said.Another area the GDS team looked at was autoscaling.Visualization of a GOV.UK traffic pattern.A lot of traffic to government services follows a predictable pattern. “People want to conduct their business with the government during the day or over weekends — almost never overnight,” said Marks. “So you would think it should be easy for us to scale up and down according to the traffic pattern. But you’d be wrong, because we experience unexpected surges of demand through what we have coined the ‘Martin Lewis effect.’”Martin Lewis, a financial journalist also known as the Money Saving Expert, is one of the most trusted men in Britain. “On what should be a boring and predictable weekday evening, Martin Lewis goes on national television and tells the millions of viewers that they can save money by going onto GOV.UK — and the surge begins,” Marks observed.To run hot all the time would be unsustainable, so GDS uses step scaling to handle unpredictable spikes in demand without resorting to over-provisioning.As we’ve covered in The New Stack before, AWS’s Customer Carbon Footprint Tool is very limited in terms of the data available for detailed optimizing. So the next step for GDS was to partner with Greenpixie, which used its methodology to acquire detailed information on energy and water consumption for every service used.“Working with the Greenpixie team, we designed and developed a dashboard that provides deeper insights into our emissions data,” Marks said. “This allowed us to target necessary improvements and enabled us to monitor the impact of changes at a team level. It also meant that any regressions or increases could be addressed early.”One Login’s Greenpixie-powered dashboard.Even more importantly, the new dashboard could show sustainability against cost. “We could suddenly prove that making sustainable decisions wasn’t just good for the planet,” Marks said, “but could actively reduce cost.”One insight from the dashboard was that one of the largest areas of resource usage was network firewalls. “One Login had hundreds of network firewalls running on hundreds of virtual machines [VMs], a setup that consumed an exorbitant amount of compute, energy and financial resources,” Marks said.Work to address this is still ongoing, but network firewalls have been drastically reduced, slashing the energy and compute required, she said: “Early indications are that this has saved us 47% of our AWS spend.”Seeking to further embed what they were learning, the team created the snappily entitled “Digital Sustainability 101 Training” course. “To be honest, we weren’t sure how it was going to be received, but over 300 people attended,” Marks said.The success of the course, she suggested, was down to its practicality. “It explained what sustainability meant for each digital data and technology profession, and what actions individuals can take into their roles.”As One Login was moving to steady state, GDS engaged GoCodeGreen to make a final assessment. The results from that assessment were outstanding, she said: “They were some of the lowest ever recorded by GoCodeGreen — within government or the private sector — for a platform at scale.”The One Login team is now implementing sustainable maintenance. “We are embedding continuous optimization,” said Marks, “ensuring efficiency scales with growth, and keeping sustainability at the heart of our platform strategy and best practice.”Green IT is a fascinating, complex, cross-disciplinary field. Gaël Duez, who organized the Green IO conference, deserves huge credit for producing an event where respect for science, people and the planet allowed for proper discussion and debate amongst its 300 attendees.We don’t have all the answers, but it was fantastic to meet so many passionate people who wanted to make a difference. “I feel like I’ve found my tribe,” one delegate told me.I explore many of these topics in my ebook for The New Stack, “The Developer’s Guide to Cloud Infrastructure, Efficiency and Sustainability,” now available for free download.The post Real-World IT Sustainability: 3 Case Studies From Green IO London appeared first on The New Stack.