Feel free to improve the article, but do not remove this notice before the discussion is closed. For more information, see the guide to deletion. Find sources:"AI datacenter"–news·newspapers·books·scholar·JSTOR%5B%5BWikipedia%3AArticles+for+deletion%2FAI+datacenter%5D%5DAFD
An AI datacenter (or artificial intelligence datacenter) is a specialized data center facility designed explicitly to support the high-performance computing (HPC) workloads required for the training and inference of artificial intelligencemodels.[1] Unlike traditional datacenters that host a variety of general-purpose computing tasks (like web services and databases), AI datacenters are optimized for the unique computational demands of machine learning, particularly deep learning.[2][3] These facilities are characterized by extreme power density per rack (often exceeding 50–100 kW)[4], advanced liquid cooling systems, and low-latency, high-bandwidth networking fabrics to facilitate parallel processing. [5][6][7]
The rise of generative AI since 2022 has triggered a global boom in the construction of AI datacenters, making them a critical and strategically important piece of national infrastructure.[8] Companies like Microsoft, Google, Meta, and Amazon are investing tens of billions of dollars to build facilities containing over 100,000 AI accelerators each.[9][10][11] This massive build out is causing a resurgence in nuclear power plants.[12] In 2025 Google spent $95 billion on capex, with much of that for AI datacenters.[13][14] In 2025 U.S. tech companies spent $370 billion on capex with much of that spending for AI datacenters.[15] OpenAI wants to create a process for new datacenter expansion every week.[16][17]
By 2026, AI data centers are projected to consume over 90 TWh of electricity annually.[18] In the U.S. energy use is growing by 33% per year attributed to AI datacenter growth.[19] A startup company was created to harness nuclear energy for the AI datacenter boom.[20][21] The largest AI datacenter in 2025 cost $7 billion , and uses 300 MW of power—as much as 250,000 households.[22] Cornell University study estimates that the AI datacenter build out between 2024 and 2025 will contribute between 24 and 44 metric tons of addition CO2.[23]
The Trump administration is promoting the build out of AI datacenters.[35] U.S. president Trump hints at deregulation to promote AI datacenters.[36][37] There is local opposition.[38][39]
History
Early AI workloads in the 2010s were often run on general‑purpose high‑performance computing (HPC) clusters or small GPU servers in conventional data centers.[40] As deep learning models and datasets grew, cloud providers began to build dedicated infrastructures for AI training, including GPU clusters exposed through services such as Google Cloud TPU[41], Amazon EC2 P-series instances[42], and Microsoft Azure’s ND‑series virtual machines[43].
Around 2022–2024, the rapid adoption of large language models (LLMs) and generative AI led to a surge in demand for specialized AI datacenters with thousands or tens of thousands of accelerators connected through high‑speed fabrics.[44] Several technology companies announced multibillion‑dollar investments in new or expanded AI‑focused campuses, often near abundant power supply or renewable energy sources.[45][46][47]
Architecture
AI datacenters are designed around clusters of accelerators optimized for parallel numerical computation. A typical facility includes: Compute: Large numbers of GPUs, TPUs, or other AI accelerators are deployed in high‑density racks. These devices are often grouped into “pods” or “nodes” that share local networking and storage and can be scaled out to thousands of accelerators for distributed training.
Networking: AI training jobs require high‑bandwidth, low‑latency communication to exchange gradients and parameters across devices. To support this, AI datacenters commonly use technologies such as InfiniBand, RDMA over Converged Ethernet, or proprietary interconnects. Storage and data pipelines: Training large models requires feeding vast datasets at high throughput. Control and orchestration: Software stacks manage job scheduling, resource allocation, and failure handling. Common AI frameworks include PyTorch and TensorFlow.
Technical architecture
Compute Density: Traditional server racks typically consume 5–15 kW of power.[48] AI racks, utilizing hardware like Nvidia H100s or Google TPUs, consume 40 kW to over 100 kW per rack.[49]
Cooling Systems: Due to the heat generated by high-density chips, AI data centers often abandon traditional air cooling (CRAC units) in favor of liquid cooling technologies, such as direct-to-chip cooling or immersion cooling.[50]
Networking: AI training requires thousands of chips to communicate simultaneously. This necessitates specialized non-blocking network architectures rather than standard Ethernet used in web servers.[51]
OpenAI
OpenAI's datacenter project is called Stargate.[52] The partnership includes Softbank, MGX and Oracle. In September 2025, it was announced the building of 5 new AI datacenters.[53] The new sites are in Texas, New Mexico, Wisconsin, and Ohio. Capacity is estimated to be 6.5 gigawatts and $400 billion investment over 3 years. OpenAI's revenue for 2025 was estimated to be less than $12 billion.[54]
The impact of these new AI datacenters is sparking concern.[55]
↑Coates, Adam; Huval, Brody; Wang, Tao; Wu, David; Catanzaro, Bryan; Andrew, Ng (2013-05-26). "Deep learning with COTS HPC systems". Proceedings of the 30th International Conference on Machine Learning. PMLR: 1337–1345.
This page is based on this Wikipedia article Text is available under the CC BY-SA 4.0 license; additional terms may apply. Images, videos and audio are available under their respective licenses.
