Thanks to the boom in AI, data centers have moved from the background of the digital economy to center stage. They are the physical backbone behind cloud computing, enterprise software, and generative AI and support daily digital activity by businesses and consumers. Under current adoption scenarios, global data center demand could almost triple between 2025 and 2030, from about 82 gigawatts to about 220 gigawatts.1
The investment case is one of ten used in the research for the McKinsey Global Institute’s report, Catalyzing competitiveness: Where investment happens and why. The report examines how variations in the basic economics of comparable projects influence investment decisions in different regions globally and the impact those decisions can have on the future of competitiveness and growth across the world.
AI data centers are experiencing a rapid increase in investment
Not all data centers serve the same purposes. Traditional cloud data centers support a broad mix of workloads such as storage, web services, and enterprise applications. AI data centers are built for more specialized workloads, housing large clusters of advanced chips to train and run AI models. This distinction matters because AI is now a primary driver of data center growth. From 2025 to 2030, non-AI demand could increase by 1.7 times from about 38 gigawatts to 64 gigawatts. Meanwhile AI-related demand is expected to grow 3.5 times from about 44 gigawatts to 155 gigawatts, accounting for approximately 70 percent of total data center demand.2 This analysis focuses on AI data centers.
AI data centers require two main types of workloads, training and inference. Training involves building models by processing vast data sets over concentrated periods of very high consumption of computing power. Inference is what happens after deployment, when a model is used to answer prompts, run searches, and generate content in real time. Many of the largest and most power-intensive facilities today are largely devoted to training, but inference may grow faster as AI adoption broadens.3 This analysis assumes that AI data centers can serve both workloads and uses a blended average of the two. Global investment in data centers excluding IT hardware may exceed $1.7 trillion cumulatively through 2030.4
The next wave of data center development will likely combine very large training campuses, retrofitted or expanded cloud facilities, and a broader network of inference capacity located closer to users. Facilities once operated at tens of megawatts but are now being planned for hundreds of megawatts or even gigawatts, especially in the United States and China.
The United States currently has the largest share of data center investment, followed by China
The United States accounts for a significant share of data center construction (table). Northern Virginia is currently home to one of the largest data center clusters globally, but grid constraints and longer time to power are contributing to new developments in other tier-one markets and new locations.5 Domestic cloud providers and expanding state-backed infrastructure support large-scale developments in China.6 China’s largest data center hubs remain concentrated in eastern and coastal regions where demand is high, such as Shanghai and the Pearl River Delta, as well as in Beijing in the north. New capacity is increasingly being pushed into western regions with abundant energy, such as the Yangtze River Basin and the Chengdu-Chongqing economic zone, under the “Eastern Data, Western Computing” strategy.7
Elsewhere, data center expansion is smaller. The United Kingdom has the most data centers in Europe, and construction is picking up in France and the Nordic countries, which have more abundant energy. The Nordics in particular benefit from lower-cost, low-carbon power, a cooler climate requiring less energy for cooling, and room to scale. Singapore is Southeast Asia’s established regional hub with a longstanding reputation for serving demand from multinational firms, even though land and power constraints have pushed incremental growth into nearby Johor, Malaysia.8
Installed hyperscaler data center and colocation capacity in 2025 by countries analyzed
| Country | Installed capacity (gigawatts, 2025) | Region examined in this analysis |
| United States | 36.0 | Northern Virginia |
| China | 20.0 | Beijing and Shanghai |
| United Kingdom | 2.2 | London |
| Sweden | 0.8 | Northern Sweden |
| Singapore | 1.4 | Singapore |
| Global | 83.0 |
Regional differences in electricity prices and cooling needs explain the difference in levelized costs
This analysis models a 100-megawatt, liquid-cooled AI data center, representing a typical phase in a larger campus build-out. It assumes a colocation model, so costs include only the physical infrastructure needed to build and operate the facility and exclude IT hardware such as graphics processing units and servers, which users who rent the space provide. This allows a like-for-like comparison of data center infrastructure economics across countries (see sidebar “Methodology”).
China is the base case in this analysis and reflects the cost in large demand-center markets where demand is high, such as Shanghai. We compare it to similar data center investments in Northern Virginia, Singapore, Northern Sweden, and London.
Benchmark: China’s eastern region
The levelized cost of data center energy, measured in dollars per megawatt-hour (MWh) of facility energy, captures the lifetime cost of building and operating physical data center infrastructure.
In the five countries in the analysis, levelized cost ranges from roughly $200 per MWh in places in China, where demand is high, to close to $380 per MWh in London (Exhibit 1).9 In China, costs break down as follows: 55 percent of costs relate to utilities, just under 30 percent to power and cooling equipment, and the remainder to construction.
Utility prices and capex are the biggest costs of building and operating an AI colocation data center.
Note: Levelized costs shown are pretax.
1Reflects a 100-megawatt AI data center facility, excluding IT hardware. “Data centre construction cost trends,” Turner & Townsend, 2025; “Global construction market intelligence,” Turner & Townsend, 2025; “Asia Pacific data centre construction cost guide,” Cushman & Wakefield, 2026. A 7-percent premium was applied to the Turner & Townsend data center cost index to convert air-cooled cloud data center costs to liquid-cooled AI data center costs.
Electricity is the main reason for the cost gap. Electricity prices are 1.9 times higher in London than in China’s major demand centers in Beijing and Shanghai, explaining nearly 60 percent of the difference. Across countries, utilities account for almost 30 percent of the variation in levelized cost, making it the single most important cost driver (Exhibit 2).
Power and cooling equipment accounts for close to 30 percent of the variation. Hotter climates like Singapore require more intensive cooling. By contrast, markets like northern Sweden and the United Kingdom can have higher equipment costs because of stricter redundancy or performance requirements. China has lower power and cooling equipment costs, reflecting its large manufacturing scale.
Energy costs account for much of the variation in the cost of an AI colocation data center across countries.
1100-megawatt tier 3–equivalent AI data center.
Our analysis suggests AI data centers are found in two categories of markets, supply-advantaged expansion markets and demand-led hubs with deep digital ecosystems. They are highly power-intensive assets, so lower-cost electricity can create a major advantage for supply-oriented markets like Sweden. But they are also highly networked assets, requiring fiber connectivity and proximity to major users, which explains why more expensive markets like the United Kingdom still attract investment.
Why chips can flip the economics
Our levelized cost analysis isolates the cost of building and operating AI data center infrastructure, which is the typical business model for colocation data center players. However, the economics for AI labs depend on infrastructure and computing hardware, which means that differences in chip or graphics processing unit availability can materially change the calculus across countries.
The United States has strong positioning in advanced AI compute due to a combination of factors such as the dominance of cutting-edge companies in AI accelerator design, close integration with leading semiconductor manufacturers in Taiwan and South Korea, and a mature software ecosystem that enables their chips to be deployed at scale. Access to the most advanced AI chips in mainland China is limited by US export controls.
These constraints have contributed to increased reliance on domestic alternatives and on building a more self-reliant semiconductor ecosystem. These approaches can still support rapid build-outs but may involve tradeoffs in performance, power efficiency, and utilization, especially for frontier AI workloads.
Chinese systems have lower compute efficiency compared to the latest platforms from leading US companies.1 The compute efficiency of their chips, measured in PFLOPS per megawatt, is ten times or more that of China’s most sophisticated chip. As a result, a larger number of Chinese chips are needed to deliver the same level of computing power. The requirement of a larger number of graphics processing units may have an impact on cost-effectiveness depending on configuration and use case (exhibit).
China’s current leading-edge chip lags other leading-edge chips in computing efficiency and cost.
1Graphics processing units.
2Petaflops (PFLOPS) are used to measure cutting-edge computing power. One PFLOP equals one quadrillion floating-point operations per second.
This highlights an important dynamic: Data center infrastructure economics and GPU economics are closely linked outside China where AI labs source the same chips—but they also affect the case in China. If the available chips in China are less efficient, more hardware is needed to deliver the same output, raising levelized cost. While China may have advantages in data center infrastructure, these advantages can diminish and even disappear after differences in chips are taken into account.
Moreover, the highest-performing leading-edge chips are subject to export controls relating to China. If less highly performing chips are used, as is generally the case in China, the energy consumption for each unit of compute increases, flipping the economics (see sidebar “Why chips can flip the economics”).
Geography matters more on the cost side, and business model matters more on the revenue side
Data center revenues are less dependent on geography than costs. Colocation pricing does vary across markets, but rental rates vary less than levelized costs and overlap substantially across major data center hubs. Pricing depends not just on location but also on customer type, contract structure, and local supply-demand balance.
Because this analysis focuses on colocation data centers, the relevant revenue benchmark is colocation pricing in dollars per kilowatt per month. Here, the biggest differences are often influenced by business model rather than by geography. Colocation data centers are owned by operators like Equinix, QTS, Digital Realty, DATA4, and Chindata, which serve both large hyperscaler tenants and smaller enterprise customers.
Retail colocation, which serves smaller or more fragmented customers, typically has higher pricing because those providers offer greater flexibility, redundancy, and managed services. At the same time, large-scale deployments, whether through hyperscaler tenants or through wholesale colocation agreements for large enterprises, are lower risk and lower priced because capacity is often preleased under large, long-term commitments from one or a few customers.
This difference has an impact on returns. Retail colocation typically earns about $200 to $380 per kilowatt per month and can generate equity internal rates of return of roughly 20 to 25 percent. Wholesale leases earn less, at about $150 to $200 per kilowatt per month with internal rates of return of about 13 to 18 percent.
This suggests that lower levelized cost does not automatically translate into higher returns. Business models and market saturation determine whether cost advantages translate into higher returns.
Markets attracting investment today are ones that combine connectivity, demand, and speed of execution
Beyond cost advantages, AI data center investment flows to markets that offer connectivity, speed of execution, and, increasingly, strategic alignment with national AI priorities.
Connectivity is important for AI data centers, especially those serving inference workloads, which need to sit within dense fiber networks and close to major users. This explains why established hubs such as Northern Virginia, London, and Singapore continue to attract investment even when they are not the cheapest places to build or source electricity.
Investment tends to favor markets with demonstrated demand. Developers are more willing to build at scale when they have access to a base of hyperscaler, enterprise, or public-sector customers and when future utilization is supported by preleasing or expansion by existing tenants. In these cases, speed of execution matters as projects become larger and more complex.10 Labor shortages, equipment lead times, permitting delays, and slow grid connections can prevent otherwise attractive markets from adding new capacity fast enough. Lead times to obtain key components such as generators, chillers, transformers, and switchgear have more than doubled since 2019, in some cases reaching more than three years. Grid connection is another bottleneck, with average wait times now exceeding four years; in some markets, connection to power can take as long as a decade.11 Such constraints are already pushing construction costs higher and extending build times, limiting how quickly new capacity can be delivered even in otherwise advantaged markets.
Government priorities are shaping investment flows more directly than before. In Europe and elsewhere, so-called sovereign AI, or plans to develop and host AI infrastructure domestically rather than relying on foreign providers, is gaining traction. For example, SoftBank’s recent pledge to invest up to €75 billion in a five-gigawatt AI infrastructure network in France underscores the role of the country’s power availability, government backing, customer access, and sovereign AI priorities in attracting AI data center investment.12
Taken together, these factors help explain why investment does not always flow only to the cheapest energy markets only and highlight the importance of creating a market where data center projects are easier to deliver and scale. As much as 50 percent of global data center capacity due to come online in 2026 could be delayed by permitting, grid connection delays, public opposition, and rising power demand.13
Reducing the time it takes to connect a data center to power may include asking data center operators to leverage long-term power contracts or come up with self-generation arrangements, as some hyperscalers are already doing. Such steps can offset the impact of data centers on local electricity costs, which have already started to rise in Northern Virginia and other parts of the United States.14 As projects scale up, solutions may involve going beyond permitting to create shovel-ready sites faster and support standardized designs that can gain approvals more efficiently. Improved schedules are often more valuable than improved costs and remain a top priority for data center developers.
A clear strategic approach, especially in more fragmented markets like Europe, may help policymakers. Without strategic alignment, lower-cost energy locations such as Iberia and the Nordic countries may remain underutilized while high-demand hubs such as Paris, Frankfurt, and London are constrained. The key is aligning data center ambitions with advantages in power, networks, demand, and execution.
Related Articles
Data centers: The race to power AI
