Nebius’s Efficient AI Cloud Claims Face a 32-Fold Emissions Increase

Nebius published new efficiency figures after its operational emissions rose 32-fold, creating a sharper conflict behind its recent Google News coverage. The AI cloud provider reported strong power, water, and heat-recovery performance across its expanding data center portfolio. Yet that expansion also pushed its reported Scope 1 and Scope 2 emissions from 2,036 to 65,001 metric tons of carbon dioxide equivalent.

That contrast matters more than any single efficiency statistic. Nebius is arguing that purpose-built infrastructure can deliver more AI computing with less supporting energy and water. Its sustainability record now shows both sides of that claim: better resource efficiency per unit of infrastructure, but a much larger total footprint.

The result puts Nebius against a difficult benchmark, not simply another cloud company. Its real opponent is the arithmetic of rapid AI capacity growth. Better cooling and denser hardware reduce waste, but total environmental pressure can still rise when a provider brings more sites and megawatts online.

The Google News Headline Hides a More Complicated Update

Nebius improved several operating metrics, but its larger footprint overwhelmed those gains at the company-wide emissions level.

Nebius released its 2025 Sustainability Report on July 30, 2026. The report covers a year when the company expanded from a relatively concentrated European footprint to seven active data center sites.

The company reported a portfolio-wide power usage effectiveness, or PUE, of 1.25. PUE compares all electricity entering a data center with the electricity used directly by computing equipment. A result closer to 1 means less power goes toward cooling, lighting, conversion losses, and other supporting systems.

Nebius compared its result with a global industry average of 1.54. The comparison suggests that its facilities directed a larger share of incoming power toward servers and networking equipment.

According to the company’s efficiency disclosure, its hardware and facility designs avoided an estimated 102 GWh of electricity consumption during 2025. Nebius equated that amount with the annual electricity use of 9,800 American households.

That estimate needs careful framing. It describes modeled consumption avoided against an alternative design, not a decline in Nebius’s total electricity use. A rapidly expanding provider can avoid energy relative to a baseline while still consuming more energy overall.

The same distinction applies to water. One company-owned facility recorded water usage effectiveness of 0.018 liters per kilowatt-hour at the end of 2025. Water usage effectiveness measures on-site water consumption against the energy used by computing equipment.

Nebius compared that figure with a projected United States data center average of 0.45 to 0.49 liters per kilowatt-hour. It attributed the result to air cooling and closed-loop liquid cooling, which recirculates coolant instead of continuously withdrawing water.

Nebius also projected that similar designs might avoid about 3.3 billion liters of water intake for every gigawatt of portfolio capacity. However, the company explicitly described that number as a design-based projection. It was not a measured operational saving across a gigawatt of deployed infrastructure.

Those qualifications do not make the engineering irrelevant. They define what the numbers can support. The report provides evidence that Nebius operates some efficient facilities, but it does not show that expansion produces a smaller total environmental footprint.

That is the essential context missing from a simple Google News headline. Nebius has presented a credible operating-efficiency story. It has not yet resolved the wider question of whether efficiency improvements can keep pace with its growth.

Efficient AI Cloud Design Starts With Heat and Cooling

Nebius treats cooling, server design, and heat recovery as one system rather than separate sustainability projects.

AI servers convert most of their electricity into heat. A data center must remove that heat while keeping processors within safe operating temperatures. Every fan, pump, chiller, and cooling tower adds overhead beyond the energy used for computing.

Nebius says it designs servers, racks, cooling systems, and parts of the data center together. That vertical approach gives engineers more control over airflow, coolant routing, operating temperatures, and equipment density.

At its Finland facility, the company can operate servers at temperatures reaching 45 degrees Celsius. This wider thermal range allows the site to use outside air for cooling during more of the year.

The method is known as free cooling, meaning ambient conditions replace or reduce mechanical refrigeration. Finland’s climate makes that approach especially useful, though Nebius describes the underlying system as adaptable to other regions.

High-density AI equipment also requires liquid cooling in many configurations. Nebius uses closed-loop systems, where coolant captures heat near the processors and circulates through a contained loop. This approach can reduce direct water withdrawals compared with evaporative systems.

The distinction matters because liquid cooling is not automatically water efficient. Some designs still reject heat through cooling towers that consume water. A closed loop reduces water demand inside the computing system, but the complete site design determines the final result.

Nebius’s approach therefore combines two methods. Air cooling handles suitable equipment and conditions, while closed-loop liquid cooling serves dense accelerator systems. The company says this combination helps it minimize both electrical overhead and water intake.

Independent reporting supports several parts of that account. A neocloud analysis found that Nebius averaged a PUE of 1.3 in 2024. Its Finland facility reached 1.15 during its strongest months in 2025, according to sustainability chief Daria Mukhortova.

The same reporting notes an important limit. PUE measures how much electricity reaches computing equipment, but not how much useful AI work that equipment completes. A data center can post an excellent PUE while running underused processors or inefficient models.

Useful-output measures might include generated tokens per megawatt, completed training runs, or productive accelerator hours. No common reporting standard currently lets buyers compare providers across all those dimensions.

Reliability also affects efficiency. Failed equipment or interrupted training jobs can force customers to repeat computation. That extra work consumes energy without producing additional useful output.

Nebius says its hardware and software integration helps reduce such waste. However, customers need workload-level data before they can separate facility efficiency from actual application efficiency.

This mechanism creates a commercial advantage as well as an environmental one. Electricity that does not power cooling can power revenue-producing GPUs instead. Lower cooling overhead can also improve the economics of a fixed-capacity electrical connection.

That alignment explains why AI cloud providers promote efficient infrastructure even when customers do not demand sustainability reporting. Energy efficiency supports capacity, cost control, and operational reliability at the same time.

The harder question is whether those advantages survive outside favorable Nordic locations. Nebius is expanding in the United States, where energy sources, weather, grid constraints, and community infrastructure differ significantly.

Heat Reuse Makes Finland the Strongest Case

Nebius’s Finland facility turns waste heat into a local product, but that model depends on infrastructure that many markets lack.

The Mäntsälä data center supplies recovered heat to the local district heating network. District heating distributes centrally produced hot water through pipes to nearby homes and buildings.

Nebius exported 19.5 GWh of recovered heat during 2025. The company says the system is designed to export more than 250 GWh annually when operating at full capacity.

That wide gap reflects both future potential and the limits of current utilization. The larger figure is a designed maximum, not the amount delivered during the reporting year.

The project has still produced a measurable community benefit. Nebius says recovered data center heat reduced heating costs for connected households by about 10 percent.

Earlier operating data provide a longer view. More than 50 GWh of heat was reused in Mäntsälä between 2022 and 2024. Independent reporting said that amount was sufficient to heat 2,500 homes during the measured period.

Heat recovery improves the sustainability case in two ways. It displaces part of the energy that a district network would otherwise obtain elsewhere. It can also reduce the amount of electricity the data center needs to remove heat.

However, the model is difficult to reproduce everywhere. A data center needs a nearby heat customer, compatible temperatures, distribution pipes, and demand that overlaps with computing operations.

Mukhortova acknowledged this constraint in industry reporting. Nebius can design sites with heat-exchange potential, but American communities often lack district heating networks. A technically recoverable resource has little value when no local system can accept it.

This makes Mäntsälä both an example and a warning. It shows how data center infrastructure can integrate with a community energy system. It also shows why one successful site cannot establish a global operating pattern.

Location therefore becomes part of the product. Nordic countries offer cool climates and electricity systems with high shares of renewable or low-carbon generation. Many also have established district heating networks.

Nebius reported sourcing 100 percent renewable electricity in Iceland, France, and the United Kingdom during 2025. It also said Finland was moving from 95 percent low-carbon electricity to 100 percent renewable sourcing.

Its Iceland deployment illustrates the location advantage. Nebius announced a 10 MW cluster in Keflavik using electricity from Iceland’s hydroelectric and geothermal resources. The company introduced that site alongside a much larger American expansion.

The regional expansion included a New Jersey facility designed to scale to 300 MW. That contrast reveals the central strategic problem.

Nordic capacity supports a strong sustainability narrative, but customer demand also pulls AI infrastructure toward the United States. The most attractive markets may not offer the same electricity mix, climate, or heat-reuse opportunities.

Competitors face the same conflict. CoreWeave operates in Sweden, while Nscale and NexGen Cloud have pursued Nordic deployments. Those locations give neocloud providers access to lower-carbon electricity and natural cooling.

The opportunity is not unlimited. AI training requires large, dependable blocks of power, and suitable grid connections are scarce. Providers often choose from available sites rather than waiting for an ideal sustainability profile.

Nebius must therefore prove that the engineering behind Finland travels well. The company can standardize racks, cooling loops, and software. It cannot duplicate Finland’s weather, grid, and district heating network at every location.

Emissions Growth Tests Nebius’s Efficiency Argument

Nebius’s central sustainability risk is absolute growth, not a failure to improve individual facilities.

Nebius reported 65,001 metric tons of market-based Scope 1 and Scope 2 carbon dioxide equivalent emissions for 2025. Scope 1 covers emissions from sources a company controls directly. Scope 2 covers emissions associated with purchased electricity, heating, and cooling.

The comparable 2024 figure was 2,036 metric tons. That means reported operational emissions increased by roughly 32 times in one year.

The increase followed the expansion of Nebius’s operational base to seven active sites. Data center operations accounted for 99 percent of its market-based Scope 1 and Scope 2 emissions in 2025.

These figures do not contradict the reported PUE result. A company can operate each unit more efficiently while adding enough units to increase total consumption and emissions.

This is a version of the rebound problem. Efficiency lowers the resource requirement for a given amount of output. Lower costs and greater availability can then encourage much more output, raising total demand.

For AI infrastructure, demand is already growing quickly. More efficient cooling lets providers place additional accelerators behind the same electrical connection. That creates more computing capacity, but it does not guarantee lower electricity consumption.

The emissions increase also exposes gaps in the available account. Nebius did not report Scope 3 emissions for 2025. Scope 3 includes indirect effects from server manufacturing, construction materials, purchased equipment, transportation, and other supply-chain activities.

The company said its reporting boundaries and supplier data systems were still being developed. It concluded that the available information would not support a representative Scope 3 inventory.

That caution is preferable to publishing a weak estimate. Still, the omission leaves a substantial part of the footprint unmeasured during an infrastructure construction cycle.

GPU manufacturing, concrete, steel, backup systems, and new electrical equipment all carry embedded emissions. A full-stack provider expanding across several regions cannot establish its total climate effect using operational electricity data alone.

Nebius had also not published quantified reduction targets for emissions and resource use when the report appeared. It said those targets were being developed alongside its data systems and governance processes.

An independent emissions review highlighted both gaps. It also examined Nebius’s planned use of behind-the-meter fuel cells in the United States.

Behind-the-meter generation supplies a facility directly instead of relying entirely on the public grid. Nebius agreed to deploy 328 MW of solid oxide fuel cell capacity at a United States site.

Fuel cells can reduce certain local pollutants and use less water than some combustion-based generation. They can also provide steady electricity without waiting for a large grid connection.

However, environmental performance depends on the fuel source. Natural-gas-powered fuel cells still create greenhouse gas emissions, even without conventional combustion at the facility.

Nebius’s public sustainability description emphasizes lower emissions options rather than claiming zero-carbon American operations. That wording is important. Lower emissions relative to a planned alternative do not equal renewable or emissions-free power.

The conflict will become more visible as the company adds American capacity. The United States offers large customers and urgent demand, but grid carbon intensity varies widely by location and time.

Nebius also faces a reporting-boundary challenge. Its 1.25 PUE figure is an IT-load-weighted portfolio average. New facilities, construction activity, partial utilization, and different ownership models can all affect future comparisons.

A stable or improving average would support the case that its designs scale. A deterioration would suggest that early efficiency depended too heavily on a small number of favorable facilities.

Google News readers should therefore treat the 32-fold increase as a scale signal, not proof that the engineering failed. It shows why operational efficiency cannot serve as a substitute for absolute emissions reporting.

The Real Opponent Is Capacity Growth

Nebius must show that its efficiency system can constrain environmental impact while it expands toward gigawatt-scale operations.

The neocloud market exists because AI developers need dense GPU capacity faster than traditional infrastructure planning often supplies it. Providers such as Nebius, CoreWeave, Crusoe, Lambda, and Nscale focus on accelerator-heavy workloads and specialized AI services.

This model creates pressure to secure land, chips, transformers, power contracts, and cooling equipment quickly. Every delay can leave expensive processors idle or send customers to another provider.

Sustainability goals operate on a different timeline. Renewable generation, grid upgrades, transmission lines, community consultation, and heat networks can take years to develop.

Nebius tries to narrow that gap through integrated design. Standardized racks and cooling systems can shorten deployment time. High operating temperatures can increase the hours available for free cooling. Software optimization can improve hardware utilization.

The strategy makes sustainability part of infrastructure economics. A lower PUE allows more of a site’s power allocation to reach GPUs. Closed-loop cooling limits exposure to local water constraints. Heat reuse can strengthen community acceptance where a suitable network exists.

Yet growth can outrun each benefit. Nebius’s Finland operation provides its strongest evidence, but the company’s future footprint will be more geographically diverse.

The New Jersey project alone was designed for up to 300 MW. That is 30 times the capacity of the company’s announced 10 MW Iceland cluster.

This does not mean New Jersey will carry 30 times the environmental impact. It does show why location-specific performance will determine the credibility of portfolio-wide claims.

Customers also influence the outcome. AI cloud buyers usually prioritize available accelerators, training speed, reliability, data residency, and cost. Sustainability can become secondary when suitable capacity is scarce.

Industry executives have described large-scale power access as a seller’s market. Providers often take viable sites when they become available, even if the local electricity mix is less favorable.

Nebius’s response is to seek lower-emissions generation and build efficient facilities wherever it operates. The approach is pragmatic, but it creates a weaker claim than a portfolio powered entirely by additional renewable energy.

Additionality means a company’s procurement helps create new clean generation rather than claiming electricity from existing resources through contractual instruments. Nebius’s current public figures do not provide a uniform additionality measure across its portfolio.

Buyers need more granular information. Annual renewable sourcing percentages can hide hourly gaps between data center demand and clean generation.

A facility might contract enough renewable electricity to match yearly consumption while drawing fossil-heavy grid power during certain hours. Hourly carbon reporting would show that variation more clearly.

Workload-level reporting would add another layer. Training a model in a low-carbon region may reduce emissions, but latency and data rules can require inference close to users.

Nebius could eventually help customers place flexible workloads according to electricity availability and carbon intensity. That would turn sustainability data into a scheduling input rather than an annual report item.

No such system resolves unlimited expansion. It would simply give customers and operators better tools to manage the tradeoff.

The strongest version of Nebius’s argument is therefore comparative. Purpose-built AI infrastructure can waste less electricity and water than a less integrated design delivering the same computing output.

The weaker version claims that efficient infrastructure makes rapid AI expansion sustainable by itself. The reported emissions increase shows why that conclusion is not yet supported.

Three Signals Will Decide Whether the Model Scales

The next test is whether Nebius publishes consistent site-level results, complete emissions data, and measurable reduction targets while capacity keeps rising.

The first signal is portfolio performance across new American sites. Nebius reported a 1.25 average PUE for 2025, with especially strong results in Finland.

Future reporting should show whether that average holds as New Jersey and other United States deployments contribute more operating hours. Site-level PUE and water figures would help readers distinguish engineering performance from geographic advantages.

An average near the current level would strengthen Nebius’s claim that its integrated design travels well. A material increase would suggest that rapid deployment or regional conditions impose efficiency costs.

Water performance deserves the same scrutiny. The 0.018-liter figure came from one owned facility at the end of 2025. Nebius should disclose annual data across more locations before presenting that result as a portfolio characteristic.

The second signal is a complete Scope 3 inventory. Operational emissions matter, but construction and hardware manufacturing are central to a fast-growing AI cloud.

Nebius needs supplier information, consistent organizational boundaries, and asset-level data to produce a representative inventory. Publishing that total will probably increase the company’s disclosed footprint substantially.

A larger number would not automatically indicate worse management. It would produce a more useful baseline for setting reduction priorities and tracking progress.

The third signal is the release of quantified environmental targets. Nebius has reported operating metrics and design projections, but it has not yet tied them to time-bound reductions.

Credible targets would specify the metric, reporting boundary, base year, deadline, and treatment of expansion. They should separate absolute emissions from intensity measures.

Both types are necessary. An intensity target shows whether each unit of computing becomes cleaner. An absolute target shows whether total climate impact is moving in the intended direction.

Nebius should also clarify how fuel cells, renewable contracts, and grid electricity fit into those targets. A lower-carbon natural gas system should not be grouped with zero-carbon electricity without a clear accounting method.

These signals matter to developers and enterprise buyers because cloud decisions transfer environmental effects rather than removing them. A customer may not own the data center, but its workloads still drive infrastructure demand.

Teams evaluating AI providers should ask for location-specific energy, water, and carbon information. They should also compare useful output, reliability, and utilization, not PUE alone.

The broader Google News story is not that Nebius discovered sustainable AI cloud computing. It is that the company published enough data to reveal a real tension between efficient design and extraordinary growth.

Nebius has shown credible mechanisms for reducing supporting energy, limiting water intake, and reusing heat. Its next reports must show whether those mechanisms constrain total impact across a far larger network.

Watch the next portfolio PUE, the first representative Scope 3 inventory, and the arrival of quantified reduction targets. Together, those results will show whether Nebius built a transferable efficiency system or a compelling Nordic exception.