Data Centre Infrastructure 2026: Where are we now, and Where are we heading to

The data centre investment theme has been in place for several years, with rapid advancements in AI continuing to fuel rising demand for data centres.

Data centre-related companies, as measured by the Global X Data Center & Digital Infrastructure ETF (DTCR), have delivered stellar performance, generating a 1-year return of 87.21% and a 3-year annualised return of 36.4%. Meanwhile, grid infrastructure companies, represented by the First Trust NASDAQ Clean Edge Smart Grid Infrastructure Index Fund (GRID), have also performed strongly, recording a 1-year return of 52.74% and a 3-year annualised return of 26.4%.

Despite the strong performance, we believe the sector’s dynamics have evolved meaningfully. Demand for computing capacity has proven to be far more robust than earlier industry expectations, while the sector continues to benefit from the structural tailwinds of the AI boom. At the same time, valuations have become increasingly demanding, while many data centres are facing risks of delay, suggesting that more prudent portfolio management and selective positioning are now required.

Forecasts for Data Centre Power Demand Continue to Rise

US data centre power demand is projected to increase from 31 gigawatts (GW) in 2025 to 41 GW in 2026, before reaching 66 GW in 2027, according to Goldman Sachs research. This demand forecast is based on estimates that US data centre capacity will expand to roughly 95 GW by the end of 2027, more than doubling from levels expected at the end of 2025, alongside an assumed capacity utilisation rate of 70%.

The US remains the world’s largest data centre market, accounting for approximately 50% of global capacity. It also continues to exhibit the fastest growth among the three major global regions, with projected supply CAGR of 17% through to 2030. The pace of expansion is significant. Year-on-year capacity additions are expected to reach 14 GW in 2026 and 36.3 GW in 2027, compared with realised additions of 6.4 GW in 2024 and 8.5 GW in 2025. This acceleration is being driven primarily by the major hyperscalers.

In the latest earnings season, leading hyperscalers including Microsoft, Alphabet, and Meta once again revised their capital expenditure projections upwards, bringing combined projected CAPEX for 2026 to more than USD700 billion. Importantly, management teams continue to cite supply constraints as a key limiting factor, highlighting that demand remains well ahead of available infrastructure capacity.

Figure 1: Approximately 100GW of new data centres will be added between 2205 and 2030.

Source: JLL Research, 2025. Supply totals include colocation, built-to-suit, hyperscale owner-occupied and on-premise.

 

Need for AI inferencing will drive data centre demand

Although AI adoption has accelerated rapidly in recent years, AI workloads represented only around a quarter of total data centre workloads in 2025, with model training accounting for the majority of demand. However, we expect this dynamic to shift significantly, as inference workloads could increasingly rival or exceed training demand over time as AI models increasingly focus on developing agentic capabilities.

While training an AI model represents a one-off or periodic investment, inference generates recurring revenue through real-world application usage once the model has been deployed. The emergence of agentic AI is expected to multiply inference demand substantially by transforming single queries into continuous, multi-step workflows.

Unlike traditional chatbots, AI agents can autonomously decompose tasks, spawn sub-agents, execute tools, and interact with APIs. As a result, Certain agentic workflows may consume materially more tokens than traditional chatbot queries, while also requiring significantly greater computational power.

Consequently, this is expected to accelerate regional data centre deployments as well as the adoption of embedded systems at the edge.

Figure 2: Inference workloads could overtake training as the dominate AI requirement.

Source: JLL Research, 2025.

Companies across the data centre stack have been benefitting

The economics of a data centre are shaped by a unique combination of massive upfront investment requirements and significant recurring operating costs. CAPEX represents the substantial one-off investment required to construct a data centre facility.

This enormous level of spending has directly benefited companies across the broader AI infrastructure stack.

AI chipmakers such as NVIDIA, AMD, and custom ASIC providers such as Broadcom are the companies with the highest direct exposure to AI infrastructure spending. NVIDIA, recently reported data centre revenue growth of 92% year-on-year in its latest quarter, with its core compute segment growing 77% year-on-year, while networking revenue surged 199% year-on-year.

Memory chip manufacturers are also experiencing extraordinary growth. Major players including Micron Technology, SK Hynix, and Samsung Electronics are expected to deliver earnings growth exceeding 500% in FY2026.

Data centre REIT giant Equinix reported occupancy rates of 82% in its latest earnings release, with approximately 25% of its 2026 retail IBX capacity expansion already pre-sold. Order backlogs also continued to expand strongly on a year-on-year basis, rising 9% year-on-year, while the company announced close to USD8 billion in CAPEX expansion plans through to 2028. Meanwhile, Digital Realty, another leading data centre REIT, reported a backlog of USD1.4 billion as at end-2025, equivalent to approximately 1.2 times its full-year revenue. The company also maintained occupancy rates of 85% alongside a future development pipeline of 5GW.

Within the broader AI infrastructure build-out, the networking layer is becoming increasingly critical, potentially rivalling the importance of GPUs themselves. NVIDIA’s networking segment recorded revenue growth of 199% year-on-year in the latest quarter.

As internal connection distances within AI data centres can easily exceed 10 metres, traditional copper interconnects are increasingly unable to meet the required high-bandwidth demands, necessitating a transition towards optical interconnect systems. Companies such as Marvell Technology have emerged as major beneficiaries of this transition. Marvell’s optical business has delivered an approximate 50% CAGR over the past five years, with growth expected to exceed 60% this year. Management has also raised its fiscal year 2027 revenue growth projection for its optical interconnect business from 30% to 50%.

As highlighted in our previous article, “Energy could be the ultimate frontier for AI”, securing a stable power supply has become one of the key bottlenecks for AI data centres. Many US power grids are ageing and continue to face lengthy interconnection lead times. As a result, many data centre operators have increasingly adopted behind-the-meter power arrangements while also exploring co-located battery storage solutions. Natural gas is expected to play a major role in alleviating US grid constraints, serving both as temporary bridge power and, increasingly, as a source of permanent on-site power generation. This trend is evident in the sharp increase in global turbine orders. GE Vernova, the electrification and power equipment business spun out from General Electric last year, reported that its order backlog reached USD18.3 billion in Q1 2026, representing organic growth of 71% year-on-year. Total backlog reached a record USD150 billion at the end of Q4 2025, equivalent to roughly four times the company’s annual revenue.

Large data centre developers accounted for 72% of corporate clean power procurement in the Americas region through power purchase agreements (PPAs) in 2025. PPAs signed by the world’s largest data centre operators represented roughly half of all corporate clean energy PPAs during the year, marking the second-highest share since annual corporate procurement exceeded 5GW.

In one of the most notable examples, Constellation Energy — the largest operator of commercial nuclear plants in the US — signed a historic 20-year power purchase agreement (PPA) with Microsoft to fully restart Unit 1 of the Three Mile Island Nuclear Generating Station nuclear plant, which has since been rebranded as the Crane Clean Energy Centre.

Figure 3: One GW AI data centre annual costs (Billion)

Figure 4: Vendor shares across all electrical equipment

The focus now shifts to execution

The outlook for AI demand remains highly robust, further strengthening the investment case across the AI data centre infrastructure value chain. This strong demand environment has resulted in many companies accumulating substantial order backlogs.

However, amid the enormous amount of capacity scheduled to come online over the next few years, there is an increasing risk that a meaningful portion of this capacity may not become operational within the expected timeframe — not because of insufficient demand, but due to execution and infrastructure bottlenecks.

A report published in late February indicated that at least 16GW of data centre capacity is scheduled to come online in 2026 across approximately 140 projects. Yet only around 5GW is currently under construction. The remaining 11GW is still at the announced stage, with little visible construction progress, despite typical development timelines ranging between 12 and 18 months. A separate report from JP morgan showed that more than 60% of data-centre capacity planned for completion in 2027 isn’t yet under construction.

At the current pace, it would not be surprising if 30–50% of the capacity targeted for 2026 were ultimately delayed. Several factors are contributing to this risk:

1)      The Power Bottleneck and Interconnection Delays

The single largest constraint remains power infrastructure. Hyperscalers are expanding faster than regional electrical grids can accommodate. As a result, project timelines are increasingly being delayed as operators face multi-year utility queues while waiting for approval to connect new high-density facilities to the grid.

This issue has become so severe that, according to industry research from Sightline, some mega-projects are now being forced to pivot towards on-site or hybrid power generation models simply to bypass grid interconnection delays.

While on-site and hybrid power approaches still represent a niche segment in terms of total project count — accounting for less than 10% of all projects — they represent nearly half of announced capacity when measured in megawatts (MW), highlighting their disproportionately large role within the next wave of AI infrastructure deployment.

Figure 5: Data centre projects by powering model

2)      A Transformer crunch

Today, the most significant delays are no longer occurring solely within the grid interconnection queue. Instead, the bottleneck has increasingly shifted downstream towards shortages of critical electrical equipment such as transformers, switchgear, and batteries. These components are required not only to support AI-related infrastructure, but also to facilitate broader grid expansion as electricity consumption rises due to the adoption of electric vehicles and heat pumps.

US manufacturing capacity for such equipment has struggled to keep pace with surging demand, resulting in growing reliance on imported electrical components among data centre developers. Although electrical infrastructure typically accounts for less than 10% of the total cost of a data centre, project timelines can be materially delayed without timely equipment delivery. Even the delayed delivery of a single critical component can materially postpone an entire project.

Prior to 2020, delivery lead times for high-power transformers generally ranged between 24 and 30 months after orders were placed. These timelines were considered manageable in the past, when data centres neither required transformers at such scale nor within compressed development schedules. However, in today’s AI-driven environment, hyperscalers increasingly require delivery within 18 months or less.

As a result, many data centre operators are being forced to source equipment from alternative suppliers. The Approved Vendor Lists (AVLs) of hyperscalers such as Microsoft and Google have historically focused on Tier-1 Western manufacturers including Hitachi Energy, Siemens, Eaton, and Schneider Electric.

However, as lead times from these vendors have expanded significantly, many hyperscalers have increasingly turned towards Chinese suppliers despite tariff concerns and alleged national security risks. In this environment, companies such as TBEA and CHINT Group could see increasing participation within the supply chain.

Figure 6: US’s AI ambitions rely on foreign imports

Source: US international Trade Commission.

 3)      Local Regulatory Pushback and Moratoriums

The sheer volume of resources required by AI data centres — including land, electricity, and significant quantities of water for cooling — has triggered an increasingly intense wave of local opposition. Organised community pushback, alongside stricter local government zoning decisions, has led to a growing number of project moratoriums that can immediately halt or delay construction timelines.

A report from Data Center Watch found that, during the research period from May 2024 to May 2025, approximately USD64 billion worth of US data centre projects were blocked or delayed due to rising bipartisan local opposition. While the specific concerns vary by region, several recurring themes have emerged, including higher utility bills, excessive water consumption, noise pollution, potential impacts on property values, and the preservation of green spaces. For instance, the state of Maine has implemented a pause on all new data centre construction through next year, while at least 11 other US states are reportedly considering similar restrictions or moratoriums.

4)      Shortages of power, grid workers

Soaring demand for workers to build data centres, transmission grids, and power plants is intensifying already strong competition for electricians, line workers, and other Engineering, Procurement and Construction (EPC) roles, particularly as a large share of experienced workers approaches retirement. Goldman Sachs Research estimates that the US power sector will require an additional 207,000 transmission and grid connection workers, alongside 300,000 workers in manufacturing, construction, and operations, in order to support the addition of 300 GW of new power capacity by 2030. In a related development, the National Center for Construction Education and Research (NCCER) forecasts that approximately 41% of the current construction workforce will retire by 2031, signalling an increasingly tight labour supply-demand balance across the sector.

Wage growth in the construction sector has now outpaced all other US industries, with average wages rising by 5.6% year-on-year compared with overall wage growth of 4.1% year-on-year.

Figure 7: US construction wage growth by segment 

Source: US Bureau of Labor Statistics. Data: SAGF Policy Group Purchase Licensing Rights.

Data Centre Theme Remains Intact, but Execution Risk Is Rising

Rapid adoption of AI among enterprises and consumers continues to fuel strong demand for data centres — the large-scale facilities that provide the computing power underpinning AI workloads — driving a broad rally in share prices across the infrastructure value chain.

While sizeable order books and a tight project pipeline have underpinned the investment case, the most immediate challenge is increasingly centred on execution risk. Close to half of the data centre capacity currently in the pipeline for this year could face delays. If these projects are ultimately postponed, this may create near-term headwinds for some companies through delayed revenue recognition and slower capacity ramp-ups.

Nevertheless, the outlook for the data centre value chain remains constructive. Structural tailwinds continue to be supported by hyperscalers’ sustained appetite to expand capital expenditure in order to secure long-term computing capacity. This is evidenced by recent earnings updates, where projected combined CAPEX from the largest hyperscalers — Meta, Microsoft, Alphabet, and Amazon — has been revised upwards to more than USD700 billion in 2026.

As such, any meaningful market correction may represent a more attractive entry point for long-term positioning.

We continue to favour segments of the value chain that appear to be the clearest structural winners, including data centre REITs, memory chip manufacturers, alternative power generation providers, grid maintenance firms, and cooling system suppliers.

Relevant ETF exposures include the iShares Global Utilities ETF (JXI), the First Trust NASDAQ Clean Edge Smart Grid Infrastructure Index Fund (GRID), and the Global X Data Center & Digital Infrastructure ETF (DTCR).

Table 1: Comparison across the ETFs

Primary Inputs	ETF Capturing This Layer	Structural Exposure
Baseload electricity, nuclear/solar generation, PPAs	JXI	Provides exposure to utilities and power generation beneficiaries linked to rising data centre electricity demand.
Transformers, liquid cooling, switchgears, power management	GRID	Grid modernization and data centre bottlenecks are deeply intertwined; industrial backlogs are heavily data-center driven. Provides exposure to grid modernisation, electrification, and cooling infrastructure beneficiaries. Read More: The energy security trade the market has not priced yet
Cloud facilities, physical co-location space, networking silicon	DTCR	Pure-play data center operators, digital landlords, and infrastructure tech. Provides exposure to data centre operators, networking infrastructure, and AI-related hardware suppliers.