Skyrocketing AI consumption to trigger 500% data center supply deficit by 2030: Study
The global annual demand for data centers is expected to reach nearly 90 gigawatts (GW) by 2030, exceeding available supply by as much as 500%, as artificial intelligence (AI) consumption and inference demand continues to skyrocket, according to the findings of an Itron Mountain-Structure Research study released on Friday.
The findings come as generative AI and agentic AI have become ubiquitous across industries, spurring new applications that require significant compute capacities. However, as demand continues to exponentially surge, the data center build-outs have started lagging, partly due to supply chain-related constraints impacting the availability of hardware components and raw materials.
“In the three years since ChatGPT launched, generative and agentic AI have become ubiquitous,” the study said, adding that investment in graphics processing units (GPUs) and data centers is soaring, leading to significant changes in adoption.
It found that the hyperscaler capital expenditures are projected to reach $375 billion in 2026, a 36% increase from 2024. Half of this investment is spent on servers and GPUs, while the other half would be spent on data center capacity. “This rapid growth will cause a massive supply deficit,” the study said.
A recent analysis from Goldman Sachs found that the scope and scale of planned AI-related capital expenditures have grown immensely. It stated that the large technology companies that are leading the buildout will cumulatively spend $5.3 trillion in capital spending from 2025 through 2030.
An S&P analysis revealed that AI-driven demand remains robust but supply-constrained, particularly across semiconductors, data center capacity, and power infrastructure. “Investors also raised concerns around demand visibility and the potential for overbuilding, citing uncertainty around the pace of enterprise AI adoption and monetisation relative to the scale of committed capital,” the financial data and analytics company said.
During 2026, inference capacity will overtake training capacity, and by 2030 it will account for 80% of all AI-critical IT load, said the Iron Mountain-Structure Research study, adding that this would mark a complete reversal of the balance in 2023.
“The cost of artificial intelligence is declining at an accelerated pace whilst the cost of the cheapest LLM has decreased by 10x every year. The declining price will not impact consumption, instead it will drive mass utilisation and innovation,” it added.
