AI Data Centre Capex Forecast Passes $3 Trillion By 2030
Data Center Knowledge reported that Dell'Oro Group now expects global data centre capital spending to exceed $3 trillion by 2030 as AI accelerators, cloud concentration and power constraints reshape infrastructure plans.
Global data centre capital spending is projected to exceed $3 trillion by 2030, Data Center Knowledge reported, after Dell'Oro Group nearly doubled its January forecast for AI infrastructure, power capacity and commodity costs.
The new outlook puts hyperscalers, sovereign AI programs and specialized cloud providers at the center of the buildout.
Dell'Oro vice president Baron Fung tied about one-third of the expected spending to AI accelerators, but the forecast treats chips as only part of the infrastructure cost.
Accelerator purchases still need servers, specialized networking and storage before they become usable training and inference capacity.
Dell'Oro's model uses a global power-availability assumption above 200 GW and places roughly half of worldwide capex with the four biggest US cloud providers.
The AI-specialized cloud category, covering model developers and neocloud providers, carries an expected CAGR of almost 60%.
Demand for general-purpose servers remains part of the forecast because inference, agentic AI and storage workloads add compute needs beyond training clusters.
Large cloud providers are using long-term supplier agreements to secure preferred pricing and capacity commitments.
Those arrangements could reduce component availability for other buyers, lengthen lead times and push prices higher, while custom chips and architectures give the largest operators lower unit costs at scale.
For enterprise buyers, that scale changes the ownership decision.
Fung expects many companies to keep variable or incremental AI demand in the cloud while moving stable, heavily used workloads on-premises once ownership becomes cheaper.
The physical data centre has to absorb the same shift.
Subzero Engineering senior CFD manager Gordon Johnson pointed to higher electrical power per rack and the need for mixed cooling strategies, with accelerator deployments forcing upgrades across electrical distribution, liquid-cooling loops, airflow containment and adjacent building systems.
Johnson's operating caution was staged capacity.
Operators need to identify where high-density compute is required before adding too much too quickly, because the cooling and power architecture has to fit the actual workload rather than the broad market rush.
Power availability remains the biggest constraint.
Norton Rose Fulbright partner Luke Edney identified procurement cost and timeline risk as central site-selection factors, with grid connection timelines in established markets stretching beyond five years.
Those delays are steering developers to sites with unused renewable supply, more flexible interconnection rules and power companies prepared to work with them before construction.
The same low-carbon reliability requirement is bringing on-site generation, fuel-cell systems and small modular reactor options into planning discussions.
The forecast leaves AI infrastructure expansion tied to a financing question as much as a chip-supply question.
Edney's caution was that enterprises still need measurable business outcomes before early commitments turn into stranded power, cooling and capacity bets.
