Gartner forecasts $788 billion in data center systems spending in 2026. That number is infrastructure for AI, and infrastructure has always been more political than it looks.
There is a building going up in the Virginia suburbs that will draw more power than a small city. It will require water for cooling measured in millions of gallons per year. It will house tens of thousands of specialized chips, most of them manufactured in Taiwan, assembled into racks by a supply chain spanning four continents. And from the outside it will look like a very large beige box.
Gartner forecasts that worldwide data center systems spending will reach $788 billion in 2026, as part of total IT spending of $6.31 trillion. Public cloud end-user spending hit $723.4 billion in 2025. Worldwide AI spending is forecast at $2.5 trillion in 2026. These numbers describe a construction project at civilizational scale. The internet needed infrastructure: cables, routers, server farms. AI needs more. Orders of magnitude more compute, orders of magnitude more power, and a chip supply chain that is currently running at full capacity and still cannot keep up with demand.
The comparison to oil infrastructure is not decorative. In the early twentieth century, oil refineries were the critical chokepoint in the industrial economy. Whoever controlled refining capacity controlled the downstream industries that ran on oil. Today, GPU clusters are the chokepoint in the AI economy. Large language models do not exist without training runs that require thousands of specialized processors running for weeks. Inference, the process of actually using a trained model to answer questions, requires sustained compute at massive scale. The organizations that control this infrastructure, the hyperscalers and a few specialized AI companies, occupy a structural position in the AI economy that resembles the position the major oil companies held in the industrial economy.
Amazon, Microsoft, and Google are investing at a pace that would have seemed implausible five years ago. Each has announced capital expenditure plans running into the tens of billions annually. NVIDIA's data center revenue has grown at rates that no semiconductor company has ever sustained for this long. The demand for GPU supply has outrun TSMC's ability to scale manufacturing. These are not technology sector stories in isolation. They are industrial policy stories, energy policy stories, and national security stories dressed in technology language.
The geopolitics of where data centers are built matters in ways that the industry has only recently started taking seriously. Compute sovereignty is becoming a real political concept. The European Union has regulations requiring that certain categories of data be processed within EU territory. China has developed domestic AI chip programs precisely because it cannot rely on TSMC-manufactured NVIDIA chips in an era of export controls. India is building out data center capacity as part of a digital sovereignty strategy. The United States has restricted semiconductor equipment exports to limit the compute access of strategic competitors. A data center is not just a building. It is a geopolitical statement about who controls the processing of information in a given territory.
The energy dimension is the one I think will become a crisis before it becomes a policy priority. Large data centers in the United States are already straining regional power grids. Northern Virginia, which hosts the largest concentration of data centers in the world, has seen grid capacity become a constraint on further development. The AI buildout is landing on an electrical grid that was not designed for this load and cannot be upgraded at the speed the AI industry is adding capacity. The water used for cooling is becoming a scarcity issue in some markets. The carbon footprint of training large models is large enough that several major AI labs have published environmental impact disclosures, though the numbers vary enough across reporting methodologies that comparing them is difficult.
From an IS research perspective, the interesting question is not just how much is being spent, but what kind of institutional foundation this spending is creating. Infrastructure is not neutral. The choices made in building infrastructure, where to locate it, who controls it, what standards govern it, who can access it and on what terms, shape what is possible on top of it for decades afterward. The electrical grid shaped twentieth-century industrial geography in ways that are still visible. The internet's architectural choices, including its decentralized design and the TCP/IP protocol, shaped what kinds of applications were easy to build and what kinds were difficult. The AI infrastructure being built now will shape what kinds of AI applications are possible, for whom, and at what cost, for a long time.
Flexera's 2025 State of the Cloud report found that 27 percent of cloud spending is estimated as waste, and 84 percent of organizations report struggling to manage cloud costs. The concentration of AI compute in a small number of hyperscaler environments creates a similar dynamic, with a structural dependency layer added. Organizations that build applications on top of hyperscaler AI infrastructure are not just paying for compute. They are accepting a governance relationship with a small number of very large companies whose interests may or may not align with their own. That is a governance question that IS research is well positioned to investigate, and I do not think the field has caught up with the pace of the buildout.
The $788 billion is not a technology metric. It is an infrastructure metric, and infrastructure is always more political than it looks from the outside.
About the author
Share
More notes
Related notes