What the Buildout Is For

What it is

The Use-Case Question

The largest technology companies are spending hundreds of billions of dollars a year on data centers before most of the revenue has arrived. Whether that spending is foresight or a bubble is the most-argued question about the buildout.

• Demand case: cloud backlogs at Microsoft, Google and Oracle total about $1.9T of signed contracts. All three say demand exceeds capacity. GPU rental prices rose in 2026 rather than collapsing, OpenAI and Anthropic revenue run-rates reached about $70B and $65B, and Google processes more than 3 quadrillion tokens a month. Randomized trials found 14-26% productivity gains in coding and customer support.
• Bubble case: spending runs several times disclosed AI revenue, much of the demand is circular (Nvidia invests in OpenAI, which buys Nvidia systems), and debt is rising: Oracle's credit spreads hit an 18-year high and Meta financed part of Hyperion off its balance sheet. Other trials found experienced developers slower with AI. The IMF and Bank of England have both warned of a correction.
• The railroad argument: builders point out that rail, electricity and the internet were all built before their uses could be named; refrigerated shipping, mail-order retail and standard time zones followed the railroad by decades. Daniel M. Sisson's Inside the Singularity (2026), by the publisher of DatacentersX, makes this case, while granting that resistance to the railroad was "not the product of misinformation." Skeptics answer with the same history: US railroads doubled their mileage from 1865 to 1872, then crashed in 1873, and a third of railroad bonds defaulted. Fiber laid in the 1990s was 97% dark in 2002, and still carries the internet today.
• Uses named today: coding agents, research and analysis, customer support, scientific tools such as AlphaFold, autonomy and robotics, and military and intelligence work. Critics point to low-quality AI-generated content, surveillance and job losses.
• China: DeepSeek's January 2025 release, trained cheaply, became the main argument that efficiency could shrink demand; the other side cites the Jevons paradox, where cheaper computing raises total use. Chinese reports said much newly built compute sat idle, and the NDRC moved to curb uncoordinated local projects.

Flagship

More Than $800 Billion of Spending, $170 Billion of AI Revenue

• Spending: Amazon, Alphabet, Microsoft, Meta and Oracle guided to about $810-840B of capital spending for 2026 or their current fiscal year after raising forecasts in July, up from $381B for the first four in 2025.
• Revenue: disclosed AI run-rates are about $70B at OpenAI, $65B at Anthropic and $37B at Microsoft, roughly $170B, and the figures overlap because lab spending is cloud revenue.
• Gap: Sequoia's David Cahn estimates the industry needs about $3T of revenue to justify 2026's roughly $1.5T of AI infrastructure spending.
• Backlog: against that, Microsoft, Google Cloud and Oracle report about $1.9T of signed future contracts, about a third of Microsoft's tied to OpenAI.
• Scale: Pulitzer-winning historian Liaquat Ahamed puts AI infrastructure at 2-3% of US GDP, against about 5% at the railroad peak before 1873.

Evidence

Spending, Demand and Doubts

Related coverage

Related Coverage

• The Pace Debate
• Business Models and Ownership
• Compute Operations
• Public Sentiment