Datacenter energy use to more than double by 2030 thanks to AI's insatiable thirst
AI’s thirst for electricity will see datacenter energy use more than double by the end of the decade – just five years from now – according to the latest forecast from investment banker Goldman Sachs.
The stark warning from Goldman Sachs Research observes that worldwide interest in generative AI has resulted in an arms race to develop ever bigger and better models, which will require more high-density bit barns, plus much more electricity to power them.
The financial services biz estimates the current total power consumed by the global datacenter industry is around 55 GW. More than half of this (54 percent) pertains to cloud computing workloads, with traditional business functions including email or storage making up 32 percent, and AI workloads 14 percent.
By modeling future demand for each of these types of workload, the projections are for power requirements to reach 84 GW by 2027, with AI expanding to comprise 27 percent of this, while cloud makes up half and traditional functions 23 percent.
If accurate, this will mean that power consumed by all those data facilities is set to increase by 50 percent in just a couple of years, and Goldman Sachs expects the trend to continue, with about 122 GW of total datacenter capacity online by the end of 2030.
Amazon spent $75 billion in capital expenditure, mostly related to AWS, in 2024 and expects to fork out even more in the current calendar year. Similarly, Microsoft is projecting a spend of $80 billion in 2025 on infrastructure to train and deploy AI, and Meta some $60 billion on AI resources.
The total increase in capacity means that electricity grids will also require significant funding to keep pace, and Goldman Sachs estimates that as much as $720 billion of spending may be needed for upgrade work between now and 2030.
That might seem like a tall order. However, the alternative is that lack of power could put the brakes on datacenter expansion in some areas.
“These transmission projects can take several years to permit, and then several more to build, creating another potential bottleneck for datacenter growth if the regions are not proactive about this given the lead time,” said senior equity research analyst James Schneider at the bank.
A report last year from management consultancy Bain & Company warned that growing electricity consumption in the US may outstrip supply in just a couple of years, unless energy companies move to boost their generation capacity and distribution.
A separate report claimed that Americans could face a 70 percent hike in their electricity bills by 2030 unless such action is taken.
Goldman Sachs expects to see the balance of supply and demand in the datacenter industry tighten in the near term, with the occupancy rate for infrastructure estimated to surge from around 85 percent in 2023 to a possible peak of more than 95 percent in late 2026.
Afterwards, the researchers believe there will be a slackening of occupancy rates starting in 2027 as more server farms come online and AI-driven growth slows. Yet this slowdown is based on an assumption by the researchers that more efficient AI models will lessen the need for quite so much infrastructure investment.
“In the long run, if we see efficiency driving lower capex levels (from either hyperscalers or new investment plans from new players), this would mitigate the risk of long-term market oversupply we see in 2027 and beyond – which we think is an important consideration that could drive more durability and less cyclicality in the datacenter market,” Schneider said.
Being a financial services operation, Goldman Sachs is as much interested in the investment opportunities presented by this potential crisis as warning about it, of course.
“Major datacenter operators with significant asset footprints that serve the likes of hyperscale cloud companies as well as large enterprise customers will be well-positioned to meet the expected surge in global demand,” the company observes. ®