AI's insatiable thirst for power is expected to skyrocket in the coming years, potentially leading to data center power shortages. According to a new report from Gartner, a new server last year required 195 terawatt-hours of power. This is equivalent to the amount used by 18 million households in one year. But by 2027, new servers could provide 500 terawatt-hours of electricity, or the equivalent of 46 million homes.
That's far more electricity than existing data centers use, which Goldman Sachs estimates will already consume 349 terawatt-hours by 2022. Without additional carbon-free power sources, pollution from AI training and use could skyrocket. Sam Altman's big bet (more than $375 million) on fusion power is starting to make more sense.