当前数据中心约有 30% 的输入电力并未用于 AI 计算，而是被冷却系统与长距离供电损耗消耗。Nvidia 指出，随着规模扩大，这些损耗只会增加，但效率提升也会放大收益；Nvidia 已投资 Emerald AI，Google 也在优化模型能效。Blackwell 芯片在 2024 年推出后，以相同能耗提升算力，但热量大增，推动行业从风冷转向直连芯片的液冷方案；研究显示，液冷可将能效提高 15%，并把化石燃料购电相关排放降低 10%。
更深层的重构集中在供电链路：电力从电网进入后需由交流电(AC)转为直流电(DC)，并从 34,500 伏降到芯片所需的 12 伏，层层转换都在损失能量。Nvidia 正测试 sidecar，将转换步骤压缩并以 800 伏 DC 供电，按 Flex 估算较现行系统节能 20%；更激进的固态变压器可提升到 27% 的效率改善。行业目标是在 2030 年前让许多 AI 工厂转向 800 伏 DC，而一旦实现，配电系统造成的总损耗可能从约三分之一降至不足 1%，并更容易接入可再生能源。
Artificial intelligence data centers are nearing the limits of power and cooling. As chipmakers such as Nvidia Corp. release ever more powerful chips, the next generation of data centers will consume many times more electricity than earlier facilities; these “AI factories” are said to use enough power to light millions of homes, potentially pushing up US electricity prices, widening AI’s carbon footprint, and slowing growth. Political pushback is already building, and Elon Musk warned that “very soon, maybe even later this year,” chip output could exceed the ability to power it; meanwhile, trillions of dollars are expected to flow into the AI build-out.
About 30% of the power entering data centers is not used for AI computation, according to Nvidia, but is lost to cooling systems and long-distance electrical transfers. Nvidia says those losses will rise as scale expands, though efficiency gains become more valuable too; the company has invested in Emerald AI, and Google is improving the energy efficiency of its models. Nvidia’s 2024 Blackwell chip raised performance at the same energy use, but generated much more heat, accelerating the shift from air cooling to direct-to-chip liquid cooling. A study found liquid cooling can improve data-center energy efficiency by 15% and cut emissions from purchased fossil-fuel power by 10%.
The deeper redesign focuses on the power path itself: electricity enters from the grid as alternating current (AC), must be converted to direct current (DC), and then stepped down from 34,500 volts to the 12 volts chips need, with losses at each stage. Nvidia is testing a sidecar that reduces conversion steps and supplies 800-volt DC, which Flex estimates can improve energy efficiency by 20% versus the current system; solid-state transformers could raise that gain to 27%. The industry aims for many AI factories to run on 800-volt DC before 2030, which could cut power-distribution losses from roughly one-third of total losses to less than 1% and make renewable integration easier.