Google is putting a data center in orbit next week. Not a sensor. Not a camera. A server that runs AI, floating above the Earth. On October 1, a SpaceX Falcon 9 will carry Google's satellite, named MVP, from Vandenberg Space Force Base in California. It holds four of Google's custom AI chips, called TPUs. This is Project Suncatcher. The goal is wild: AI data centers in space, powered by the sun.
Google will become the first tech heavyweight to take this step. Others talk. Google is launching. The satellite is a prototype, a single server in a metal box. But it is the first real move toward outer-space compute by a company you know.
I love this. I do. The engineering is clean. The ambition is clean. You point a dish at the sun and you run models where there is no grid to fight you. It feels like science fiction that decided to show up early.
And then the bill arrives.
Tech firms are taking on massive debt for the sake of AI. Not small debt. Not careful debt. The commitments sit off the balance sheet, in leases and purchase contracts that do not count as liabilities until the money is actually spent. A Wall Street Journal analysis found roughly $3 trillion in these hidden obligations across nine major tech companies. That is triple their reported debt. Meta signed a 20-year lease for a data center in Louisiana but does not show the full risk on its books. Oracle's AI debt packages were so large that banks struggled to spread the risk.
This is not a sideshow. It is the engine room of the AI boom. And the engine room is running hot.
A new projection from economist Stijn van Nieuwerburgh, published by the Brookings Institution, estimates total AI-infrastructure investment at $10.3 trillion from 2025 to 2032. That is an average of 3.6% of U.S. GDP every year. By that measure, this buildout would beat the investments in railroads, electrification, highways, and the internet. Never before has the U.S. economy leaned this hard on a single industry.
Three point six percent of GDP. Every year. For eight years.
Say it out loud. It does not sound like a normal tech cycle. It sounds like a national project. It sounds like the kind of bet that changes prices, wages, and what gets built next. The Wall Street Journal notes that data-center spending is already greater than that for canals, railroads, and the grid combined. It is creating jobs and stock-market wealth. It is also boosting inflation and crowding out housing.
I get why this is happening. AI needs power. AI needs space. AI needs chips that do not exist yet. The ground is full. The grid is full. So you look up. You put servers in orbit. You finance the rest with leases that do not show up on the balance sheet until they have to. You tell yourself the demand will be there. You tell yourself the models will keep getting better. You tell yourself the revenue will cover the debt.
But what if it does not?
The catch is simple. The scale of this feels sudden and large relative to familiar investments like railroads or highways. Those booms took decades. This one is compressed into eight years. The financing is clever. Clever financing works until it does not. When Oracle's lenders could not find enough buyers for its AI debt, the deals stalled. When Meta's joint venture raises $27 billion in bonds to build a data center, that debt sits with the landlord, not Meta, even though Meta promised to pay for 20 years. This is not fraud. It is structure. And structure can hide risk until the moment it cannot.
I am not saying this will crash. I am saying the size of the bet matters. $10.3 trillion is not a product launch. It is a reshaping of capital flows. If AI demand slows, if models hit a wall, if power constraints bite harder than expected, the leases do not vanish. The purchase commitments do not vanish. The satellites do not come back down.
Google's satellite is a beautiful first step. It is also a symbol of how far this has gone. We are not just building faster chips. We are building an economy around them. We are putting servers in space because the Earth feels too small. That is thrilling. It is also a warning.
Wonder without the catch is advertising. The wonder here is real. The catch is the debt, the scale, the speed. The question is not whether AI will work. The question is whether the bill can be paid without breaking something else.
I want this to work. I want the models to keep getting better. I want the power to show up. I want the satellites to hum in orbit and answer simple queries from space. But I also remember how it felt the last time a boom turned into a burden. The enthusiasm is honest. The concern is honest too.
More on: Tech