AI is often discussed as software: models, chatbots, apps.

But behind the software sits something much more physical. Data centers the size of small towns. Power plants and transmission lines. Warehouses full of specialized chips.

A new paper by economist Stijn Van Nieuwerburgh tries to put a price on that physical layer. Its estimate: building 188 gigawatts of AI capacity in the United States by 2032 would require nearly $9 trillion.

That is roughly 3.2% of US GDP every year from 2025 to 2032.

At that scale, the question isn’t just what AI can do. It’s who pays for it, and who carries the risk.

One gigawatt of AI capacity costs about $41 billion

The paper starts with a single building block: a one-gigawatt AI campus.

A gigawatt is a large amount of electricity, roughly the output of a large power plant. An AI campus of that size combines data center buildings, power infrastructure and the computing equipment inside.

The paper estimates that one such campus costs about $41 billion.

The 188-gigawatt scenario then scales this up to the size of the national buildout that the industry’s plans imply.

Tech giants’ spending now exceeds their cash flow

For years, the largest technology companies, the “hyperscalers” that run giant cloud platforms, could pay for data centers largely out of their own profits.

The paper finds that this is no longer the case. Their capital spending now exceeds their internal cash flow.

So the money has to come from elsewhere. The paper describes financing moving toward:

  • leases, where someone else owns the data center and the tech company rents it
  • joint ventures, where costs and ownership are shared with partners
  • project debt, borrowing tied to a specific facility rather than the whole company
  • private credit, loans from investment funds rather than banks
  • securitization, bundling future payments into tradable investments
  • special-purpose vehicles, separate legal entities set up to own and finance specific assets

Each of these lets more money flow into AI infrastructure than the tech companies could provide alone.

The new financing moves risk rather than removing it

The paper’s central point is that these arrangements don’t make risk disappear. They move it.

Several risks stand out:

Debt piles up at the asset level. A data center held in a separate entity can carry a lot of borrowing, even if the tech company using it looks financially conservative.

Obligations become harder to see. Leases and joint ventures can create commitments that don’t show up clearly as debt on a company’s balance sheet.

A few tenants carry a lot of weight. Many facilities depend on a single large customer. If that customer pulls back, the investors behind the building are exposed.

The equipment ages quickly. AI chips can become outdated within a few years. If the technology moves on faster than expected, assets may be worth far less than the loans secured against them.

Construction can be delayed. Large projects depend on power connections, permits and supply chains, any of which can slip.

No one knows what the buildings will be worth later. If demand for AI computing falls short, the value of these facilities after current contracts end is highly uncertain.

Big technologies have always needed new kinds of finance

There is a broader pattern here.

Large technological ambitions have often required new financial structures. Railways, electrification and telecommunications networks all drew in investors, lenders and financing arrangements far beyond the companies that built them.

Those structures made the buildouts possible. They also spread the consequences widely when expectations ran ahead of reality.

The AI buildout appears to be following a similar path. As spending outgrows what any single company can fund, the investment and the risk are distributed across a wider set of investors, including pension funds, insurers and private credit funds that may never use the technology themselves.

The $9 trillion is a scenario, not a forecast

The headline figure needs care.

The $9 trillion is the cost of a specific scenario: a 188-gigawatt buildout completed by 2032. It is not a forecast that this spending will happen.

The actual figure could be very different. Chip prices, construction costs, energy supply and, above all, demand for AI could all change. If AI usage grows more slowly, or if more efficient models need less computing power, the buildout could be smaller. If demand grows faster, it could be larger.

The paper is an NBER working paper and has not yet been peer reviewed, although it is forthcoming in the Brookings Papers on Economic Activity.

It is also a framework for understanding AI investment, not investment advice.

The AI boom is being built in contracts as well as labs

Most people experience AI through a chat window. This paper is a reminder that the chat window sits on top of one of the largest infrastructure projects of the century.

How that project is financed will shape who profits if AI delivers, and who absorbs the losses if it doesn’t.

The AI boom is being built not only in labs, but in contracts.