Is it a bubble? The trillion-dollar AI build-out, examined

The largest companies on earth will spend close to a trillion dollars on AI infrastructure this year. A sovereign wealth fund just warned it could lose a third of its value if the bet sours. Both statements are true at the same time, which is why "bull or bear" is the wrong frame. The more useful question, the one a builder rather than a bettor asks, is simpler: when the dust settles, who is still standing?

The spending, in numbers

The five biggest US cloud and AI players, Microsoft, Alphabet, Amazon, Meta and Oracle, have collectively guided to between $660 billion and $690 billion of capital expenditure in 2026, close to double 2025, pushing the wider industry into its first trillion-dollar year of compute capex (Futurum).

The warnings are just as loud. On 18 March 2026, the head of Norway's $2.1 trillion sovereign wealth fund cautioned that an AI bubble could erase up to 35% of its value. Analysts point to structures like Meta's roughly $30 billion Hyperion project, where only about 20% sits on Meta's own balance sheet and the rest is financed through a separate vehicle (Man Group, Goldman Sachs).

What history says a bubble actually destroys

Set aside the verdict for a moment and look at the pattern. Railways, electricity, the telecoms fibre glut of 2000, every one was a genuine bubble, and every one left behind infrastructure the world then used for decades. The mania burned the speculators and the over-leveraged. The tracks, the grid, the fibre survived and quietly powered the next era. Bubbles rarely destroy the technology. They destroy the people who financed it badly.

That reframes the worry. The sceptics aren't wrong that some of this trillion is built on circular financing, capex underwritten by demand commitments from a handful of model labs whose own cash flows are deeply negative for now. But "some of it is fragile" is not the same as "the compute won't be needed." If demand keeps compounding, the asset endures even if particular balance sheets don't.

The two things that survive a correction

If you strip the debate to what protects an operator through a downturn, it comes down to two levers. First, secured demand: capacity sold before it's built, not capacity built on the hope buyers appear. Pre-commitment is the difference between an asset and a liability when the cycle turns. Second, a low cost floor: when revenue per unit of compute falls, the operator with the cheapest power is the last one still profitable, because everyone with a higher cost base goes underwater first.

Notice that neither lever requires predicting whether this is a bubble. They're the right moves in either world. In a boom, low costs and pre-sold capacity maximise margin. In a bust, they're what keeps the lights on while overextended rivals write down. Build for the correction and you win the boom too.