Environmental, power supply and longevity risks demand a flexible approach to investment in AI data centres and infrastructure.
When a large language model (LLM) that generates replies to AI searches was asked a simple spelling question, it laboured for longer than a schoolchild. The task was to count the number of letter Rs in the word strawberry. In response, the LLM conducted a labyrinthine investigation.
“It searched through the entire body of knowledge online – everything stated on the internet – to find evidence of people discussing this question,” says James Phare, CEO of Neural Alpha, a fintech consultancy specialising in sustainability issues.
The efficiency of the search for queries, of course, determines the amount of fossil energy the LLM uses – a concern for investors with sustainability-focused mandates. The pathway of AI development, which could alter this relationship, is being mapped as each day passes. But given rocketing returns due to market dynamics, many investors have jumped on the AI bandwagon.
Insatiable demand
The AI investment boom has been making headlines throughout 2025, with the valuations of the ‘magnificent seven’ big tech stocks driven by expectations of direct and indirect profits from the transformative impact of the technology on productivity across the global economy.
This week US tech firm Oracle announced a US$15 billion increased in planned spending on data centre construction as part of an effort to compete more effectively to service AI groups such as Open AI and Anthropic.
“Insatiable demand for data means there’s a significant amount of growth in this space – more demand than supply can keep up with,” says Lee Belfield, Investment Director at GLIL Infrastructure, a fund set up by local government pension funds in the UK.
It is an emerging field fraught with uncertainties, not only over AI efficiency and energy use, but power availability. Few predictions go beyond the next five years. According to recent International Energy Agency (IEA) estimates, electricity demand from data centres worldwide will more than double by 2030 to around 945 terawatt-hours – equivalent to Japan’s electricity consumption today. Data centres currently account for 0.5% of CO2 emissions, reaching 1%-1.4% in five years.
Due to its potential volatility, some have screened out the AI sector altogether, preferring to steer towards data centres servicing cloud providers – a steady, well-understood customer base growing more slowly. That is certainly the case for GLIL. According to Belfield, cloud-based data centres are a reliable investment on a par with many renewable energy projects.
“These types of data centres generate returns closer to those of solar PV installations, providing more of a lower-risk core return,” he says. With 14 assets in different infrastructure sectors, GLIL is an open-ended fund with an ultra-long time horizon, and only considers the acquisition of existing lower-risk projects with predictable income flows from established customers.
Thirsty chips
Environmentally, cloud-based data centres are also less harmful. “The power needs of AI data centres are quoted in gigawatts as opposed to traditional data centres where they may be less than 100 megawatts,” says Miranda Beacham, Head of UK Responsible Investment at Aegon Asset Management. In addition, AI data centres experience a much higher voltage, and need dedicated cooling.
“The next generation of chips has to be cooled with liquid directed at the chip itself, whereas data centres would traditionally have acted as cold storing rooms with air conditioning to keep the chip at the right temperature. That has required a complete change in the data centre architecture,” she says.
Environmental impacts are under scrutiny in Phoenix, Arizona, a hub for data centres of all types, according to a report by Ceres, a non-profit organisation which works with investors and corporates on environmental issues. Real estate prices or natural gas availability are major reasons why operators select cluster locations – which can lead to higher costs to local inhabitants.
According to Ceres, annual water use from data centre electricity demand is expected to rise by 400% in the region, enough to supply Scottsdale – a city of a quarter of a million people – for over two years. Water demands and the proliferation of data centres could increase water stress in some areas by as much as 32% if all currently planned facilities come online, it adds. What is more, in many cases water cooling demands the use of PFAS (per- and polyfluoroalkyl substances) – dubbed ‘forever chemicals’.
Existing data centres, which tend to be closer to urban areas, are unlikely to be suitable for retrofit to be converted to AI data centres. “The power and cooling intensity needs of the next generation computer processing units are much greater, and that requires greenfield development,” says Beacham.
AI for humanity
These factors would raise alarm bells for any investor with a green remit. Beacham indicates concerns can be managed with appropriate investment selection, suggesting that AI holds great promise for solving many social and environmental challenges, partly due to the efficiency and productivity gains. “AI is possibly one of the harder sustainability challenges that we have had to analyse but it’s undoubtable that there are huge benefits to humanity from using AI,” she states.
One example is the increase in data throughput by pharmaceuticals producers, enabling speedier drug development to reduce disease. “There are massive ESG benefits and, yes, ESG detractions as well… but what we look for is how companies are mitigating those risks and managing energy use and cooling systems.”
Among the greatest unknowns is the number of data centres needed, and the amount of energy required to power them over the long term. Some investors, as well as analysts at the McKinsey Institute, have warned against “stranded assets” due to overinvestment in data centre infrastructure. This could emerge if demand for computing power is misjudged by investors, it says, estimating nearly US$7 trillion in capital outlays needed by 2030. That equates to double the GDP of France in 2025.
At the same time, the present layout of data centres may not be adaptable to software developments in coming decades, or to equipment to reduce environmental impact. The resultant longevity risk for investors, most of whom hold the investment for at least five years, is evident. Nevertheless, warnings of data centre underuse may be overstated. As Belfield of GLIL points out, this is familiar territory for real estate and infrastructure investors, and has its advantages. “There’s a significant amount of construction and demolition activity taking place, and this may be reflected in an improvement in returns because investors are taking on more risk with AI data centres.”
Infrastructure backlog
Curiously, most of the AI speculation has been ignoring prosaic issues like energy supply, which also affects construction plans, partly because AI company valuations are not correlated to data centre valuations. Yet grid infrastructure constraints in many countries could hit developers, delaying the data centre commissioning date and adding to competition with renewable energy developers for grid connection. The knock-on effect for big tech companies would soon become clear – they would not be able to monetise their products as soon as expected.
In South Korea, for example, the government announced in 2022 on data centre energy connection ban due to power restrictions. The problem is not going to disappear any time soon according to analysts at US data consultancy Green Street. “We assume this is a long-term problem. The only solution would be an ‘off-grid’ power option…but we are likely five-plus years away from anything like that being widely adopted,” they note.
Cutting fossil energy
As fossil fuel use in the field has grown, proposed solutions have mushroomed ranging from more speculative suggestions – such as small modular nuclear reactors serving big tech companies directly – to the direct connection of green electricity sources to power data centres. For instance, Downing Renewable Developments in the UK is seeking to develop the Green Data Solar Farm, a project in Harlow, Essex consisting of ground-mounted solar photovoltaics.
The project will have an installed capacity of up to 40MW providing power directly to a local data centre. On a larger scale, First Solar and Nextpower are two of several suppliers providing cleaner power under long-term contracts to big tech companies, such as Microsoft in Arizona. Comfort Systems, a large US electrical contractor, provides another potential impact investment opportunity by supplying flexible grid equipment to big tech data centres.
At the other end of the size spectrum, Exmouth Leisure Centre in Devon announced it was the first site in the UK to benefit from heat recapture by cloud data centres in 2023. The surplus heat donated to the leisure centre reduces its swimming pool’s gas requirements by 62%, creating financial savings for the leisure company and reducing carbon emissions.
If global green energy plans continue to bear fruit, most countries will be supplied by more gigawatts of renewables through the grid in the longer term, reducing the severity of pollution from fossil fuels. Water and chemicals impacts may remain, however. One solution, perhaps, is to finance data centre suppliers with a range of customers beyond big tech while another is impact investment.
Unequal time horizons
Nevertheless, rapid changes in AI development ill-matched to the long-term horizons of data centre investors creates a unique conundrum, potentially locking investors into fossil fuel use for decades or making data centres redundant. As Phare points out, AI could take surprising leaps, dramatically reducing energy use per query.
“A lot of research is going into reasoning models within AI, in order not to necessarily have to train it on huge volumes of data,” he says. But fund managers expect demand to match supply of energy, whatever that amounts to. “It’s a case of the number of queries increasing as AI gets more efficient. That’s typically what we see with tech markets,” says Beacham.
