Elon Musk did not predict the end of artificial intelligence on Sunday. He predicted the end of a particular kind of company: the one whose product is a data center and whose customer is the boom.
More than 200 firms in AI computing infrastructure could fail or be consolidated, he said, as financing gets harder for businesses without significant revenue beyond the halls of chips they borrowed to build. Chamath Palihapitiya expects 2027 to bring the usual sequence — dead investments, capacity that never clears its cost of capital, and a smaller industry than the pitch decks promised. That is not mysticism. It is what happens when a herd finances the same idea with debt.
The rebuttal sounds reassuring and answers a different question. Nvidia’s Jensen Huang says demand for neocloud services is very high and that working capital, not customer interest, is the constraint. Executives at Nebius and Iren say demand exceeds capacity into 2028. Take them at their word. A full order book is not a solvent company, and “working capital is the constraint” means the firm dies when the next loan gets expensive. Demand and distress are roommates in any boom financed faster than revenue.
Some of that demand is also circular. An enterprise paying to run a real workload is demand. A startup renting GPUs with money raised from investors excited about GPU rental, to serve customers living on the same AI capital, is a loop. Loops look like growth until the financing window narrows. Musk’s line — firms without significant revenue beyond AI data centers — is the line that matters. Huang can be right about appetite and still be describing businesses that cannot stand when appetite has to be financed at a market rate.
Musk is talking his book. SpaceX competes in rented computing capacity. He gains if overextended rivals fold and capital stops carrying his competitors’ slide decks. Discount him for that. Do not insult yourself by deciding the incentive makes the arithmetic optional.
If consolidation comes, people who mistook a capital cycle for a birthright will call it a crisis. It isn’t. Power, chips, and talent are spread across hundreds of firms that may not earn what their lenders were promised. Moving those inputs to operators who can service the debt is the mechanism working. A sector in which nobody may fail is not competition. It is a queue.
The antitrust reflex gets this backward. Keeping two hundred undercapitalized landlords of rented GPUs alive does not protect the public from monopoly. It protects a financing structure. Excess capacity, if that is what was built, should push prices down and put hardware back on the market — the opposite of a choke point. If a survivor later abuses a real bottleneck, name it and deal with it. Insolvency is not a market structure. It is a loss.
That loss is where politics will try to climb in. “Two hundred American AI companies may fail” is one press release away from strategic infrastructure, China, jobs in your district, and a request: guarantees, credits, pressure to roll the paper. Both parties know the script. A faction of the Right has learned to treat the word China as a credit enhancement — any industry in a threat briefing becomes entitled to the taxpayer as a silent partner. Competing with Beijing is a real duty. It is not a put option for equity that bet on rented GPUs and no customers.
The compute that matters for national power will not be saved by keeping every neocloud on life support. It sits with firms that have balance sheets and with the few operators who can raise money because someone pays them in revenue. Letting a leveraged middleman fail does not hand the future to the Chinese Communist Party. Writing the lender a public check and calling it patriotism would. A serious strategic policy is unglamorous: permit power plants and transmission, stop treating energy abundance as a zoning heresy, and keep adversary capital from scooping distressed American compute at a bankruptcy auction. A rescue for firms with no significant revenue beyond the data center is moral hazard in a flag pin.
Towns cutting abatements should read the warning more carefully than the deal sheet. A data center recruited with waived taxes and a grid upgrade the public will pay for is a public bet on a private capital structure. When the structure fails, the substation stays.
Financing ought to get harder for businesses that cannot exist without it. The scandal is not consolidation. The scandal would be discovering, once again, that in American politics a bad boom never has to end if enough people can spell national security.
How it may affect me
Broad stock and bond funds can already hold this risk through chipmakers, hyperscalers, and private-credit funds that financed data-center debt — not only through famous AI stocks. You will not vote on a rescue; you would pay as a taxpayer or a ratepayer. On local data-center deals, tax abatements and grid upgrades can remain on your bill after a project is sold or scrapped. The practical question is who eats the loss: the lenders who funded it, or your rates and taxes.
