历史上，创新扩散常受最慢环节限制：Edison 的灯泡在 1880 年问世后，两年后纽约下城的首座中央电站只供应 400 盏灯，四十年后美国家庭仍仅约三分之一有电。AI 亦然：Amazon、Alphabet、Meta 的在建工程支出都大幅上升，但大型资料中心完工速度仍落后；例如 OpenAI-Oracle 的 Stargate Abilene，在近两年后仅 8 栋规划建筑中有 2 栋运作，且 600 MW 扩建已于 3 月被取消。
供给缺口正在快速量化。Goldman Sachs 预估到 2028 年电力短缺将达 45 GW，并估算到 2030 年美国传输与配电需要新增 207,000 名技术工人，而 IBEW 现有 887,000 名成员还包含加拿大与退休者。AI 公司押注能力提升后会出现足够获利的客户，但更先增加的是算力需求；若资料中心、变电站、高压电缆与钢材跟不上，成本将持续上升、现金流持续被烧掉，且 xAI 等案例也恶化了公众观感与地方反弹。
Anthropic has agreed to lease all capacity at SpaceX’s Memphis data center for $1.25 billion per month, showing how AI firms are becoming constrained by compute rather than by model ambition. The article frames this as a structural bottleneck: when software’s scalability meets physical infrastructure limits, the limiting factor is the built environment, not demand.
History suggests innovation diffusion is gated by the slowest complementary component. Edison’s bulb (1880) did not translate into rapid electrification: the first central power station in lower Manhattan served only 400 lamps two years later, and after four decades only about one-third of U.S. homes had power. Likewise, AI data-center construction is lagging despite rising capital spending by Amazon, Alphabet, and Meta; OpenAI-Oracle’s Stargate Abilene has only 2 of 8 planned buildings operating after nearly two years, and a 600 MW expansion was cancelled in March.
The supply gap is enormous and measurable. Goldman Sachs projects a 45 GW power shortfall by 2028 and estimates a need for 207,000 additional skilled U.S. transmission and distribution workers by 2030, versus an IBEW membership of 887,000 including all specialties, Canada, and retirees. Meanwhile, AI agents will likely require even more compute, while data-center buildouts rely on substations, high-voltage cable, steel, and labor that are already in short supply; the result is rising costs, worsening cash burn, debt-heavy financing, and growing public backlash such as xAI’s turbine-heavy Memphis and Mississippi deployments.