As demand for artificial intelligence computing surges globally, US AI data centers (AIDCs) are increasingly hindered by prolonged delays in securing power grid connections. The average wait has extended to approximately five years, significantly outpacing the typical 18 to 24 months required for data center construction, creating "completed but idle" facilities and prompting urgent calls for regulatory overhaul.
Microsoft to adjust power arrangements amid electricity price concerns
US President Donald Trump recently announced that Microsoft Corporation intends to implement significant adjustments to ensure that electricity costs from AIDCs do not adversely affect ordinary Americans. The statement also hinted at forthcoming government intervention to mitigate the financial burden on consumers caused by the substantial power consumption of these data centers.
Traditional grid infrastructure struggles to meet AI computing demands
Research from Princeton University's ZERO Lab, published in December 2025, highlights that traditional grid constraints—specifically limited transmission bandwidth and generation capacity—are the main barriers slowing AIDC grid connection approvals. The conventional approach requires data centers to wait passively for power providers to upgrade regional grids, often resulting in multi-year delays.
Flexible connections and self-supplied capacity could reduce wait times
The Princeton study proposes a novel model combining "flexible grid connection" contracts with "Build Your Own Cloud/Capacity (BYOC)" strategies, potentially reducing power delivery times by three to five years. Flexible grid connections involve a layered contract structure that guarantees an uninterrupted baseline of 24-hour power service. Simultaneously, data centers agree to actively mitigate peak grid demand by switching to battery energy storage systems or backup generators during rare grid stress events, which empirical data suggests occur only 40 to 70 hours annually. This arrangement can expedite the issuance of grid access permits by more than three years.
The BYOC approach encourages data center operators to take responsibility for their supplemental power needs. Operators can enter power purchase agreements, develop virtual power plants, or invest directly in renewable energy infrastructure, thereby increasing generation capacity without relying solely on grid expansions. This strategy aims to align electricity costs with the sources of demand, adhering to the principle that entities driving consumption should bear corresponding expenses.
Implications for Taiwan's energy and AI sectors
Experts note the US model offers a valuable framework for Taiwan's evolving energy and AI landscape. Taiwan's policymakers could adopt flexible grid connection contracts that prioritize operators who use off-peak power and manage peak loads independently. Additionally, institutionalizing self-provided capacity would encourage domestic AI companies to invest in local renewable energy and long-duration energy storage, mitigating the impact on residential electricity tariffs. The proposal also advocates digitalized grid dispatching, using software solutions to precisely manage peak loads during critical hours. This shift from hardware-centric expansions to dynamic, software-driven scheduling could accelerate the operational readiness of AI computing facilities.
From static to strategic: reimagining electricity supply
The evolving challenges underscore a paradigm shift in electricity supply for AI data centers. Power grids are transitioning from static, infrastructure-based assurance to dynamic, strategic resource allocation. In this context, effectively managing limited peak hours becomes a key determinant of competitive advantage in the AI era, emphasizing the importance of flexible power solutions to sustain rapid AI development.
Article translated by Jingyue Hsiao and edited by Jerry Chen
