By: Andreas Kuehn and Tara Chawla
The artificial intelligence (AI) infrastructure buildout that is currently underway is one of the largest capital investment cycles in corporate history. The capital expenditure (capex) for global data centers — the corporate spending directed at purchasing, building, or substantially upgrading long-term assets — reached $455 billion in 2024. U.S. hyperscalers have since raised their company-wide 2026 capex plans, much of which is directed toward AI and cloud infrastructure. Amazon is leading with $220 billion, closely followed by Alphabet at $195-205 billion, Microsoft at $175 billion, and Meta at $130-145 billion. Chinese technology firms are investing at a smaller but still substantial scale. Alibaba has announced $53 billion in AI and cloud infrastructure spending over three years, while ByteDance reportedly plans $23 billion in capex for 2026; much of it for AI infrastructure. Annual data center spending is projected to exceed $1 trillion by 2026.
These eye-popping financial investments might grab headlines, but the global AI infrastructure buildout changes how AI competition should be understood. The relevant unit of analysis is no longer only who trains the most advanced AI language models, or designs and fabricates leading-edge AI accelerators, but who can assemble and control the full AI stack, including data centers, electricity generation and transmission, and communications networks.
The magnitude of the spending becomes clearer when looking at the cost of IT infrastructure. The average selling price of an AI server was nine times that of a traditional server in 2024. A single AI server rack costs between $1.5 million and $4 million, and a leading AI data center can house up to 10,000 such racks. A rack can contain 8-18 compute trays, each housing one or two AI accelerators, primarily graphics processing units (GPUs). These specialized chips are optimized for parallel processing and are essential to training and running AI models. The AI buildout involves not only more racks, but much more expensive, densely packed, and power-consuming IT equipment. A single training cluster can include more than 100,000 GPUs working in concert. A networked cluster of 100,000 Nvidia H100 GPUs, an example of frontier-scale AI training infrastructure, can cost more than $5 billion.
Data centers are the physical facilities that house this equipment. Roughly 12,000 data centers operate worldwide, with the United States hosting about 4,500-5,000 facilities and nearly half of global capacity. They generally fall into three categories: enterprise data centers operated for a company’s own needs; colocation and service provider facilities that host equipment for multiple customers; and hyperscale data centers operated by major cloud and technology firms. Their design depends on the intended workload — the type of work they perform. Conventional data centers are typically CPU-heavy and support enterprise applications, web hosting, email, database, virtual machines, and file storage. In contrast, AI data centers are built around GPUs and other accelerators, optimized for AI training and inference.
The current wave of data center expansion accelerated in 2022 and 2023, particularly in the established cloud and data center corridors of the United States, including Northern Virginia, Texas, the Pacific Northwest, and the Midwest. These regions remain central because they already possess fiber connectivity, experienced operators, land, and power infrastructure. U.S. hyperscalers and other operators are also expanding internationally to serve new customers, reduce latency, improve resilience, and comply with data-localization and digital-sovereignty requirements. In 2025, for example, Google announced a $15 billion investment in an AI hub in Visakhapatnam, India, combining gigawatt-scale data center capacity with new energy and subsea connectivity. Also in 2025, AWS and Saudi Arabia’s HUMAIN announced a joint investment of more than $5 billion to build an ‘AI Zone’ in Saudi Arabia. This would allow more workloads to be processed locally rather than through AWS regions abroad. China, meanwhile, is pursuing a more state-directed buildout. Beijing is reportedly preparing plans to spend $295 billion on data centers over five years, and is expected to rely increasingly on domestic chips, a consequence of U.S. export controls and Beijing's push to prioritize indigenous technology.
The massive global AI infrastructure expansion is no longer confined to conventional terrains or data center facilities. Hyperscalers are directly investing in electricity generation and transmission, as well as international connectivity. Meta announced Project Waterworth in 2025, a multi-billion-dollar, 50,000 kilometer subsea cable system, connecting five continents. More speculative proposals envision extending the buildout into space. SpaceX filed plans with the U.S. Federal Communications Commission to operate a low Earth orbit data center constellation consisting of up to one million satellites.
The AI buildout is not simply a race for more advanced accelerators, but an infrastructure-centered contest to turn technology, energy, and capital into usable computing capacity. Chips matter only when installed, powered, cooled, networked, and connected to customers. As bottlenecks shift from chips toward power, land, and connectivity, advantage will accrue to states and firms capable of coordinating investment and supply chains across the full AI stack. Data centers and their supporting networks are consequently becoming strategic infrastructure, with implications for economic security, resilience, and national policy. The decisive measure will not be announced spending or planned capacity, but how quickly those commitments become reliable, operational computing capacity.
Dr. Andreas Kuehn is a Senior Fellow for the Cyberspace Cooperation Initiative and Tara Chawla is a Summer 2026 Intern for the Technology Policy program at ORF America.