OpenAI Chief Executive Officer Sam Altman warned that Washington is underestimating the speed and scope of China’s advance in artificial intelligence, arguing that export controls on chips alone are unlikely to slow Beijing’s progress. “I’m worried about China,” Altman told reporters, adding that U.S. policy needs to confront multiple layers of the competition, from research talent to computing infrastructure. Altman’s comments come as fresh data show the United States struggling to build the data-center capacity that underpins large-scale AI systems. Commercial real-estate firm JLL said in its mid-2025 market review that North American data-center vacancy has fallen to 2.3%, with 73% of all new capacity pre-leased and developers waiting an average of four years for grid connections. Individual interconnection requests have ballooned to as much as 5 GW—enough to power several million homes. The supply squeeze is colliding with soaring demand. Goldman Sachs projects U.S. AI-training workloads will consume about 50 GW of electricity by 2030, equal to more than 5% of the national grid. With utilities unable to add generation fast enough, companies including Meta, Microsoft and Amazon are financing their own natural-gas, nuclear and renewable plants to keep pace. Infrastructure constraints are increasingly spilling into local politics. A tally by Data Center Frontier found that US$18 billion in proposed facilities have been blocked and another US$46 billion delayed this year as communities from Virginia to Texas cite water use, noise and land impacts. In Nevada’s high desert, for example, a cluster of projects seeking nearly 6 GW of power has triggered warnings from environmental groups about aquifer depletion in the nation’s driest state. Industry executives and policymakers now face twin imperatives: expand reliable, low-carbon power and streamline permitting, while avoiding the community backlash that is stalling projects. Altman’s warning underscores the strategic stakes; unless the United States removes these bottlenecks, he argues, the physical limits of its grid—not the brilliance of its researchers—could decide who leads the next era of AI.