When bits outpace electrons

Data centers currently account for 1-2% of global electricity demand, more than the country of Japan. AFRY projects a share of global demand of around 5% by 2030. In March 2025, Nvidia unveiled its Rubin Ultra Graphics Processing Unit (GPU). Each rack may include 576 GPUs with a projected power demand of 600kW. The average European house has a demand of 1-2kW.

A confluence of maturing technologies

According to Statista, the AI market will reach USD 244 billion this year and grow by a 26% CAGR to USD 1 trillion by 2031. Enabled by the confluence of the cloud, devices, appification, communications networks, and chip processing power, mankind is taking full advantage. In addition to computation and storage, the inference capabilities of generative AI and large language models allow rapid adoption by consumers and businesses focused on productivity, customer experience, and data-driven insight.

The energy bottleneck

The supply of reliable, low-cost cost and low-carbon energy is an emerging physical constraint. A byte created, moved & stored needs energy. A standard Nvidia H100 GPU consumes roughly 3,740 kWh of electricity annually, roughly double the average European household. AI workloads concentrate this demand in specific cloud regions, accounting for 20% of current demand in Ireland. Constraints are already evident in large urban centers, such as Frankfurt, London, Paris, Amsterdam, and Dublin.

Focusing on the power demand from the chip ignores the other energy costs, like powering the networks.

Beyond AI, quantum computing is coming, offering a revolutionary leap in computational power using qubits, reliant on cryogenic environments with energy-intensive cooling and extensive error correction due to their superposed state. Energy operates under the physical laws of thermodynamics.

These can't be exponentially improved or ignored. There is no Moore’s law for energy. This largely unforeseen data center demand growth (the IEA’s 2023 World Energy Outlook did not mention AI) has upset energy industry planning assumptions and been compounded by policy emphasizing increased demand through electrification.

Can we imagine a human constraint on creating and exploiting data? We may have to.

Clean energy supply isn’t keeping up

Hyperscalers have ambitious sustainability goals. Google aims to achieve 24/7 clean energy, and Microsoft carbon negativity. But there is simply not enough reliable green energy to satisfy the need for firm power and back-up through the data life cycle.

In the US, this demand surge has led to the delay of over 9 GW of coal plant retirements and the planned addition of 10.8 GW of new gas-fired generation. At the former Homer City Generating Station in Pennsylvania, seven gas turbines are set to deliver up to 4.5 GW directly to data centers.

The sector could emit 2.5 billion tonnes of CO₂ annually by 2030, with 60% of emissions stemming from energy use and 40% from infrastructure. Microsoft’s emissions rose by 29% in 2024, and Google’s have increased nearly 50% over the past five years.

To satisfy demand, data centers have pragmatically used whatever reliable power sources are available. By default, that has meant more thermal power generation. In the US, gas will be the solver, if the supply of engines and turbines can keep up. In Europe, it is not yet clear. Mitigation efforts are scaling. Microsoft dominates the carbon direct removal market, comprising more than 60% of the total volume in 2024. Amazon has been the single largest corporate purchaser of clean energy PPAs since 2019.

Big tech investment in clean tech is surging. Geothermal funding nearly tripled to $558 million in 2024, while nuclear investment almost doubled to $1.9 billion.

Data center annual electricity consumption in household electricity consumption equivalents

Spatial concentration of various facilities versus proximity to urban areas

The sustainability challenge of water, land, and biodiversity

Data centers are like giant electric heaters. The generated heat must be cooled, often using potable water from the local watershed, with a significant proportion lost to evaporation. According to the World Economic Forum, by 2027, global AI demand will account for 4.2-6.6 billion m3 of water, more than four times Denmark’s total annual water withdrawal.

Land use is another growing concern. Campuses have large land areas, typically more than 20 hectares, displacing ecosystems and harming biodiversity as they require highly specified buildings, consume significant raw materials, and produce rising volumes of e-waste. With clusters of data centers springing up in peri-urban or rural areas, concerns about ecological loss are mounting. Data centers often offer limited long-term jobs but have high local impacts, including noise and visual pollution.

Big Tech and Big Energy can drive sustainable solutions together

Sam Altman, CEO of OpenAI, has stated that "eventually, the cost of AI will converge to the cost of energy." In other words, the future of AI is dependent on the availability of energy. Energy’s availability will depend on increasing the understanding and collaboration between Big Tech and Big Energy. The wave of US capital investments in energy is a positive step, currently not mirrored in Europe.

To avoid an inevitable crunch in meeting demand, we see specific opportunities to deepen collaboration and accelerate sustainable progress:

1. Convergence of data centers and energy companies.

This is increasingly apparent in the US. It will likely be repeated elsewhere and will go far beyond the supply of energy and capital investment for clean energy offtake. We may see the beginnings of vertical integration and blurring of the lines between energy generation, transmission, and consumption.

2. Revitalising and renewing heavy industrial sites.

Heavy industrial activity has declined in Europe in recent decades. If they can support new data centers, underutilized brownfield industrial sites may have significant potential for a data center operator and represent a much-needed diversification opportunity for struggling industries.

3. Locating near constrained renewables.

High-latency tasks like model training can be relocated closer to remote renewable energy sources which are often grid-constrained. Reconciling the sustainability challenge will be critical but would potentially reduce curtailment and system costs and improve renewable generation utilization.

4. A potential nuclear renaissance.

Nuclear is a reliable and clean source of energy, but the knowledge and supply chains needed to support rapid scaling have been lost in the OECD. Capital costs and timescales are prohibitive to private investment. Working with the government, Big Tech could underwrite the capital needed to derisk private investment and bring some of the much-needed competencies to support a new wave of nuclear construction.

5. Accelerating decarbonization solutions.

Data center growth can support developing and scaling the use of green construction materials, resource-efficient HVAC, circular design, and zero-carbon backup systems. Against very public ESG commitments, the growth in demand could be what’s needed to turbocharge new approaches to emissions reduction.

6. Expanding sustainability beyond energy efficiency.

As data center growth becomes more dependent on thermal power, the industry can lead the sustainability agenda beyond energy into water stewardship, planning for biodiversity net gain, and designing compact sites with lower material intensity.

This article is part of our AFRY Insights publication series, where experts from AFRY Management Consulting share their insights into emerging global trends across the energy and bioindustry sectors, as well as sustainability transformation.