AI’s environmental footprint: why investors should care

Artificial intelligence is widely viewed as a software-led growth theme. In reality, however, it is increasingly becoming an infrastructure-heavy business model with a meaningful environmental footprint. For investors, the key question is no longer simply whether AI demand will remain strong, but if the sector can scale in the context of constraints on power, water and critical materials without creating bottlenecks, delays or social tensions, while also creating opportunities in efficiency and resource management.

Water is becoming a strategic constraint

The first debate around AI focused on electricity demand, grid capacity and carbon intensity. That remains important. But water is emerging as an additional and increasingly material constraint, particularly for data centres, the physical infrastructure that powers AI models. Water consumption is expected to rise significantly, with forecasts indicating +50 trillion liters by 2050.

The water footprint of data centres has two dimensions. First, there is direct water use, mainly for cooling servers that generate significant heat. Second, there is indirect water use, linked to semiconductor manufacturing and energy generation. Together, these exposures are becoming more relevant as AI infrastructure expands. However, a distinction should be made between consumptive and non-consumptive water use. Although vast amounts of water consumption in AI deployment are non-consumptive, that is returned after ‘use’, the environmental and societal impact is not negligible as water is often returned warm or altered and therefore likely to impact ecosystems and put constraints on other users.

Forecast water demand in new economy industries

Source: GWI (2026)

For investors, the issue is not only the scale of consumption, but the geography. Water stress is highly uneven. In the US, availability can vary significantly across states and even within them, affecting site selection, operating continuity and permitting. In Europe, increasing heatwaves and droughts are making the operating environment more challenging for an industry still building out its AI capacity (under the Cloud and AI Development Act, the European Commission aims to triple the EU’s data centre capacity in the coming five to seven years). Moreover, cooling needs are also typically highest during periods of greatest stress, for example, in the summer.

This turns water into both an operational and a political issue. Facilities located in water-stressed basins may face higher costs, less reliable access, slower approvals or reputational pressure.

Yet, beyond chronic climate risks, acute risks are also jeopardising operations. First Street, a climate risk analytics firm, recently reported that '79%, of data center capacity faces high acute climate risk from flood, wind or wildfire’. Swiss Re, the reinsurance company, confirmed the trend, stating more and more data centres are being constructed in zones prone to acute climate risks such as tornado activity.

The upstream risk: semiconductors are exposed too

The water challenge does not stop at the data centre gate. Semiconductor manufacturing is also highly water-intensive, and production is concentrated in regions, especially in Asia, that already face water scarcity. A 2026 case study report published by the Task Force on Nature-related Financial Disclosures (TNFD) estimated that the sector uses roughly 210 trillion liters of water per year, almost half of which is consumed in areas facing higher-than average water scarcity. With processing continuing to intensity, as advanced chips require more water, IDTechEx, an international market research and business intelligence company, expects the industry’s water usage will double by 2035.

So what about costs? S&P research suggests that climate-related water risks, including shortages, flooding disruptions and community friction, could cost major tech and semiconductor value chains up to USD 24 billion annually by 2050.

This matters because the risk is not limited to chipmakers alone. Companies such as TSMC, UMC, Micron and Vanguard International Semiconductor have already been affected by droughts. Even where the largest producers are able to absorb short-term costs (although disclosures on cost impacts and resilience strategies remain diverse), supply-chain disruption can still affect broader AI infrastructure, equipment and deployment timelines. For investors, the question is therefore not just who uses water, but where that water is sourced, how exposed the business is to local stress, and which downstream actors might be impacted.

Zooming in on data centres: cooling efficiency helps OpEx, but may require more CapEx

Within data centres, one of the central investment issues is cooling economics. More efficient systems can reduce operating costs by improving thermal performance and lowering both Power Usage Effectiveness (PUE) and Water Usage Effectiveness (WUE). A Jefferies publication estimates that every 0.1 reduction in PUE, can save around USD 1.2 million per year in operating costs for a 20 MW data center. This helps explain the shift away from traditional air-based cooling toward more advanced liquid-cooling solutions.

These systems should gain market share over time and can improve both energy and water performance. But they also require higher upfront capital expenditure and greater operational complexity. From an investment perspective, the question is whether management is investing early enough in resilience CapEx to reduce future resource risk. The best operators may not be those with the lowest initial cost, but those building the most adaptable infrastructure.

According to BNP research, out of every USD 100 CapEx in the AI value chain, USD 7.5 is heading to cooling, and USD 20 to power. This picture is likely to change significantly moving forward.

Permitting and social license are increasingly material

While semiconductor manufacturing is still largely concentrated in Asia, public scrutiny of data centres operated in the US and EU is intensifying globally, especially in regions where major facilities are located in water-stressed areas. A 2025 investigation by The Guardian and SourceMaterial identified 38 active data centres operated by Microsoft, Amazon and Google in regions already facing water scarcity, with 24 more under development. A 2026 Guardian report confirmed the trends, with planned centres continuing to be built in drought zones.

The hyperscalers themselves acknowledge the issue, although disclosure remains uneven. In 2024, Microsoft said that 42% of its water came from areas with water stress, while Google said 15% of its water consumption was in areas with high water scarcity. Later reports confirmed the trend, with Amazon disclosing that 22% of its water use was in areas with high or very high water stress.

Developers with the most data centers in areas of drought

Source: The Guardian (2026)

The investment implication is clear: water-based cooling, once seen primarily as an efficient solution, is now being evaluated through a broader lens that includes local (and seasonal) availability and hence operational risk, social acceptance and regulatory risk. Studies also suggest that data centres can compete with other societal and environmental users for water. But what is market news telling us?

In some markets, this is already affecting project execution. In Chile and Uruguay, Google data centre projects faced community and government pushback between 2022 and 2024, leading to changes in cooling strategy. In the US, Amazon’s USD 3.6 billion investment in Arizona drew opposition because of drought risk. In Europe, friction looks to be less prevalent, although in Ireland water supply-related investments to meet data centre cooling demand are triggering debate, such as the Shannon Water Supply Project in the Greater Dublin area. The project aims to increase water supply in the region, and hence resilience, yet faces significant opposition due to the environmental impact, drought risk and costs. Finally, in the UK, a 600 MW data center proposal was delayed over summer, with Scottish ministers requiring a full environmental impact assessment by the developer before reopening the consultation on the project, after 14,000 submissions were received during earlier consultation.

And although hyperscalers are innovating by looking at closed-loop technologies, their ‘holy grail’ for tackling water risks, these systems require more energy, which is again linked to vast amounts of water being consumed indirectly. Hyperscalers are therefore increasingly turning to fossil fuels, as the fastest way to secure power. Apart from the balancing act with emissions targets, the switch would also fuel the power debate even more.

For corporates, permitting risk is no longer just a planning matter. It can become a material factor in timing, returns and local operating stability. Developers able to demonstrate credible plans for power procurement, water stewardship, infrastructure investment and community engagement may be better positioned to navigate permitting process and avoid costly delays.

In Europe, a new rating system for the sustainability of data centres, adopted on 21st September and applicable as of 2027, might serve as guidance for local and regional authorities during the permitting process, with the focus placed heavily on water and power usage effectiveness. The actual effect of the label on permitting, connectivity and financing however remains highly uncertain. This is due to insufficient coverage for a proper assessment by authorities and network operators, as specific elements like reuse of waste heat or the impact on the wider energy system are omitted, as while waste heat is referenced, it is not integrated.

Other resource constraints matter too

AI’s resource intensity is not limited to energy and water. The value chain also depends on critical materials, where the issue is often supply concentration rather than absolute scarcity. High-purity quartz and copper are both essential inputs for chip manufacturing and AI infrastructure, and both can be vulnerable to local disruption, extreme weather and tight supply conditions. The same applies to conflict minerals such as tin, tantalum, tungsten and gold. As AI build-out continues, these constraints could support higher input costs and slower project delivery.

Furthermore, another resource might be at risk in the future, yet not due to concentration risk or resource constraint. PFAS supply, still largely used for advanced server cooling systems, specialised fire suppression agents, and the microchips powering servers, might be at risk as nations worldwide consider PFAS-bans/phase-outs due to the negative environmental and societal impacts and continued societal pressure. With only limited alternatives available, it is difficult to forecast how regulators, and the AI value chain, will tackle this.

What investors should focus on

AI remains a compelling structural theme, but environmental constraints belong in the long-term investment analysis, with data centres for example typically expected to be in operation for 20 to 30 years. Investors should look for companies that are actively managing resource risk rather than assuming it away.

Overall, this calls for investor dialogue on environmental scenario analysis (stress testing) to complement traditional fundamental views and the assessment of short, medium and long term resilience. Key areas to monitor include power sourcing and community concessions strategies, resilience and resource efficiency CapEx, water stewardship, insurance, supply-chain diversification and hard metrics on water use, reuse and efficiency.

To conclude, AI’s environmental footprint is not only a risk story, as the above-discussed points also create opportunities for companies that improve efficiency, reduce water intensity or enable better resource management. The investment challenge is to identify which parts of the value chain are best positioned to scale responsibly and profitably within tightening environmental limits. Furthermore, resource constraints can also drive innovation and create value for companies that deliver more computing capacity with less energy, water and infrastructure. AI may also help optimise the systems it is straining. According to the IEA, wider use of existing AI applications in the electricity sector could save up to 110 billion dollars a year and unlock transmission capacity. The next phase of AI may therefore be shaped not only by scale, but by efficiency, resilience and responsibility.