Worldwide data center capital spending is projected to surpass $3 trillion by 2030, according to a new forecast from Dell’Oro Group, as hyperscalers, sovereign AI programs, and specialized cloud providers continue expanding infrastructure to support AI workloads. The research firm says its 2030 outlook has nearly doubled since its January 2026 forecast, reflecting higher hyperscaler spending guidance, larger estimates for global data center power capacity, and rising commodity costs.
Baron Fung, vice president at Dell’Oro Group, said AI accelerators are expected to account for roughly a third of the projected $3 trillion in capex. That figure does not capture the full cost of AI infrastructure, since operators also need servers to host the chips, specialized networking for AI clusters, and storage for training and inference workloads.
The forecast assumes global data center power availability will grow to more than 200 GW. Dell’Oro estimates the four largest US cloud providers could account for about half of global data center capex, while its AI-specialized cloud category, which includes model developers and neocloud providers, is projected to grow at a compound annual rate of nearly 60 percent. General-purpose server demand is also expected to rise as inference, agentic AI, and storage needs expand.
Hyperscaler Buying Power Could Squeeze Others
Large cloud providers are using long-term supplier agreements to secure preferred pricing and capacity commitments, which Fung said could reduce component availability for other buyers, extending lead times and raising prices elsewhere in the market. Growing use of custom chips and architectures by hyperscalers is lowering their own costs while pressuring server manufacturer pricing more broadly.
Fung expects many enterprises to adopt a hybrid approach, keeping variable or incremental AI demand in the cloud while shifting stable, heavily utilized workloads on-premises once ownership becomes more cost-effective than renting GPU capacity.
Cooling and Power Needs Grow With Density
Gordon Johnson of Subzero Engineering said higher-density AI systems require more than additional accelerators. AI workloads draw more electrical power per rack than traditional computing, often requiring a combination of cooling strategies. He recommended operators add capacity in stages, identify where high-density compute is actually needed, and integrate new cooling methods with existing infrastructure, warning that building too much too quickly is the biggest risk.
Luke Edney of Norton Rose Fulbright said power availability is becoming the primary constraint on AI infrastructure expansion, with some campuses now designed at gigawatt scale. Grid connection timelines in established markets can exceed five years, pushing developers toward locations with surplus renewable generation and utilities open to proactive partnerships. Edney said enterprises must weigh the risk of falling behind against the financial exposure of committing to infrastructure too early, adding that AI investment should be tied to measurable business outcomes rather than market pressure alone.