What Is Actually Inside the AI Data-Centre Buildout?
“Data-centre capacity” sounds like one thing. It is not. Conventional cloud workloads, enterprise systems and AI-optimised compute can share the same buildings while demanding very different servers, power densities, networking and cooling.
AI is still only part of the installed base.
JLL estimates AI represented about 25% of global data-centre workloads in 2025, with traditional workloads still accounting for the majority. By 2030, JLL expects AI could reach roughly half of workloads.
JLL 2026 Global Data Center Outlook. These are workload-capacity estimates, not a universal physical split of every facility.
Cloud and AI are not cleanly separate buckets.
A hyperscaler can run AI inference inside its cloud platform. The same workload can therefore be described as “cloud” in one dataset and “AI” in another. Source definitions matter.
An earlier Goldman Sachs baseline split used 54% cloud, 32% traditional workloads and 14% AI. Useful as a mental model, but not a current universal market share. Goldman projected AI's share to rise materially over time.
Your instinct that roughly half of the historical workload base is cloud is reasonable in at least one major published framework.
It does not mean 54% of every data centre is “cloud space” and another separate 14% is “AI space.”
AI is mainly a compute-and-power story, not a storage story.
AI systems need accelerated processors, high-bandwidth memory, very fast networking, large electrical feeds and increasingly liquid cooling. Storage remains important, but it is not the defining component of AI infrastructure.
Runs training and inference
Keeps model data close to compute
Links thousands of accelerators
Supports much denser racks
AI-optimised compute is driving the incremental power growth.
Gartner's 2026 estimate makes the change visible. Conventional-server electricity demand is almost flat, while AI-optimised server demand jumps sharply.
In Gartner's estimate, almost all of the server-side electricity growth from 2025 to 2026 comes from AI-optimised systems, not conventional servers.
Gartner estimates AI-optimised servers account for about 31% of total global data-centre electricity consumption in 2026. Cooling and other infrastructure consume power too, so workload share and power share are not the same thing.
Not according to broad vacancy data.
CBRE reported global data-centre vacancy of 6.7% in Q1 2026 across the 16 largest markets it tracks. That suggests the current market is tight rather than sitting on large amounts of visibly unused space.
But vacancy is a real-estate measure. It does not reveal GPU utilisation, how much future AI capacity is pre-committed, or whether announced 2027–2030 capacity will ultimately be absorbed.
Will downstream demand absorb what is being built?
That cannot be answered from megawatts alone. We need to watch whether adoption broadens, deployment gets deeper, businesses keep paying, usage intensity rises and economic value becomes more visible.
The explainer describes the infrastructure side. The monitor follows the demand side. Together they form the AI Infrastructure & Demand inquiry.
Directional, not a single definitive market split.
Different sources define “workload,” “capacity,” “server type,” “power consumption” and “AI infrastructure” differently. The figures below should not be combined into one precise market-share calculation. They are used to build a more useful mental model.
AI workload share and 2030 outlook ↗
Cloud / traditional / AI workload framework ↗
Electricity consumption by conventional and AI-optimised servers ↗
AI infrastructure server vs storage spending ↗
Global vacancy and available-capacity context ↗