Delivering more, with less

We monitor many environmental metrics including energy consumption at different levels (data center, hardware, and software layer), water use, heat recovery volumes, GHG emissions intensity, and material recovery rates for retired equipment. We track core infrastructure metrics like PUE (Power Usage Effectiveness) and WUE (Water Usage Effectiveness). Internally, we track utilization and workload efficiency. Key metrics we are working to define and track consistently are AI output per unit of energy, as well as AI goodput — the share of compute that results in useful output.

Metrics such as GPU-hour usage, utilization, and energy consumption are available in the client console and provide insight into workload performance, including how long it took to complete tasks and how compute resources were used. We are also working to provide customers with on-demand insights into the energy use and carbon footprint associated with their workloads.

We approach energy efficiency as a full-stack engineering challenge. Because we design and operate across hardware, software, and data center infrastructure, we can optimize the system end-to-end. This approach has led to 20% lower server electricity draw compared to off-the-shelf units¹, efficient cooling architectures, and reducing energy spent on non-compute functions (for sites we design and own, 3-4x lower overheads vs a typical data center)². Combined, and factoring in compound gains, these innovations help us avoid considerable unnecessary electricity use: over 120 GWh across our sites in 2025, enough to run one of our smaller facilities 24/7 for a year, or to power approximately 12,500 American homes over the same period.³

Software determines both how efficiently hardware is used and how long workloads run. Our orchestration layer improves workload placement, reduces fragmentation, and maximizes GPU and fabric utilization, cutting idle time. Reliability features like auto-healing prevent failed jobs from restarting from scratch, avoiding wasted computing and energy. Faster time to outcome reduces total energy consumed per task. In addition, our inference platform includes post-training model optimization tools to ensure models run as efficiently as possible on the hardware. Together, this ensures more of the energy used translates into AI output.

Customers benefit from higher utilization and faster job completion⁴, which lowers total compute time and energy use per task. Effective orchestration reduces idle GPU time, failed runs, and provides faster access to computing resources. The combined effect is improved performance per kWh and reduced cost.

We prioritize developing in locations with stable grids and strong transmission infrastructure. We also work closely with local utilities to ensure capacity is available without destabilizing the grid. For example, we are financing critical infrastructure, including substation (switchyard) development near our site in Alabama, which will increase interconnection capacity and improve load balancing.

Yes, Nebius does use renewable energy to power data centers, and assesses access to low-carbon energy from the outset. The share of low-carbon and renewable electricity is disclosed in our sustainability reports. We use instruments like power purchase agreements (PPAs) to source renewable energy, and some of the sites we operate, including in Iceland, France (Paris) and the UK, have 100% of their electricity supply covered by renewable sources.

Data centers we design and own use a combination of air-based cooling and direct liquid cooling, depending on the workload and hardware. Our liquid cooling systems operate in closed loops, meaning no water intake is required for day-to-day operations, and parts of our fleet use air-based cooling that does not rely on chillers or water loops. Cooling systems are designed with integrated heat recovery, allowing us to capture heat as a byproduct of operation. This heat can be reused internally, transferred to district heating networks where infrastructure exists or used for other functions.

Our facilities rely on near zero water use — no water is needed for day-to-day cooling operations at the sites we design and own. Our direct liquid-cooling systems use closed loops that do not rely on water intake, and parts of our fleet rely on air-based cooling that requires neither chillers nor water loops. In practice, this results in very low Water Usage Effectiveness (WUE): at our Finland site, WUE was 0.018 in 2025, meaning just 0.018 liters of water per kWh of IT power, significantly below reported industry benchmarks.

We design cooling systems with heat exchange built in, so if there is local demand, heat recovery can be easily enabled. Our approach is geography-agnostic and not limited to cooler climates. One practical application is capturing data center heat and supplying it to district heating networks where such infrastructure exists. This was implemented over a decade ago by our team in Mäntsälä, Finland, as one of the first of its kind in the region. In the past years the system has been capturing heat associated with 20-30% of the electricity consumed at the site, which is then donated to the local network, displacing other energy sources used for heat production. In 2025, this also contributed to reducing local heating costs by around 10%.

Our geography-agnostic designs enable downstream opportunities to utilize excess heat. Where heat recovery is implemented, part of the heat is reused internally. For example, in Finland, warm air is blended with cooler outside air to achieve the right input temperature for air-based cooling. Heat exchange with the district network runs year-round, as there is constant demand for hot water. Any remaining heat is discharged in line with standard data center practice.

Across our deployments, we design systems for repair and reuse, with components that can be swapped or upgraded without replacing entire servers. Our tool-less hardware and cable-free connections to motherboards are another key component of how we build. Our servers are built to run at higher temperatures, reducing the energy needed for cooling, while reducing failure rates. We also extend hardware lifespans at the workload level, for example, by shifting older generation GPUs from training to inference, enabled by our inference platform.

We prioritize extending hardware lifespans first, as this has the largest impact on reducing waste. When equipment is decommissioned, components are sent through certified recycling channels.

We design our operations to avoid putting pressure on local electricity prices. Sites are selected with available grid capacity and connected at high-voltage levels to avoid impacting low-voltage networks that serve commercial and residential demand. We pay the full cost of our power through contracts and PPAs, so costs are not shifted to residents. We also invest in infrastructure, for example funding a new switchyard in Alabama that improves grid reliability. We also take opportunities to contribute to local energy systems. In Finland, we supply data center heat to district heating networks, which in 2025 helped lower heating costs for households by 10%.

We engage early, listen first, and focus on building long-term local value, which is a differentiator in our approach. Engagement includes working with municipalities and local authorities, organizations, and residents to understand their priorities and turn them into community benefit plans that have concrete actions. We also prioritize clear, transparent communication through channels like open houses, community panels, dedicated websites, and local government updates.

Through our local engagements, we work to understand what actions and resources would best benefit the community. In practice, this can include local hiring and training, education partnerships with schools, universities and technical programs, along with AI literacy support for educators, and access to reskilling programs for professionals. Other opportunities include integrating into local energy systems, for example, by donating recovered data center heat to district networks, which in Finland contributed to household heating costs being 10% lower in 2025.

Access to compute is one of the biggest constraints in AI innovation. Nebius aims to lower that barrier for startups, developers, researchers, and organizations working in critical sectors. In 2025, Nebius awarded $850,000 in cloud credits to startups to help scale production. Our programs and partnerships support practical AI adoption through infrastructure access, technical enablement, academic collaboration, and upskilling. Nebius AI Discovery awards spotlight new technologies, foster collaboration, and propel advancements in healthcare and Nebius Academy provides focused training, expert led courses, and research initiatives to help teams put AI into practice.