AI's Rising Energy and Grid Pressure

• Local opposition is no longer a footnote to data center expansion. In Q1 alone, community pushback blocked or delayed at least 75 U.S. data center projects worth more than $130 billion, matching the total for all of 2025
• Air pollution is emerging as the next front in that pushback. At least 82 gas burning power plants are now planned or under development specifically to power data centers, shifting the debate from electricity demand alone to public health
• Grid strain is also showing up off the books. Developers are increasingly turning to behind the meter fossil fuel generation, with one estimate putting potential U.S. capacity above 40 gigawatts by 2028, a trend a new tracker is now attempting to quantify by provider
• Transparency gaps are structural, not incidental. Only 36 percent of data centers in Europe currently meet their sustainability reporting obligations under the EED directive, and there are no penalties yet for the other 64 percent
• The macro trend line has reversed. Rather than continuing a shift toward decarbonization, the AI compute build out is now actively increasing the internet's carbon intensity, a pattern one analyst called recarbonisation
• Efficiency gains are not translating into lower total consumption. Jevons Paradox, the pattern where efficiency improvements increase overall usage rather than reduce it, is being cited as the reason renewable capacity and location strategy matter more than efficiency alone
• Regulatory frameworks are trying to catch up. The UK has classified data centers as Critical National Infrastructure with a goal to triple capacity by 2030, while the water sector highlights that data centers remain excluded from some existing usage protections

‍

Cooling, Power and Physical Infrastructure

• Liquid cooling is delivering measurable efficiency gains. Transitioning from air to liquid cooling systems like the IR7000 rack lowers power usage effectiveness from roughly 1.4 to between 1.1 and 1.2, and grid aware workload scheduling cut emissions by 32 percent in a UC San Diego pilot
• China has switched on the world's first underwater AI data center off the coast of Shanghai: 192 racks powered by an offshore wind farm, running at over 95 percent renewable energy, saving an estimated 61 million kilowatt hours annually
• China's broader strategy pairs compute location with renewable supply directly, routing AI workloads to facilities near wind and solar generation in regions like Ningxia rather than optimizing after the fact
• Water is becoming as contested as power. AWS data centers in India are designed for zero water use in cooling, with replenishment efforts returning more water to communities than is consumed
• Not every water strategy survives contact with reality. OpenAI's planned Australian data center originally aimed to use recycled water for cooling, but pipeline and infrastructure constraints have reportedly forced a shift to waterless cooling instead
• Materials innovation is entering the conversation. Engineered timber is being positioned as a lower carbon alternative to conventional data center construction without sacrificing structural performance
• Market growth is following the modular route. The global modular data center market is projected to reach $96.84 billion by 2034, driven by edge computing, AI workloads and faster deployment timelines

‍

Measurement, Disclosure and Standards

• A survey of 250 IT leaders across EMEA found that 45 percent say their current data center cannot meet its own energy and carbon goals, 92 percent now choose infrastructure partners specifically to cut energy use, and 99 percent say data sovereignty will determine where their data lives
• Standardization is catching up to measurement needs. The ISO/IEC 30134 series now defines a common set of KPIs intended to let the industry measure data center sustainability consistently rather than by provider specific metrics
• Meta's exit from RE100 has exposed a deeper accounting problem: annual clean energy matching allows companies to offset gigawatts of gas use on paper while opposing more rigorous hourly and local accounting reforms to the GHG Protocol
• An Uptime Institute survey shows tracking is improving unevenly. 87 percent of operators now track energy consumption and 79 percent track PUE, but only 31 percent track carbon emissions from electricity use, and server utilization tracking has stalled at 37 percent
• Hyperscalers' reluctance to disclose AI data center energy and emissions data traces to more than reputation management. Google representatives cited a lack of standardized measurement methodology and limited telemetry as the actual barriers to apples to apples comparison across providers
• Finland's planned 30 million euro data center tax refund scheme has been delayed by at least six months after the government paused rollout to bundle it with a new mandatory sustainability registry law
• Measuring actual electricity consumption is being positioned as the starting point, not the endpoint, for credible Green AI claims, a distinction increasingly drawn by providers building AI tools specifically around energy transparency

‍

Green AI Research and Model Efficiency

• A new whitepaper introduces the Red AI versus Green AI framing, arguing that token usage alone fails to capture the real environmental cost of continuously running, always on AI agents, a problem it terms the blast radius
• A Q1 journal review paper distinguishes between Green AI, Sustainable AI and Efficient AI as three separate concepts, covering techniques including pruning, quantization, knowledge distillation and federated learning
• Efficiency is showing up directly in model architecture choices. DeepSeek V4 Flash scores 50 on the Artificial Analysis Intelligence Index, close to GLM 5.2 and just behind Claude Sonnet 5's 53, while using only around 13 billion active parameters, meaning less compute and less energy per request
• Programming language choice remains a measurable lever. A benchmark study evaluating 27 programming languages found Java among the most energy efficient virtual machine languages tested, despite its JVM overhead reputation
• Carbon aware scheduling is moving from research to publication. A new IEEE published Hybrid Carbon Aware Optimization framework predicts regional carbon intensity to optimize cloud workload placement
• Applied green AI research is extending into infrastructure monitoring, with a new framework proposing energy efficient AI methods specifically for structural health monitoring applications
• Enterprise deployment is also testing efficient models against real workloads, with one platform adding a new model choice specifically to let organizations process complex compliance data securely without the compute overhead of larger frontier models

‍

Cloud, Software and FinOps Efficiency

• The Green Software Foundation has launched GSF Labs, a new space for practical sustainable software experimentation that includes an agentic AI carbon calculator among its first tools
• A Green Software Foundation practitioner course is now training professionals directly in GreenOps concepts and standards
• Gartner predicts that by 2026, organizations combining FinOps and IT asset management will report 60 percent less financial waste from software and cloud investments than those managing the two separately
• A completed eight part series on cloud economics closes by naming a waste problem exceeding $100 billion industry wide and distilling ten core principles for continuous, validated cloud optimization
• New research published in ACM Transactions on the Web on web front end energy consumption found that caching strategy and cookie handling have more impact on a website's energy footprint than most other design choices
• Simple, high impact sustainable IT measures are being reframed as accessible starting points rather than requiring large infrastructure investment, timed to this year's Earth Overshoot Day

‍

Applied Green AI Across Sectors and Regions

• Credit Agricole's DataLab Groupe became the first company of national scale to hold both LNE certification and the RSE Labelia label for its AI development methodology, a standard set in 2023 that now shapes how each new AI project is scoped
• Canal de Isabel II, which manages the full water cycle for around seven million people, is using AI and real time data to reduce water losses and energy use across utility operations
• Africa currently holds under one percent of global data center capacity, with the entire continent running on less than half a gigawatt of active data center white space, a gap being framed as an opportunity to build a decentralized, renewable powered AI infrastructure rather than replicate the hyperscale model
• Green hydrogen production and AI data centers requiring round the clock clean power could add 15 to 20 gigawatts of additional annual solar demand in India starting in fiscal year 2029, pushing the country toward nearly 85 gigawatts of annual solar installations by fiscal year 2030
• A testimonial from Carbon 2C founder Nolwenn Godard describes using an AI and climate community specifically to apply AI capability toward ESG and climate risk analysis rather than treating the two as competing priorities
• Sustainability outcomes in modern data centers increasingly depend on intelligent low voltage systems, including security, fire protection, building management and environmental monitoring, not power and cooling capacity alone

‍

Programs, Community and Events

• The Bundesverband Green Software has opened submissions for its annual Green Software Initiative Award, with entries due October 15 and the winner receiving the trophy on December 10 in Frankfurt
• Climate Collective has launched the second cohort of its AI for Earth: Women in Climate accelerator, following the first cohort's graduation in May
• GCF Academy has built a 40 hour Sustainable AI Value Added Programme that universities can run in a single semester with zero infrastructure cost, requiring only personal laptops and free AI accounts
• Speaking engagements at FOST, apidays and Green IO covered greener software architecture, sustainable API design and responsible generative AI use
• A recap of the Green IO Munich conference highlighted its workshops and networking sessions focused on responsible AI practice
• A panel at Adelaide Climate Action Week stressed that sustainable IT progress depends on direct collaboration between IT, finance and sustainability teams rather than siloed initiatives
• A Green AI Hub panel at the Handelsblatt conference examined responsible AI use specifically within the energy sector

‍

Thanks to Aditya Manglik, Aldrick ZAPPELLINI, Alyson Freeman, Anita Schüttler, Anna Lerner Nesbitt, Anne Currie, Aparna K., Arpit Sharma, Arwel Owen, Ashish Khaira, Aydin Mir Mohammadi, Ben Hillier, Boel Stier, Boris Gamazaychikov, Cal Innes, Cas Burggraaf, Chong Yi Yew, Clifton Reddy, Dalima Negi, David Linthicum, Devesh Sharma, Dhruvi Srishwal, Diana Todea, Dr. Eva Riesenhuber, Dr. Niladri Choudhuri, Dr. Sachin Arun Thanekar, Elias Grass Brandelik, Gabriela Lujan Brollo, Hannah Herbst, Hussain Al Buloshi, James Martin, Janne Kalliola, John Kamara, Jonathan Mattis-Glender, Julie Chenadec, Karen Paul, Karl Koesser‍ and Zap CHENG for contributing insights to this edition. Find the full list of posts and voices on LinkedIn: