Can data center supply keep up with AI demand?

New research from Goldman Sachs points to a precarious balancing act for data center operators

Data center energy consumption concept image showing a digitized server room with racks on either side with servers glowing blue.
(Image credit: Getty Images)

Data center capacity will probably hold up in the face of increasing demand, according to Goldman Sachs – but that's by no means guaranteed.

The bank's analysts are forecasting that AI’s share of the overall data center market will double to 30% over the next two years, eating into the share from conventional cloud workloads.

By 2030, the firm said in a new report, overall power consumption from data centers looks set to jump 175% from 2023 levels – more than the 165% that the firm previously forecast.

The financial services giant has drawn up four possible scenarios to show how the data center boom may play out.

Goldman Sachs' base case is that the balance between supply and demand is set to narrow significantly over the next 18 months, with occupancy predicted to remain at peak levels through 2026.

After that, it expects supply constraints to ease off, with occupancy rates gradually falling back to around 90% by 2028 and then leveling out.

“What we’ve seen over the last nine months is that both supply and demand forecasts have gone up, but the demand side is increasing a bit more than the supply forecast," said Jim Schneider of Goldman Sachs Research.

In another scenario, demand for compute capacity to support data-hungry AI video content could jump significantly, while a new generation of GPUs could also consume more power than expected.

"The power efficiency is improving, but power demand is outstripping what a lot of people thought," Schneider says.

​In this scenario, data center occupancy rates would exceed 100% in peak regions by 2030 and come in 17 percentage points higher than the base case.

Another possibility, the firm said, could see AI demand falling, particularly from end users. If demand drops by 20%between 2025 and 2030, occupancy would run at eight percentage points less than the base case – leading to excess supply in data centers, and potentially forcing operators to lower their lease rates.

Deteriorating macroeconomic factors could also play a role in determining the trajectory of the industry, Schneider added.

“If the economy slows a bit and companies decide they need to be a little tighter with the usage of cloud services and they optimize, then growth rates could decelerate.”

Most of the capacity in data centers still supports run-of-the-mill cloud computing, with cloud and traditional workloads currently making up about 85% of data center demand.

However, Goldman Sachs warned that in a scenario where corporate spending on these basic services declines, occupancy in data centers could fall by four percentage points from the base case, even though AI demand remains steady. Efforts to optimize the use of cloud services could also reduce demand.

If heavy demand tips the occupancy rate into a significant lack of supply, researchers said hyperscalers will be leasing or utilizing their capacity as fast as they can build it.

"We believe occupancy will continue to tighten through the medium term before the market loosens," said Schneider. "And the market could remain tighter for longer, but that is yet to be seen."

Make sure to follow ITPro on Google News to keep tabs on all our latest news, analysis, and reviews.

You can also follow ITPro on LinkedIn to stay up-to-date with all our coverage.

MORE FROM ITPRO

Emma Woollacott

Emma Woollacott is a freelance journalist writing for publications including the BBC, Private Eye, Forbes, Raconteur and specialist technology titles.