Data center sustainability has a signal-to-noise problem

The debate around AI data centers and sustainability has a signal-to-noise problem. Earlier this year, Carbon Direct hosted a webinar to cut through the noise.

The conversation explored the complex intersection of accelerating AI demand, constrained grid capacity, evolving policy environments, community dynamics, and rapidly advancing technology.

The ICEF 2025 Sustainable Data Centers Roadmap, considered one of the most comprehensive analyses of data center sustainability published to date, served as the basis for conversation among several of the top academic and industry experts in the space.

The panel included David Sandalow, Inaugural Fellow at Columbia University's Center on Global Energy Policy and Co-Director of the Roadmap Project; Julio Friedmann, PhD, Chief Scientist at Carbon Direct and Roadmap Co-Author; Colin McCormick, PhD, Chief Innovation Officer at Carbon Direct and Roadmap Co-Author; and Sara Neff, Partner and General Manager for Sustainability at Microsoft. Zara Ahmed, DrPH, Chief Operating Officer at Carbon Direct, moderated.

Here is what the panel found.

Construction decisions lock in sustainability outcomes

Data center energy consumption is expanding at a pace the grid has not seen in decades. Global electricity demand from data centers doubled between 2017 and 2024, from roughly 200 TWh to more than 400 TWh. The ICEF roadmap projects that data centers could account for 1.8–3.4% of global electricity use by 2030, up from roughly 1.5% in 2024.

Over 86% of global data center energy use is concentrated in the US, China, and Europe, and facilities are breaking ground across all three markets right now. In the US, home to roughly 45% of global data center capacity, that figure could reach 7–12% of total power demand by 2028. In some regions, concentration is already extreme: data centers accounted for 22% of Ireland's national electricity use and 25% of total electricity demand in Northern Virginia in 2024. This growth has created backlash from communities, and has prompted governors, utilities, and hyperscalers to reconsider new structural and commercial agreements.

The decisions operators make during construction, on siting, energy procurement, and cooling systems, lock in the environmental performance of those assets for their entire operational lives.

As Sandalow emphasized, "The stakes of this moment are high, and the timeline is short." With the current US administration placing low priority on environmental standards, he noted that action at the operator level and in state jurisdictions becomes all the more critical. The ICEF roadmap calls on operators to integrate energy and environmental strategies centrally into planning and operations, and on governments to require mandatory disclosure and minimum environmental standards.

Power and emissions are not the same thing

As Dr. Friedmann noted during the webinar, "power does not equal emissions." A data center's electricity emissions depend heavily on the carbon intensity of its grid. According to the ICEF roadmap, data centers on the lowest-carbon grids can have scope 2 emissions almost 100 times smaller than those on the highest-carbon grids. Design choices and operational management, from cooling technology to workload scheduling, shape a facility's greenhouse gas footprint just as meaningfully.

Leading hyperscalers already understand this, and many practice it. Some (e.g., Google, Meta, Amazon, Microsoft) have become the world's largest procurers of renewable energy, using long-term power purchase agreements for solar and wind and, increasingly, contracts for small modular reactors (SMRs) and geothermal energy. Some are also using natural gas with carbon capture and storage (CCS), or procuring natural gas with unusually low fugitive methane emissions.

Embodied emissions, those generated during construction and IT equipment manufacture, are harder to see but increasingly consequential. In data centers powered predominantly by low-carbon electricity, these embodied greenhouse gas emissions, classified as scope 3, can exceed 40% of a facility's total lifetime greenhouse gas footprint, according to analysis by Dr. Friedmann and Dr. McCormick. The largest single source is chip manufacturing, which relies on fluorinated gases between 100 and 24,000 times more potent than CO2.

Dr. Friedmann explains that in many new facilities, embodied emissions represent a significant portion of lifetime emissions, making early procurement decisions as consequential as operational ones.

Data center siting shapes climate and financial risk

Smart data center siting reduces a facility's energy consumption, water footprint, and greenhouse gas emissions simultaneously. The ICEF roadmap recommends that operators prioritize locations with low-carbon power, low water stress, and favorable cooling conditions. Where conditions allow, waste heat from operations can be channeled into district heating systems, turning an unavoidable byproduct into a community resource.

Community engagement is essential

Local opposition to data center development is rising and increasingly costly. In the past two years, local opposition has resulted in the cancellation of more than $170 billion in announced AI data center investment. As Dr. McCormick highlighted, "the number of localities opposing data centers has increased substantially over the last year."

Communities are raising legitimate questions about electricity prices, water availability, noise, and grid strain.

The ICEF roadmap recommends that operators engage collaboratively with communities throughout the project lifecycle, from site selection to post-construction operations. It also recommends that governments fast-track approvals for well-located, well-designed, and well-managed projects, making good environmental decisions a competitive advantage.

Water risk is local and concentrated

On a global basis, data center water use accounts for less than 0.008% of freshwater withdrawals. Agriculture consumes roughly 12,000 times more. To put it another way: the water used to produce one hamburger is roughly equivalent to the water needed for 19,000 ChatGPT queries. In water-stressed regions, however, even a small global footprint can translate into significant local impact. That local risk is concentrated where data centers are actually being built. Two-thirds of data centers built or in development in the US since 2022 are in areas of high water stress, according to the ICEF roadmap.

The shift to liquid cooling for data centers is directly relevant here. Roughly 95% of facilities still use air-based cooling, which relies on evaporation. But graphics processing unit (GPU)-heavy AI workloads are pushing rack densities past the approximately 30 kilowatt (kW) threshold, where air cooling begins to fail. Liquid cooling systems handle racks exceeding 100 kW while using substantially less water. As Dr. McCormick explained, liquid cooling is "driving even more efficiency and, paradoxically, reducing water consumption" as energy densities increase.

Microsoft has already demonstrated what is possible at scale. Neff was direct: "All of our data centers since August of 2024 have been zero water." That benchmark sets a high bar for the industry. She is equally clear on the transparency imperative: water is deeply personal to host communities, and operators have a responsibility to proactively engage, not just report annually. Recently, Microsoft’s CEO Satya Nadella stated that its new data center consumes the same amount of water as one restaurant.

A key challenge is standardization and relevance of water data and metrics. Water Usage Effectiveness (WUE), the most commonly used metric, captures direct on-site consumption but misses indirect use from power generation. In coal- or nuclear-heavy grids, that indirect use can exceed on-site consumption entirely.

Data gaps undermine progress

Despite years of growing scrutiny, data center environmental disclosure remains fragmented, inconsistent, and difficult to compare across operators, regions, and building vintages. The ICEF roadmap is explicit: facility-level disclosures provide significantly more benefit than company-level disclosures.

Greenhouse gas emissions data are uneven across all three scopes. Scope 3 sources, including embodied emissions from construction materials and chip manufacturing, are particularly underreported. Water data present similar challenges: WUE is often reported but captures only direct on-site consumption.

The data gap has direct commercial consequences. As Neff put it, customer expectations have moved well beyond annual sustainability reporting: "What they want is more granular. What about my grid? What about my community? What are the exact emissions of my workloads?"

That pressure will only intensify. Agentic AI systems consume up to 100 times more compute than traditional chatbot interactions, according to the roadmap, meaning workload-level emissions accountability is shifting from a differentiating capability to a baseline expectation.

Without reliable, standardized information, communities cannot make informed decisions, regulators cannot set meaningful standards, and operators cannot credibly demonstrate progress against their own commitments.

Getting these decisions right is imperative

The ICEF roadmap offers something rare: rigorous, independent analysis of what actually drives data center sustainability. The key conclusion is clear. Decisions made before a shovel hits the ground, on where to build, how to power it, and how to engage communities, determine environmental outcomes for the life of the asset.

Neff put the collective challenge plainly: the solutions exist, and the technology is moving fast, but no single operator can shift the market alone. Broader adoption of low-carbon equipment, transparent reporting, and coordinated advocacy across the industry are what will move the needle.

Those efforts also need to account for the emissions already being locked in. The carbon removal capacity needed to offset decades of new AI infrastructure emissions is not growing fast enough to close that gap, and the cost of addressing it will likely only increase over time.

The scientists and advisors at Carbon Direct work with data center operators, policymakers, and investors on exactly these decisions, from power system modeling and grid strategy to scope 2 and scope 3 carbon accounting, load flexibility and grid impacts analysis, CCS pathway evaluation, and community impact assessment.

Explore our data center work to learn more.