AI now needs customers to spend like a new economy

There is a number buried inside the AI boom that is much harder to explain than another trillion-dollar valuation.

$636 billion.

That is roughly how much annual AI-related revenue Goldman Sachs estimates the major hyperscalers would need if they want a 30% return on the enormous capital they are pouring into data centers and related infrastructure.

Their current incremental AI cloud revenue is only about $70 billion annualized.

So the industry isn’t looking for another 20% or 30% of growth.

It needs something closer to nine times the current revenue acceleration.

And that is before asking whether the AI companies sitting on top of all this infrastructure make money too.

Goldman’s calculation gets to roughly $1 trillion of annual AI application spending for the hyperscalers to earn solid returns while the application companies maintain healthy margins. If everyone wants returns closer to the historical norms that investors became accustomed to, the required consumer and enterprise spending can approach $2 trillion.

That changes the question.

The AI industry has spent the last several years proving that it can build.

Now it has to prove that the economy can consume.

The numbers are getting enormous.

Goldman estimates the largest U.S. hyperscalers will spend roughly $800 billion on capital expenditures in 2026, with consensus around $1.1 trillion for 2027. The announced revenue backlogs exceed $1.5 trillion, which sounds enormous until you remember that a backlog is a promise of future revenue, while the capital expenditure happens now.

This is where the usual AI discussion gets backwards.

Everyone asks whether AI will eventually become a gigantic business.

Fine.

But how gigantic does it have to become to justify the infrastructure already being built?

At a 30% ROIC target, the answer is not “pretty big.”

It is $636 billion a year.

And if the software companies need their own margins, too, the amount of money businesses and consumers ultimately have to spend starts approaching $2 trillion.

That is not a normal software adoption curve anymore.

That is an enormous new spending category.

Think about what has to happen for that money to appear.

Companies have to decide that AI produces enough additional revenue or saves enough labor and other costs to justify spending tens of thousands, hundreds of thousands or millions of dollars on it.

Consumers have to start paying substantially more for AI products in a world where much of the internet trained them to expect digital services to be free or nearly free.

And the AI companies themselves have to stop subsidizing usage long enough for the revenue to travel back through the chain.

The whole system has to work at the same time.

That is the part I find more revealing than the usual “AI bubble” argument.

The technology can be real.

The demand can be real.

The productivity gains can be real.

And the investment can still outrun the revenue needed to earn an acceptable return on the capital.

Those things are not contradictory.

In fact, that may be exactly what is happening.

Look at what the financing machine is already doing.

The Financial Times reported this week that hyperscalers could borrow as much as $1 trillion through 2030 to finance their data-center expansion. The problem bond investors are increasingly focused on isn’t whether Amazon or Microsoft can repay their debt today. It is how much more borrowing is coming as the companies repeatedly increase their capital-spending plans.

And now Amazon is reportedly considering selling roughly $8 billion of Nvidia chips to an outside investment vehicle and leasing those chips back.

That is a remarkable development.

Amazon would still use the hardware.

But an outside investor would own the assets and finance them through the special-purpose vehicle.

The AI infrastructure remains in operation.

The financing simply moves somewhere else.

That tells you something about the economics without requiring anyone to declare an AI bubble.

The industry is finding increasingly creative ways to finance the machines before the full revenue stream exists.

And that brings us back to the $636 billion.

If the hyperscalers really can generate it, the spending will eventually make sense.

If AI becomes productive enough that businesses willingly spend $1 trillion or $2 trillion a year on applications, today’s infrastructure may turn out to have been necessary preparation.

But if the end market stops far short of that number, the problem isn’t that people suddenly discover AI doesn’t work.

The problem is that the machines may work better than the business model.

That distinction matters.

A technology can be revolutionary and still produce terrible returns for the people who overbuild it.

Railroads changed the world.

Telecommunications changed the world.

The internet changed the world.

That never guaranteed that every dollar invested in building them earned a good return.

Right now, the AI industry is making an enormous capital commitment based on a future in which the rest of the economy spends at an extraordinary scale.

The $2 trillion figure is therefore not a prediction that consumers will spend $2 trillion on ChatGPT subscriptions.

It is the size of the economic demand that eventually has to emerge somewhere in the system if everyone wants to get paid.

And that is the part investors should be arguing about.

Not whether AI is real.

Whether the customer base can become large enough, fast enough, to pay for what everyone is building today.

Not financial advice.