Making AI Sustainable? From Frugal Models to Responsible Governance

Is the rapid expansion of artificial intelligence compatible with environmental and social concerns? With major model providers reporting that their active user base had tripled and their revenue had increased fivefold over the past year, the question was high on the agenda at the 2026 Natixis CIB Green Summit, featuring several panel discussions.

This surge in AI adoption is putting growing pressure on resources and raises broader societal concerns. Starting with energy: the IEA reports that data-center electricity consumption rose by 17% in 2025 - and by 50% among AI-focused hyperscalers. It projects that consumption will nearly double by 2030, reaching around 3% of global electricity demand.

Water use and land footprint also raises challenges. By 2030, AI-related water consumption could equal the annual basic domestic water needs of 1.3 billion people, while its land footprint could exceed 14,500 km² - roughly twice the size of the Jakarta metropolitan area.

The debate also extends to public acceptance, amid concerns about the location of data centers, the risk of losing control over AI systems, and the potential job displacement.

Against this backdrop, can AI and sustainability be compatible - and under what conditions? How can the infrastructure supporting AI’s expansion be developed in a responsible manner, involving environmental limits and local communities? Can good governance keep AI aligned with sustainability and societal interests? And can AI itself also be part of the solution?

At the Green Summit, we addressed these questions from a variety of perspectives. We looked at the infrastructure itself: the environmental and social concerns of data center developers and investors (“Data Center Financing: Screening for Sustainability Issues”) and how they can secure local population acceptance (“License to Operate & Business Impacts: New Learning from the Data Center Industry?”). We took a step back: what is this infrastructure being developed for? (“In the Spotlight: Can We Build a Frugal and Sustainable AI?”) How is AI impacting our societies, through our jobs? (“AI and the Future of Work: Modern Times Revisited”) and how do we ensure that we don’t lose control over AI? (“Living in Science Fiction: How will AI Impact our Organizations and Governance Models?”)

Georgia Dunne opened the discussion by highlighting the need to balance sustainability commitments with meeting delivery timelines to avoid penalties. “Grid connection is a major development bottleneck. We have a strong preference for renewables, but procurement depends on location”.

She illustrated the impacts of climate change on the operation of a data center, highlighting the need for proper physical risk assessment: rising temperatures leading to additional cooling requirements, cooling failures leading to operational disruption, and flooding leading to damage and business interruption.

Location also determines which solutions can help reduce a data center’s environmental footprint. Recovering waste heat for district heating, for example, can be challenging if there is no nearby user for the heat.

She also outlined solutions developed by STACK, including sourcing renewable electricity through green power purchase agreements (PPAs) with utilities while retaining backup generators to ensure uptime. At STACK reduced generator-testing frequency also helps cut related emissions by 46%.

Another option, developed by STACK’s US subsidiary, is to adapt natural-gas fuel-cells to run on hydrogen or biogas, as the technology becomes economically viable.

Birgit Lundem Jakobsen opened the discussion from the investor perspective, observing that while data centres were traditionally seen as liquid, safe-haven assets backed by strong tenants, investor appetite is shifting. The sheer volume of financing needs, alongside concerns around residual values and technology obsolescence, has contributed to wider spreads and greater differentiation between transactions.

Beyond tenant credit quality, investors are increasingly focused on the lease, including its length and termination provisions, as well as site-level exposure to flooding, extreme heat, water stress and power disruption, and the availability of adequate insurance. “Physical and environmental risks are considered as part of credit risk.”. She concluded: “Investor analysis has become more forensic.”

This is driving greater selectivity. Samiya Jmili explained how her team focuses on top-tier, well-connected, high-quality European assets with quality plans to improve their environmental performance. Climate risk tools are also key to the preliminary screening of projects, assessing current and future exposure and quantifying expected losses, considering measures from the design stage. Conscious of the shortcomings of these tools, they can be complemented by on-site expertise.

Another focus of sustainability analysis is the decarbonization of data-center assets, in particular of their energy supply. “We are comfortable holding an asset in our portfolio temporarily, provided it has a credible transition plan that preserves its potential to decarbonize and avoids locking in carbon-intensive activities”, explained Samiya Jmili. “For this, we rely on an NZIF-aligned framework”.

Both panelists stressed that a transition plan needs more than ambitious targets: it should include clear milestones, achievable measures, and the capital expenditure needed to deliver them. They also highlighted that decarbonization can make business sense: it can reduce operating costs by improving efficiency, support innovation and competitiveness, and lower operational risks - including exposure to volatile gas prices.

Birgit Lundem Jakobsen concluded on the importance of context when interpreting environmental and operational performance metrics such as PUE and WUE; they need to be considered alongside location-specific resource constraints such as water and power availability, as well as data on emissions from construction, IT equipment and operations. She also stressed the value of a pragmatic approach, with the need to sometimes accept some short- term trade-offs, when there is a clear decarbonization strategy, supported by evidence of implementation.

The discussion extended beyond environmental impacts to the social acceptance of data centers within local communities. Birgit Lundem Jakobsen underscored that investors have now identified local communities and the risk of NIMBY as key elements of their analysis. Indeed, community opposition can lead to permitting delays and regulatory changes, which was also identified by Georgia Dunne as a key challenge DC developers are aware of.

Understanding the concerns behind opposition - such as water stress or rising electricity prices - is essential. Community acceptance may also vary between regions where data centers are just emerging and areas that already host a large cluster, she noted.

“Yet it can also create a barrier to entry,” Birgit Lundem Jakobsen noted, “which may be credit-positive for established operators by making it harder for new competitors to enter the market.”

It is therefore essential to communicate at the very early stage about a project’s impacts and benefits, “including job creation and its contribution to data sovereignty”, Georgia Dunne specified.

Responsibility rests with developers, and tenants also have a primary role to play through their community engagement program. As data centers are often pre-leased, tenants face reputational risks if local concerns go unaddressed. According to Georgia Dunne, there is a perception gap around data centers: they don’t all need the same resources or have the same environmental footprint.

“This is where sustainable finance has a role to play” commented Rebecca Smith. One of the key missions of sustainable finance is to provide more information to the market, and a green labeling on a Data Center project financing can highlight key metrics in terms of energy and water efficiency, renewable energy sourcing, much needed information for all stakeholders involved, from local communities to investors, both at the moment of financing and through annual reporting.

The panel concluded with the question: what is the single most critical factor to ensure the expansion of AI and data center infrastructure remain both investable and responsible? Georgia Dunne mentioned power, Samiya Jmili and Birgit Lundem Jakobsen opened with a reflection on: what are we using the data centers for? A perfect connection to the session on frugal AI.

For decades, traditional high impact sectors have learnt, with diverse levels of success, to secure their "license to operate". This informal acceptance of the project by local communities and stakeholders may be critical for some industries’ operations and then influence investment decisions.

As data center development accelerates, securing a “license to operate” is becoming as important as obtaining funding and planning permissions.

Beyond formal approvals, developers must earn the trust of the communities that host their facilities and demonstrate how those communities can benefit from them. This pivotal insight, identified in the “Data centers: power, place and execution matter?” panel, deserved its own deep dive.

The panel “License to operate & Business Impacts: New Learnings”, moderated by Anne-Claire Lejeune, Director, Origination & Advisory, Green & Sustainable Hub, explored what makes local acceptance specific to the data center industry, using social science insights and real-life examples. The discussion brought together Chloë Voisin-Bormuth, doctor in Sociology and Urbanism and Managing Director of the think tank Paris Île-de-France Capitale Économique, and Roland Chedlivili, CEO of Verne France, the French operations of a European developer of data centers powered by low-carbon energy.

How did we get here so fast? When local opposition and moratoria start to hit financial performance.

A few years ago, the central question for data center development was financing, Chloë Voisin-Bormuth noted. In her view, three forces have reshaped the debate since then: the growing scale of developments, the pressure to deploy infrastructure quickly, and intensifying global competition.

To illustrate the sector’s expansion, Chloë Voisin-Bormuth cited figures from RTE: in May 2026, data center capacity reached 18 GW, compared to only 5 GW about a year and a half prior, at the end of 2024.

According to Roland Chedlivili, the increasing physical scale and footprint of facilities which accompanied this sectoral growth, has contributed to greater public scrutiny: “Data centers used to be small, and no one really cared”, he noted. Today, concerns range from the size of the building to the risk of power outages, noise and air pollution.

At the same time, international competition is increasing the pressure to move quickly, with the recurrent rationale being “if we don’t build the capacity here, someone else will”, an argument that does not actually help with efficient decision-making, Chloë Voisin-Bormuth warned.

Managing these pressures, the panelists emphasized, requires early stakeholder engagement, while a facility’s location or design can still be adapted to community concerns. Once financing is secured and permitting or construction is underway, changes become significantly more difficult and costly, leaving less room to incorporate local feedback.

Engaging communities late is therefore structurally riskier and can lead to major delays, or even stop a project altogether. As of mid-2026, public opposition has already contributed $64bn of blocked or delayed US data center projects in the last two years.

Chloë Voisin-Bormuth argued that public debate often treats data centers as a homogeneous category, overlooking significant differences in their characteristics and local impacts.

This can create a twofold challenge: it may fuel opposition to new developments while making it harder for local authorities to assess proposals and make informed decisions.

Roland Chedlivili also stressed that those concerns can differ considerably among stakeholders, from mayors and municipal authorities to regions, states and local communities, requiring a tailored response from operators.

Indeed, a key challenge of this industry is a mismatch of scales. Data centers can generate significant local impacts - on land, electricity, water, noise and infrastructure - while many of their benefits are felt at regional or national level. These impacts can also accumulate across regions and nationally.

Greater transparency from operators on impacts and benefits can support mutual understanding and help decision-makers address concerns such as environmental impacts and job creation.

Opposition can arise when local communities bear the impacts while the value is captured elsewhere. Construction creates temporary jobs, but ongoing operations typically generate fewer, making investment alone an incomplete measure of local benefit.

The central question is therefore how local communities can participate meaningfully in the value created by data centers and the wider AI economy, and share that value, rather than simply receive compensation for their impacts

Verne’s approach, as described by Roland Chedlivili, offers a practical example of how to address this challenge. It illustrates how facilities can be integrated into the local ecosystem, including by reusing waste heat to heat local buildings, partnering with local universities to develop skills, disclosing clear metrics on water consumption, noise and heat emissions, and providing transparent, science-based responses to residents’ environmental concerns.

At its Seine-et-Marne site, Verne provided detailed estimates of the jobs created, broken down by role and implemented a closed-loop cooling system that significantly reduces water use. Roland Chedlivili also highlighted the company’s exploration of alternative fuels for backup power, including vegetable oil and waste-derived fuels, although noting that these remain at an experimental stage.

These efforts point to a broader opportunity: treating data centers not as isolated infrastructure, but as part of the local economy. The benefits of hosting a facility do not materialize automatically; they require deliberate planning and collaboration. Done well, this can create opportunities in energy engineering, cooling, heat reuse, real estate development, research and skills development, strengthening both the facility’s competitiveness and its contribution to the host community.

Sylvain Beaudoin, IT practitioner and contributor to the Shift Project’s Digital Technologies program, described frugal AI as an approach to designing IT projects around constraints. He distinguished “efficient AI”, the standard approach that tries to optimize a solution to be the most efficient, from “frugal AI”, where the initial considerations are around constraints on resources (including physical limits) and needs.

This, he said, represents a change in mindset to approach and tackle IT projects. "Frugal AI is being ready not to use AI. It requires to take a more open approach, defining the needs with the lowest acceptable level of performance, allowing any solution, without explicitly or implicitly mentioning whether the solution needs to be AI-based."

Another key concept in frugal AI is the measure of impact: when designing a solution, we must take into account both direct/first order (the amount of energy used to run a model) and indirect/second order impacts (for example, the number of cars avoided thanks to the AI solution). The additional challenge is the third order impacts, such as rebound effects: if an AI solution dedicated to aircraft efficiency leads to more air travel, can it be considered “frugal”?

Adrian Popescu, Research Director at the CEA and Associate Professor at Université Paris-Saclay, echoed the approach of frugality from a model design perspective. Frugality starts with engineering constraints and identifying where the bottlenecks are: computing capacity, data, model size, and cost.

The key question is therefore not which large language model to use, but rather which minimum resources are needed to achieve the required level of performance. “For some tasks, a much smaller model can deliver comparable results with far less computing power.” He added that performance gains are not linear: it is easy to get to 90-95% level of performance, but the last percentage points often require significantly more resources.

Pierre Fontaine, COO of French AI company Delos, which develops AI workspaces and autonomous working agents, highlighted the need to balance performance, cost and latency: "frugality cannot come at the expense of performance". The company's main constraint is economic: to keep prices low for its customers, it has to be efficient in its choice of models. "We pick the right model for each task, to keep costs low and latency short. At Delos, we use over 60 models for our different tasks."

Delos has two ways of encouraging its customers to be more frugal. The first is frugality by design: Maestro, Delos's model orchestrator, selects the right model for each task its customers perform. The second is visibility: Delos collects usage metrics and shares them with its customers, and has observed that once informed, customers adapt and become more frugal.

At Engie, AI is seen not only as a resource-intensive technology, but also as a potential lever for the energy transition. “Frugality also means creating value through greater efficiency across the whole production lifecycle,” noted Mihir Sarkar, Head of AI and Research & Innovation.

“AI models enable faster optimization, real-time simulations and decision-making - capabilities that existing models couldn’t support because of computational constraints.”

He also cited AI-powered autonomous drones that capture and analyze images to detect issues, such as corrosion on wind turbine blades or soiling on solar panels. The technology is now mature enough for deployment, he added.

“AI is disrupting the energy business. In five to ten years, competition may well come from hyperscalers and AI-native startups rather than pure energy players.”

AI enables better forecasting of demand, production and maintenance needs, helping reduce costs and avoid emissions by preventing unnecessary energy production. The test, he said, is whether the gains - in efficiency, avoided emissions and better decision-making - justify the resources AI consumes. And the key challenge is in accurately measuring them.

Even if we address the environmental impacts of AI, there remain critical questions on its societal consequences. The rise of AI is prompting strong reactions about its impact on work.

Artificial intelligence promises productivity gains and new ways of working, but it also raises concerns about job displacement and the dehumanization of workplace relationships and working conditions.

During the roundtable moderated by Fleur Masterman, Head of Green & Sustainable Hub Réseaux BPCE, the speakers shared concrete examples of how AI is changing tasks, jobs and skills within their organizations.

A shared guiding principle emerged: the challenge is to redesign work and processes with employees so that AI’s individual productivity gains translate into collective value while safeguarding the skills, know-how and career pathways organizations will need in the future.

The sociologist Yann Ferguson, Scientific Director of LaborIA at Inria, explained why it is difficult to establish a straightforward link between the use of AI and a potential increase in productivity.

AI assistants, for example,  can generate productivity gains at the individual level, but it is difficult to scale up or translate them into organization-wide benefits.

Bearing in mind the distinction between generative and agentic AI, the question is therefore not simply what these tools enable employees to do faster or better, but, above all, what value the organization can create through their use.

Jérémie Manigne, CEO of Niji, stressed the importance of a company’s “plasticity”: organizations must adapt and reorganize rather than treat AI simply as another tool to deploy.

At this mid-sized IT consultancy firm, equipping developers with AI tools was not enough to deliver the expected gains. The company had to fundamentally rethink its production processes - including identifying which tasks could be automated or transformed, and which should remain the responsibility of employees - in order to achieve a significant reduction in development costs.

AI adoption requires change management, with HR playing a key role. Drawing on her experience at Salesforce, HR Executive Yaëlle Leben explained how employees were placed at the heart of the process, in a bottom-up approach starting with their day-to-day activities and needs.

For example, all employees completed a certification in creating AI agents, helping them better understand the tools and how they could be used. Teams were also invited to identify tasks that AI could take on and determine which should remain the responsibility of employees.

In general, AI is increasing the importance of soft skills - particularly critical thinking and interpersonal skills -, with a lesser focus on hard skills and technical expertise.

Organizations therefore need to anticipate and plan for changes in jobs and skills requirements in order to make informed decisions about training, management support, and recruitment.

At MACIF, Guillaume Mesmin, Director of the HR and Future of Work Lab, described the focus on employee skills and employability. MACIF has reviewed job profiles and career paths and revised them for priority roles, including developers.

The recruitment and development of junior employees is another key issue. Yaëlle Leben cautioned that reducing junior recruitment because these roles are among those most easily replaced by AI, may serve short-term objectives but could undermine the transfer of know-how and the development of critical skills - with negative longer-term impacts on career paths and on organizations themselves.

AI can also cause a lot of fear, in particular of losing control over its development and usage. In the session “Living in Science Fiction: How Will AI Impact Our Organizations and Governance Models?”, the panel addressed to the question of how to integrate AI systems in business operations and building robust governance models enabling to harness the transformative potential of this technology.

Moderated by Adrian Ramon, Origination & Advisory​, Green & Sustainable Hub, the discussion offered a grounded yet urgent look at AI development, corporate governance and risk management. It featured Cristina Oprean, Co-founder and CEO of BleuAI; Stéphane Distinguin, Founder of Pantomeno; David Djaiz, President of Ascend Partners and author of La Révolution obligée and La Méthode; and Michael Herskovich, Global Head of Stewardship at the Sustainability Center, BNP Paribas Asset Management.

Rather than focusing purely on science-fiction scenarios of rogue machines, the speakers emphasized material risks, such as cybersecurity gaps and the need to set proper checks and balances for the AI sector, ensuring verifiability, control and accountability for product developers.

A recurring theme among the experts was the inadequacy of purely declarative governance. To successfully navigate the AI transition, organizations and Boards of Directors must move beyond declarative committees and implement active technical controls, rigorous audits, human-in-the-loop validation, and standardized resilience metrics. Boards should be equipped with the skills to assess AI-related risks across company functions and operations.

Ultimately, the consensus was clear: the primary danger is not that AI will inevitably outsmart humanity overnight, but that society will authorize and deploy these powerful systems without protecting and nurturing collective knowledge, know-how and interpersonal skills.  As AI commodities, competitive advantage arises from human capital and critical thinking.