What is the building making?
Call a building a data center and people picture rows of racks. Call it an AI factory and the picture moves to output: data arrives, processors train or run models, and something useful leaves. That shift of attention is the phrase’s appeal.
NVIDIA uses the term for purpose-built infrastructure and software for AI, joining accelerated computing, networking and the software stack into one approach. It is a company describing its own products. Facilities that carry the label can still differ widely in architecture, operating practice and capability.
The term has also reached independent analysis. The International Energy Agency’s April 2026 report applies it to “cutting-edge data centres specifically designed for AI”. Its satellite tracking of 21 large AI data centers under construction in the United States found that their capacity, estimated from completed floor space, tripled within 18 months to more than 6 GW.
The work around the processors
The factory image helps most when it points to the work around the machines. Data has to be prepared, models evaluated and deployed systems monitored. Cisco’s explanation of the term covers this wider workflow, from data intake to inference and continued improvement.
A kitchen with an expensive new oven still waits if ingredients are not reaching the counter. In the same way, useful output can depend on the data pipeline, storage, network and software as much as on the accelerators. Where the bottleneck sits varies from one workload to the next.
Training and inference ask different things, too. Training tends to reward fast progress through a large distributed job. Inference has to deliver dependable responses at a latency and cost the service can support.
The rest of computing
Data center remains the larger category. These buildings also support applications, storage and connectivity for enterprise IT, cloud services and colocation, and one building can host AI alongside other workloads.
The IEA’s figures show both sides. It estimates that data centers used 485 TWh of electricity in 2025, and that consumption by AI-focused data centers grew 50% that year, against 17% for data centers overall. In its Base Case for 2030, AI-focused facilities reach around 465 TWh out of roughly 950 TWh: about half, with the rest in data centers that are not AI-focused.
Visual guide
How much data center electricity is AI-focused?
Choose a step, or select a part of the diagram.
- Total in 2025. The IEA estimates that data centers used 485 TWh of electricity in 2025. This total is shown without an AI split.
- Growth in 2025. By the IEA’s estimate, electricity use by AI-focused data centers grew 50 per cent during 2025, against 17 per cent for data centers overall.
- AI-focused in 2030. In the IEA’s Base Case projection for 2030, AI-focused data centers use around 465 TWh out of roughly 950 TWh, about half.
- The rest in 2030. In the IEA’s 2030 Base Case, the rest of the roughly 950 TWh is in data centers that are not AI-focused. AI-focused facilities are growing far faster, yet about half of data center electricity is still used outside them.
Sources
- AI FactoriesNVIDIA · Accessed 23 October 2026
- What is an AI factory?Cisco · Accessed 23 October 2026
- What is a data center?IBM · Accessed 23 October 2026
- Key Questions on Energy and AIInternational Energy Agency · April 2026
Company sources describe their own work and terminology. Our editorial approach
