Data center construction boom enters a new chapter: inference becomes the focus, power becomes a bottleneck.

Industry Background

The rapid development of the global AI industry is driving data center construction into a new phase. According to a recent analysis by Bain & Company, the data center market has exhibited six significant trends over the past 12 months, marking a shift from pure scale expansion toward a more complex and refined development model.

Large enterprises are accelerating the deployment of production-grade AI applications, prompting hyperscalers to continuously expand their computing capacity. The momentum and growth prospects of AI remain strong, which is expected to translate into sustained robust demand for data center construction.

Market Impact

New construction growth is stabilizing: Although there were earlier expectations that hyperscaler investment might decline, the reality is that investment has increased significantly in 2025 and is expected to continue growing in the coming years. However, vendors are beginning to focus more on capital efficiency and becoming more selective about new deployments, especially for AI training.

Data centers are becoming larger but more flexible: "Gigacampuses" with at least 1 gigawatt (GW) of power capacity will become the standard for frontier model training, but only a limited number of such data centers will be needed globally. Meanwhile, inference workloads have relatively modest scale requirements, enabling a network of smaller, distributed data centers. New data centers are being designed to flexibly switch between training and inference workloads—for example, by deploying multiple cooling solutions to avoid stranded assets. Distributed training is also being explored and could reshape training architectures.

Inference is becoming the new focus: AI workload patterns are shifting, with the importance of scaling inference rising significantly alongside frontier model training. This shift is partly driven by the clear implementation of enterprise AI use cases. Test-time compute is reshaping infrastructure strategy, economics, and architecture, with profound impacts on data center build-versus-lease decisions, chip diversity, and power configuration.

Competitive Landscape

Growth is concentrated but globalizing: North America holds the largest data center capacity, primarily driven by hyperscaler capital expenditure. At the same time, sovereign AI strategies and enterprise adoption are activating regional markets around the world. Enterprises face decisions about which markets are suitable for which workloads. They are seeking geographic flexibility to align computing infrastructure with latency, data sovereignty, and energy access factors.

Power is becoming a growth bottleneck: Even as GPU and construction constraints ease, power access has become a key limiting factor for growth. Behind-the-meter (BTM) power generation is altering construction decisions and timelines. Currently, BTM projects are most common in the U.S., relying mainly on standalone gas-fired generation, though other technologies such as solid oxide fuel cells are being explored. Utilities, developers, and regulators face urgent coordination pressures, and there have already been cases where utilities and data center operators cooperate to plan large-scale electricity demand.

Enterprise Insights## 企业启示

• Enterprises should reassess their infrastructure strategies for AI workloads: the differing requirements for training and inference call for more flexible design and site selection.
• Power supply will become a key factor in future site selection, and enterprises need to coordinate in advance with energy suppliers and regulatory agencies.
• As inference becomes the focus, enterprises should pay attention to inference cost optimization and the possibility of distributed deployment.
• In global deployment, it is necessary to consider data sovereignty, latency, and energy availability simultaneously to build a diversified layout.

未来展望

In the next 12 to 24 months, we expect:

• Hyperscalers will continue to increase data center investments, but capital allocation will become more refined, for example, prioritizing the construction of inference centers.
• Power bottlenecks will drive more innovative power solutions, including behind-the-meter generation, small modular nuclear reactors, and other technology explorations.
• Regional markets such as the Middle East, Southeast Asia, and Latin America will accelerate construction due to sovereign AI and local enterprise demands.
• Data center architecture will evolve towards more flexibility, multiple cooling solutions, and the ability to dynamically allocate training/inference resources.
• It is expected that within three years, global data center capacity will still maintain double-digit growth, but the construction model will shift from "rush building" to "precise construction."

For the entire AI industry, the evolution of data center infrastructure is both a support and a constraint: only when infrastructure bottlenecks are effectively overcome can the large-scale application of AI truly unleash its potential.