Driving more efficient and sustainable AI
Artificial intelligence, spanning machine learning, generative AI and agentic AI offers powerful opportunities to increase productivity and accelerate progress on global challenges. At the same time, its growth brings increased attention to its energy use, greenhouse gas emissions, water consumption, and the material intensity of advanced computing systems. IBM is addressing these considerations through a full‑stack approach that embeds efficiency and sustainability into every layer of AI technology.
AI’s footprint is shaped by interconnected factors: model architecture, training methods, workload orchestration, hardware performance and the materials used to manufacture chips. Because these elements are interdependent, no single intervention is sufficient. That is why we are designing for AI systems holistically, applying various efficiency techniques and our own innovations to minimize unnecessary compute and make the most effective use of infrastructure. Our open, high-performing and trusted Granite models use hybrid architectures and mixture of experts techniques to deliver high accuracy with significantly lower memory requirements. Rightsized models enable organizations to select the smallest model capable of meeting the performance requirements of each task.
Deploying rightsized models on open, scalable AI platforms allows workloads to run on the most appropriate infrastructure, whether on premises or in the cloud. Through contributions to the open-source ecosystem and the delivery of enterprise AI platforms such as Red Hat OpenShift AI and Red Hat AI Inference, we help advance more efficient training and inference at scale. At the infrastructure layer, systems including IBM Power11, LinuxONE 5 and the IBM z17, powered by the Telum II processor and innovative cooling and routing technologies, provide high AI throughput at lower power. When combined with the Spyre accelerator, these systems further extend performance gains while improving energy efficiency for demanding AI workloads.
By optimizing models, platforms and hardware together, we improve how efficiently intelligence is delivered, achieving more throughput per system and unit of energy. This full-stack efficiency enables AI systems to scale while supporting environmental management and long-term operational resilience.
Driving energy conservation and emissions reductions in IBM’s operations
IBM’s greenhouse gas reduction goals are based on scientific benchmarks established by the United Nations Intergovernmental Panel on Climate Change (IPCC). In 2023, we achieved our goal to reduce operational GHG emissions by 65% by 2025 (from 2010, adjusted for acquisitions and divestitures).
We are now working towards achieving our goal to reach net-zero operational GHG emissions by 2030. Our net-zero ambition is supported by initiatives to conserve energy and increase our procurement of renewable electricity.
Energy consumption
In 2025, IBM decreased total energy consumption by 8.0% versus 2024, driven by increased operational efficiencies and a continued focus on energy conservation.
Energy conservation
IBM maintains a global energy conservation program focused on reducing energy consumption through ongoing efficiency programs. In 2025 we implemented 634 energy conservation projects avoiding an estimated 123,000 MWh of energy consumption and 35,000 mtCO2e. The primary driver of energy savings was IT equipment upgrades across our data centers.
Data center energy efficiency
To reduce the environmental footprint of IBM data centers, we focus on improving energy performance through space optimization and continued modernization of energy-efficient infrastructure, increasing compute delivered per unit of energy consumed.
We track performance versus our data center cooling efficiency goal by measuring the Power Usage Effectiveness (PUE) of our data centers. In 2025, the estimated weighted average PUE of our data centers was 1.39, representing a 28.0% improvement in cooling efficiency compared with our 2019 baseline. This performance surpasses our goal to achieve a 20% improvement in data center cooling efficiency by 2025 compared to our 2019 baseline.
In parallel with efficiency improvements, we continue to increase the share of renewable electricity powering our data centers. Overall, 85% of the electricity consumed in our data centers came from renewable sources. Globally, 44 data centers were supplied with 100% renewable electricity in 2025, reflecting our ongoing focus on lowering operational emissions through cleaner energy procurement.
Renewable electricity
In 2025, 84.5% of IBM’s total electricity consumption came from renewable sources, exceeding our 2025 renewable electricity procurement goal. Our renewable electricity purchases including 73.4% contracted directly from power suppliers or obtained via landlords and 11.1% already in the electricity mix we received from the grid. 1
GHG emissions inventory
IBM calculates greenhouse gas (GHG) emissions according to The Greenhouse Gas Protocol Corporate Accounting and Reporting Standard. Information on our 2025 emissions inventory can be found on our Data and Policies page.
Assurance
IBM engaged an external third party to perform an attest review engagement for certain environmental performance metrics disclosed on our website as of December 31, 2025 and for the year then ended. The external third-party report is available at the Greenhouse Gas Independent Limited Assurance Statement.
Product energy efficiency
In line with IBM’s goal to reduce power consumption per unit of delivered work for each new generation of enterprise server and storage products, we introduced the IBM z17 and IBM LinuxONE Emperor 5 systems in 2025. Both systems were developed to deliver performance and security while optimizing energy use.
- The largest IBM z17 model (Max208) improves maximum system capacity per kilowatt by approximately 22% compared to the largest IBM z16 model (Max200) under a worst-case scenario.1
- The Telum II data processing unit reduces power required for I/O management in a large IBM z17 system by over 90% compared to a similarly configured IBM z16.2
- For cloud-native, containerized workloads, IBM LinuxONE Emperor 5 can reduce power consumption by up to 65% compared to a comparable x86 solution.3
We have also expanded sustainability‑related monitoring capabilities within the Hardware Management Console and enhanced configuration‑specific product carbon footprint estimation tools to support customer decision‑making.
To learn more about IBM Z and IBM LinuxONE sustainability, see
https://ibm.com/products/z/sustainability, https://ibm.com/products/linuxone/sustainability and
https://ibm.com/support/z-content-solutions/sustainability-linuxone.
Advancing sustainable infrastructure with IBM Power11
In July 2025, IBM launched the Power11 server, introducing significant advancements in data center energy efficiency. Engineered to decouple performance from power consumption, the Power11 delivers up to 37% more rPerf per Watt4 compared to the Power10.
A key feature of this generation is the new "Energy Efficient Mode," which allows clients to reduce power usage on demand, achieving up to 28%5 better server efficiency compared to the system’s maximum performance mode. These innovations enable enterprises to scale their mission-critical and AI workloads while strictly managing their energy footprint and operational costs.
ENERGY STAR
IBM maintains a foundational partnership with the U.S. EPA’s ENERGY STAR program. We continue to prioritize third-party validation of our system’s efficiency; in 2025, our portfolio included 9 certified enterprise servers and 7 certified storage products, demonstrating continued adherence to the program’s rigorous environmental specifications.
Footnotes
Our renewable electricity reporting counts only generation from the same grid regions where our consumption occurs, as defined by U.S. Energy Information Administration power balancing authority territories (or equivalent in other jurisdictions). We do not count unbundled renewable energy certificates from other grid regions. Reported renewable electricity includes wind, hydropower (including large hydro), biomass, solar, and geothermal, and reflects all contracted purchases regardless of vintage or “additionality.” Due to timing differences between generation and use, electricity delivered at any moment may include non‑renewable sources, even though total renewable procurement equals the amount claimed.
Note 1. System capacity based on LSPR data available at ibm.com/support/pages/ibm-z-large-systems-performance-reference. Power consumption is computed using the Power Estimation Tool for 3931 and Power Estimation Tool for 9175, available at ibm.com/support/resourcelink/api/content/public/PowerEstimationTool-legacy.html. Uses worst-case power conditions with the absolute maximum system power configuration at the maximum utilization and for the system environment driven maximum power condition. Results may vary.
Note 2. Comparison based on IBM lab measurements for the difference in power required for supporting I/O for FICON and OSA in an expected large IBM Machine Type 9175 configuration based on an actual historical large IBM Machine Type 3931 configuration. IBM Machine Type 9175 is Max 208 with 23 TB memory, 56 active processors, 3 IBM Virtual Flash Memory, 14 ICA-SR 2.0, 7 PCIe+ I/O drawers with 69 FICON Express32 – 4P LX, 12 OSA-Express7S 1.2 GbE SX, 18 Network Express LR 10G, and 4 Crypto Express 8S (2 HSMs). The IBM Machine Type 3931 is configured to provide the same hardware capability. Results may vary.
Note 3. IBM internal performance tests for the core consolidation study compared an IBM Machine Type 9175 Max136 with 136 configurable processor units with an x86 solution that used a commercially available enterprise server with two 5th gen Intel Xeon Platinum 8592+ processors and 64 cores per CPU. Workloads consisted of a containerized OLTP WebSphere Liberty v25 application running on Red Hat OCP v4.17 and an EDB Postgres for Kubernetes v1.25 on the same OCP cluster. Both solutions used Red Hat Enterprise Linux v9.5 and KVM. Test results were extrapolated to a typical, complete customer IT solution that included production and non-production IT environments isolated from each other. The IBM Machine Type 9175 solution required one Max136 and the x86 solution required 23 compared servers. Results may vary.
Note 4. Based on maximum configuration at 100% utilization under typical operating conditions across all Power11 systems.
Note 5. Based upon IBM measurements of performance per watt on servers comparing Maximum Performance Mode to Energy-Efficient Mode while running compute-, disk-, and memory-based workloads on Power11 systems with fully configured sockets and memory as follows: E1180 with 4x10c / 64x64GB DDIMM, E1150 with 4x16c / 64x32GB DDIMM, S1124 with 2x16c / 32x32GB DDIMM, S1122 with 2x16c / 32x32GB DDIMM.