Data centre infrastructure outlook: AI is crunching on power demand!

The hype surrounding artificial intelligence has reached a fever pitch, positioning AI as one of the most transformative and talked-about technologies of the decade. At the heart of the AI frenzy is the meteoric rise of generative AI, sparked by models like OpenAI’s ChatGPT, which debuted in late 2022, and amplified by subsequent innovations from companies like xAI, Google, and Anthropic. In 2024, global investment in AI surged, with estimates suggesting that hundreds of billions of dollars have been poured into startups, infrastructure, and R&D.

The generative AI breakthroughs and widespread adoption have had a profound ripple effect on the global infrastructure landscape, most notably, an insatiable demand for data centres. As AI transitioned from an experimental technology to a cornerstone of business, government, and consumer applications, the computational power required to train, run, and scale these systems skyrocketed, placing data centres at the epicentre of this tech revolution.

The core driver of this demand is the sheer scale of computing resources AI requires. Training large language models (LLMs) and other advanced AI systems involves processing massive datasets. For instance, models like those powering ChatGPT or xAI’s creations rely on thousands of GPUs (graphics processing units) or specialised AI chips running in parallel for weeks or months. Once trained, deploying these models for real-time inference (e.g., answering queries or generating content) demands equally high-performance infrastructure.

This computational hunger translated into a construction boom. Hyperscale data centres—massive facilities operated by tech giants like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud—expanded rapidly to meet AI-driven cloud computing needs. Although the emergence of DeepSeek has sparked questions about the huge spending on AI, prompting some to anticipate a decline in infrastructure needs, we believe this development could support longer-term demand for computational power.

The adoption of efficient lighting offers a blueprint for understanding AI’s future. When energy-efficient lightbulbs entered the market, many expected a dramatic reduction in electricity consumption. Instead, efficiency and lower costs spurred widespread adoption, unlocking new applications and ultimately increasing power consumption and infrastructure needs. The Jevons paradox shows that as AI becomes more affordable and efficient, we expect adoption to surge and generate novel use cases, driving continued and potentially exponential demand for computing power and infrastructure.

Figure 1: Data centre Inventory has been growing

AI Data Centre demand will continue to grow

The data centre industry stands at the precipice of a transformative era, driven by the relentless advancement of artificial intelligence. Initial data indicates that households are adopting AI at a rapid pace—potentially even faster than previous transformative digital technologies. From the emergence of ChatGPT in 2022 to now, more than 5 similar widely used LLMs are in the market, and it is estimated that by 2025, there will be 750 million apps using LLMs.

However, the specific ways AI is utilised by both households and businesses over time will influence overall energy demand. For instance, video generation requires significantly more energy than text generation or AI-powered search. Additionally, the trajectory of AI model development is crucial, as some methods are far more energy-intensive than others. Nevertheless, we believe that the wider adoption of AI will eventually lead to more demand for data centres.

Currently, most of the new capacity coming online is preleased to Microsoft, Google, AWS, Meta and Oracle, which has shown in their massive increase in their CAPEX projections for 2025, with most of them flowing into developments of data centres. This will help to limit the risk of introducing a significant surplus of unoccupied capacity into the market. Additionally, new colocation data centre capacity is being developed for small to medium-sized tenants who pay higher lease rates on a per kilowatt per month basis. Large utility companies such as Siemens AG and Schneider Electric SE have also cited a very upbeat outlook for demand for data centres.

Overall, meeting this growing demand for computational power will require US power producers to expand capacity by approximately 50GW by 2030. This challenge is further intensified by increasing energy requirements from semiconductor manufacturing and electric vehicle adoption, potentially necessitating nearly $60 billion in new power generation investments. Based on this current pace of under construction and planned developments, the global data centre market will likely expand at a baseline 15% CAGR through 2027 – with the potential to reach 20%.

Figure 2: Capex by hyperscalers

Figure 3: Global data centre capacity projected to grow at 15% per year but this will not be sufficient to meet growing demand

Figure 4: Household adoption rates of digital technologies in the United States

Electricity demand to increase accordingly

With AI developments expected to grow rapidly alongside increasing data centre demand, a key consequence is rising electricity consumption. Most of the new data centres are being designed to house a combination of both AI and traditional workloads. An average data centre is quite small in power terms, with demand in the order of 5-10 megawatts (MW). But large hyperscale data centres, which are increasingly common, have power demands of 100 MW or more, with annual electricity use comparable to that of around 350,000 to 400,000 electric cars.

Training foundational AI models further amplifies the power cost to AI as well. OpenAI consumed around 50 gigawatt hours (GWh) of electricity to train GPT-4, equivalent to the annual energy usage of 6,000 US households. This is a fiftyfold increase compared to the electricity required for training its predecessor, GPT-3. Individual GPUs are also consuming more power, with Nvidia’s Blackwell (GB200) chip requiring nearly seven times the power draw of the A100 chips used for training GPT-3, despite notable improvements in energy efficiency.

As such, data centres are demanding higher electricity consumption, and this can grow even larger when the number of data centres continues to increase, especially with AI receiving wider adoption. For example, due to the amount of electricity required to answer a single query, ChatGPT needs ten times as much energy to process on average as a simple query (Figure x). Thus, AI will drive data centre power demand to grow by 160%.

Today, data centres account for around 1% of global electricity consumption, and annual electricity consumption from data centres globally is about half of the electricity consumption from household IT appliances, like computers, phones and TVs. Data centre electricity usage is forecasted to double by 2026, accelerating global electricity demand faster over the next three years, and growing by an average of 3.4% annually through 2026.

Figure 5: Electricity consumption of LLMs VS Google Search

Figure 6: Data centres may consumer 12% of US Electricity by 2028

Challenges ahead: Power transmission bottlenecks

It is expected that global data centre energy demand could double over the next five years. While data centres consume large quantities of power, they are one component of the complex global power challenge. Power infrastructure bottlenecks are a major impediment to data centre development.

Power scarcity garners most of the headlines, but equally as significant are the extended timelines required to build transmission lines. These challenges will continue to intensify as the data centre sector expands rapidly into new geographies.

In many markets, extending high-capacity power lines to new development sites can take four years or longer, with the primary cause of delays being the process of obtaining easements and regulatory approvals. While supply chain challenges persist, especially for transformers and switchgear, equipment procurement is not the main factor behind transmission delays.

For the US, much of the grid dates to the 1960s and faces growing stress. The average transformer is over 40 years old, and many power lines are operating beyond their capacity. This results in frequent maintenance issues and power outages, and historically, power companies have expanded generation and transmission capacity at a steady pace, driven by long-term factors such as population growth over several decades. In contrast, data centres follow a much faster timeline, frequently transitioning from groundbreaking to full operation in less than two years. This highlights the urgent need for a robust overhaul of the infrastructure to meet growing demand and ensure reliability.

The US has more generation capacity in its interconnection queue than installed nationwide. If all the power projects that are seeking to connect to the grid were to come online right now, we would solve power demand problems for years to come.

Another challenge for the energy system is navigating the decarbonization of the grid, which clashes with the need to rapidly scale power production infrastructure. Given the transition towards clean energy, traditional coal-fired power plant capacity is set to be gradually phased out, while solar and wind alone will not be enough to meet current demand. As such, major data centre regions may face significant power shortfalls by 2030.

Figure 7: Global growth in final electricity demand by use in the Stated Policies Scenario, 2023-2030

Opportunities ahead

Despite the challenges facing the US power grid, it also presents significant opportunities. Planned investments in transmission expansion have risen from USD 9.2 billion in 2022 to USD 15.1 billion in 2024. However, investment levels must accelerate further to prevent delays in electrification efforts. Between 2024 and 2030, an estimated USD 1 trillion in grid infrastructure investments may be required.

Renewables. Recently announced data centres are set to rely on natural gas turbines for dependable, on-site power generation. While these turbines provide a steady electricity supply, their emissions pose challenges to tech companies’ carbon neutrality goals. Meanwhile, the adoption of solar and wind energy is increasing, but their intermittent nature still falls short of meeting data centres' continuous power demands. However, energy storage technology advancements are gradually helping close this gap. A potential opportunity here could be nuclear, which has garnered a lot of interest lately, driven by small modular reactors (SMRs). Unlike traditional nuclear plants, SMRs are compact, scalable, and faster to deploy with much quicker construction timelines. Tech companies are the largest occupiers of data centre space, and they have among the most aggressive net-zero targets. Nuclear provides a solution to both challenges.

Figure 8: More than 100 sites worldwide are being evaluated for SMR installations

Smart grid infrastructure. Another solution is deploying smart grid technologies. A smart grid is an advanced electricity network that leverages digital and cutting-edge technologies to monitor and regulate the flow of electricity from various generation sources, ensuring it meets fluctuating consumer demands. By synchronising the needs and capacities of generators, grid operators, end users, and electricity market participants, smart grids optimise overall system performance. This enhances efficiency, reduces costs and environmental impact, and improves reliability, resilience, flexibility, and stability across the power network.

There are nearly 11,600 projects representing 1,570 gigawatts of generator capacity and 1,030 GW of storage actively seeking interconnection, or 2,600 GW in total. That’s more than double the 1,279 GW of installed capacity currently available in the country. All these projects can be unlocked if grid infrastructure is upgraded (smart grid + transmission lines). Such capacity could solve the issue of lack of power.

Companies poised to benefit from this surge in investment include those specialising in grid-related products and services, such as Hubbell Incorporated, Quanta Services, Siemens AG, Schneider Electrics EA and Eaton Corp.

The First Trust NASDAQ® Clean Edge® Smart Grid Infrastructure Index Fund is designed to track the performance of common stocks in the grid and electric energy infrastructure sector. The index includes companies that are primarily engaged and involved in electric grids, electric meters and devices, networks, energy storage and management, and enabling software used by the smart grid infrastructure sector.

Table 1: GRID Information

Figure 9: Subsector Weightings

Key Takeaway

Overall, the boom in AI will drive rising demand for more data centres. This backlog suggests that all proposed projects were connected, the US would have ample capacity to meet future energy demands. However, the current infrastructure and regulatory processes remain inadequate to accommodate this rapid growth in renewable energy projects.

To address this issue, significant investments in grid modernisation, streamlined interconnection procedures, and enhanced transmission infrastructure are essential to fully harness the potential of renewable energy and ensure a reliable, sustainable power supply.

Given these dynamics, we see significant investment potential in the digital infrastructure value chain as companies ramp up spending on AI.

As such, investors looking to capitalise on the growth of digital infrastructure, the First Trust NASDAQ® Clean Edge® Smart Grid Infrastructure Index Fund (NASDAQ: GRID) offers direct exposure to companies enabling smart grid and energy infrastructure advancements, making it a timely opportunity to align with the growth of AI-driven digital infrastructure.

Table 6: GRID