Compute units: the new real estate of the AI boom

Compute units: from capex to rental, the AI gold rush

Compute units have stopped being a capital expense and become the new infrastructure asset. The thesis is attractive and unsettling: clusters are rented out for periods, like offices, and Nvidia sells both the hardware and the software to manage them. Amid so many compute units, some still wryly insist on calling it a computer. But the underlying problem is who ultimately pays the rent at the end of the chain.

The play consists of treating computing capacity like real estate. Funds set up vehicles with money from sovereign wealth funds and pension plans to buy land, build data centers and purchase chips from Nvidia. Compute units are rented out under contracts that generate revenue and can be securitized and sold in packages. This is the neocloud business: reselling capacity the way one sells land.

Meanwhile, the frontier labs (OpenAI, Claude, xAI) keep burning money. Four scenarios are on the table, ranging from profitability to extinction. The most uncomfortable one for the bubble is that no one shows up to pay for subscriptions: Chinese AI is already good enough for everyday use, Google will give away an advanced version to 99% of users, and companies are seeing their cloud spending spiral. Without a final tenant, the rental contract is worthless.

And as if one limit weren't enough, there's energy consumption. Without gallium nitride (GaN) and silicon carbide (SiC) power semiconductors, the party ends due to electricity shortages. The bonanza can hold up as long as the labs can find funding. When it ends, someone will have to explain who pays the rent. Plenty of GPUs, yes, but still no solid tenant.