Forget chips, this could be the biggest AI opportunity

Increasing data demand driven by the rollout of AI has necessitated the rapid development of data centre processing and storage capacity globally. This rapid growth in data centre development has caused a step change in power load demand, creating some challenges for electric network operators.

Despite the challenges, 4D believes that the step change in load demand required by new data centre and hyperscaler (1) developers presents a significant opportunity for infrastructure investment across multiple sectors.

This is an abridged version of the article. Read the full article with more detail about global impact and company-specific analysis, here.

The AI phenomenon and further development of data centres

AI has been exciting investors with its potential application across virtually every industry. It’s also ignited demand for the technology to support it, such as new generation processing technology, communication towers and data centres infrastructure. What has been somewhat overlooked by investors is that power and energy are the lifeblood of AI, and the associated roll-out of data centres.

It comes as no surprise then that data centre build out has been rapid over the past decade, with the number of data centres globally shown below.

The ongoing development of AI will require enhanced processing technology within data centres, such as next generation processors, which have the capacity to process data on a magnified scale. This new processing technology, and the more advanced cooling technology required to support it, have both driven increased data centre development across the globe. Cutting edge data centres that utilise new processing and cooling technology to facilitate AI include the hyperscale data centres.

Traditional co-location data centres are established to meet the varying needs of multiple customers within a single data centre facility. By contrast, hyperscale data centres are usually much bigger data centres in terms of capacity, and are designed to meet the specific technical and operational requirements of companies like Amazon, Meta, Google, Alibaba, IBM and Microsoft. These same hyperscale companies leading the AI revolution, are driving exponential growth in data centres. 

Analysis by Synergy Research Group shows that about 20% of global data centre capacity is already dedicated to AI. The global hyperscale data centre market is estimated to reach approximately $935.3 billion by 2032, with a forecast CAGR of 27.9%

Expected load demand growth

An important aspect of dealing with AI processing technology is the need for cooling of this new technology. According to the International Energy Agency (IEA), processing power accounts for about 40% of a data centre’s power needs; while cooling requirements to achieve stable processing efficiency makes up another 40%. The remaining 20% comes from other associated IT equipment.

AI technology increases the average power utilisation of a rack (2) within  a data centre from around 8 kW per rack in 2020, to 40 kW/rack for more modern processor racks (typically used for cloud computing), and up to 132kW/rack for liquid cooled latest edition Nvidia processors. This means the energy intensity of a data centre with the same number of racks has increased between 5-15x since 2020.

Future trends of the data centre sector are complex to navigate, as technological advancements and digital services evolve rapidly. Depending on the pace of deployment, the range of efficiency improvements, AI and cryptocurrency trends, and impediments to development, the IEA expects global electricity consumption of data centres, cryptocurrencies and AI to range between 620-1,050 TWh in 2026, with the base case for demand at just over 800 TWh – up 74% from 460 TWh in 2022 (almost 2% of total global electricity demand). This corresponds to an increase in power demand roughly equivalent to Germany.

Facilitating data centre development

The incremental power demand, and strain that data centre development places on the whole electrical network, are significant.

Some network operators and utility companies globally are struggling to facilitate this significant load from data centres while ensuring that the associated development costs will not cause affordability issues for other customers. Utilities are actually looking to ensure that the new data centre load is a benefit for other electrical service customers, which requires negotiation and planning on behalf of the utility companies.

Community concerns related to data centre development typically include power scarcity, public costs and water consumption. Data centre providers and hyperscalers believe that through community/stakeholder engagement and proper project design, many of these concerns can be alleviated, and the benefits of a data centre investment can be fully appreciated.

Waiting time to facilitate load demand and interconnection

The lead time for a data centre interconnection to the grid will vary based on where the data centre is located. There are examples globally where constraints on available load supply and/or network (transmission and distribution) limit the interconnection of new data centre projects to the system, such as in Dublin, Ireland, and Northen Virginia, US.

Hyperscalers with load demand in excess of 25 MW are experiencing a wait of between two to seven years in the US for incremental supply to satisfy their operational needs. Based on TD Cowen recent checks, they forecast the lead times for procuring data centre power across the FLAPD (Frankfurt, London, Amsterdam, Paris, and Dublin) markets:

• 3-5 years in Frankfurt,
• 4-5 years in Paris,
• 3-5 years in Amsterdam,
• +10 years in London (absent buying reserved power from another operator), and 
• Indefinite lead time in Dublin (Eirgrid is not accepting applications) (3).

Data centre developers are looking at alternative energy solutions such as fuel cells or supplemental natural gas turbines to facilitate projects on shorter schedules. The ability for data centres to operate as a microgrid can accelerate the development timeline as the grid’s main limitation is a lack of network capacity during peak demand – usually the summer season.

Hyperscalers in some cases have gone to lengths to secure power supply through directly contracting and connecting specific power generation facilities on co-generation sites. A recent example was in March 2024 when Talen Energy announced its sale of a 960 MW data centre to Amazon Web Services (AWS) for $650 million. The 1,200 acre campus is directly connected to and powered by the adjacent Susquehanna Steam Electric Station (nuclear facility), which generates 2.5 GW of power (4) By contracting directly with the Susquehanna facility, AWS is able to obfuscate the network connection wait time.

Location sensitivity of data centres in facilitating AI

Through the iterations of data centre development from facilitating internet, then to cloud, the importance of minimising latency has been central to the service proposition for customers. This latency minimisation requires deployment of data centres closer to major population centres where network density already exists. This led to a clustering effect that drives dense pockets of data centre power load in major metros, leading to localised transmission constraints.

Considering the exponential demand in the most recent iteration of data centre development to facilitate AI, despite deployments being much more power-intensive than their predecessors, they are not nearly as latency and location-sensitive. To date, hyperscalers have preferred to deploy data centres near major population centres due to the flexible use of the data centre capacity (if AI does not generate the returns hyperscalers need, they can repurpose the data centre capacity to support latency sensitive workloads such as cloud). However, as power constraints across markets grow, hyperscalers are looking to deploy capacity in more diverse areas.

In 2023, the rise of the ‘location agnostic’ data centre development reflects the lower latency sensitivity of the power intensive AI centres. There has also been a shift in the land acquisition strategies of hyperscalers as they look to build in less populated areas where power is available, and land is cheaper. A great example of this is Microsoft's acquisition of +1,000 acres in the Wisconsin Innovation Park, US, that will support a massive data centre campus where it will invest $3.3 billion in the next two years to support national cloud and AI infrastructure.

Power constraints in China began to emerge in major cities including Beijing, Shanghai and Shenzhen, which limited the ability of co-location data centre operators, including GDS, to secure power for new developments. As power constraints intensified, the Chinese government established the ‘Eastern Data, Western Computing’ plan in 2022. The plan looked to split latency sensitive workloads. Non-sensitive data centre projects, such as AI focused and/or hyperscale developments, were shifted to less densely populated regions in the western part of China, where there was greater land available and access to renewable power (5). 

Data centres operate around the clock, so have a requirement for consistent, uninterrupted load. Many of the large hyperscalers have also made ambitious commitments to net-zero, which are summarised in the table below. The operational requirement for uninterrupted load, is somewhat is conflict with decarbonisation ambitions due to the intermittency of most sources of renewable generation. 

Opportunities for infrastructure companies

The incremental load demand requirements of data centres are creating a challenge, but also an opportunity for network utilities and IPPs across the world. The incremental load required by data centres globally is going to drive:

• additional generation development opportunities, probably at higher power prices (as supply catches up with demand);
• network investment to facilitate interconnections; and
• investment in network modernisation to support higher peak network load.

The obvious beneficiaries are IPPs who will target data centre developers and hyperscalers to construct generation to fulfill their requirements, or contract existing supply available in merchant markets at higher prices than is currently available. Hyperscalers need for speed to market, and uninterrupted security of supply has meant that they have been willing to offer PPA prices in excess of what is available in current merchant markets. This has been shown by Microsoft contracting separately with Constellation Power for nuclear energy is being contracted at prices rumoured to be far above the $60/kWh futures price 2026 (6) which is currently prevailing in the PJM zone merchant price market in the eastern US. Analysts predict these prices could reach as high as $100/kWh.