New Delhi: The more than 200 young startups leasing access to AI computing power are in for a rough period, according to Elon Musk. On the other side of the coin, in a post on X now Twitter on October 26, he claimed that debt markets would be almost entirely shut to any company that isn’t making a lot of money outside of AI data centers.
The post, which had hit 10 million views by midday, was a response to an analysis by Chamath Palihapitiya, a Social Capital investor and founder. After a couple of words, Musk broke out with “Pretty much.”
A funding squeeze
The credit door has been closing during the last few weeks, said Musk. But his own businesses are in a different situation, he said. Earlier this month, SpaceX successfully completed a $40 billion debt issuance to purchase GPUs, special chips used to train and operate AI models. SpaceX’s success, he wrote, is due to them being “investment grade” and having Starlink and their launch business to back them.
He forecasted guppy-level death and consolidation for those that don’t have such a foundation. There will be some that survive, he said, but not many.
What Palihapitiya laid out
Palihapitiya, who was first to post about 18 hours prior to Musk, refers to himself as being a “portfolio holder of energised land assets”, which includes power already arranged for data centres. Up there, he made three observations.
The problems with neoclouds, the smaller companies providing GPU space, are that they are disorganized and poorly capitalized. Some clearly aren’t as good.
Second, hyperscalers like huge cloud computing providers are bureaucratic. They will sign an NDA in 3 months, and will close an escrow in 18 months, he said. His own portfolio is trying to compete by being quick and simple in that regard, he said.
Third, he said the existing pipeline delay will come as a surprise to people in 2027. Hyperscalers will react by focusing on their own internal compute needs over third parties. If the third party is a model company, then they will ask it to go through their model router, and then again take a non-trivial cut. This may cause Anthropic to lose access to the compute it needs, he said, and rumored numbers in the S-1 filing indicate it might already be the case.
Stranded compute, then a shortage
Then Palihapitiya’s main point is that the industry seems to be equipped with a lot of partially stranded compute, or capacity that’s built but hasn’t yet been fully utilized, as well as compute in inventory in 2027 that’s more than it’s projected to be able to use. That capacity will be purchased and paid for but underearning, or it will be underleveraged or not be used at all, that is, at less than their total cost of capital. This applies to all key companies, he said, from model companies to hyperscalers and neoclouds all have one of these issues.
In his opinion this is a dislocation that’s working in favour of two types of players. One is the non-hyperscaler big scalers who can quickly build an industrial-scale capacity. The other is the quick and agile operators that supply them, who are well-capitalized. He said 2027 will be a “noisy year” that will wipe out “wrong shots and bad shots” for some, and will also be a “consolidation year” for a few.
Musk had made this same remark the day before. On October 25, he wrote on X that he believes it’ll be “a lot of stranded compute out of completed data centers”. This post had nearly 2.3 million views.
Why it matters
The swaps indicate a strain that has been developing in the AI sector’s infrastructure; the spending race is being funded primarily by debt and lenders are now choosing sides. A neocloud is usually funded with a loan, usually from GPUs or long-term contracts from customers. When the contracts are expected to expire, and credit dries up before then, or when customers shift their business to the largest platforms, the numbers start getting harder to compute.
Neither man provided any numbers as to how many companies might cease to exist and Musk didn’t give any names. Nor has his “200+” count for neoclouds been independently verified. These are thoughts from investors with big chip at stake as is the norm with posts on X. Palihapitiya has land suitable for data centers, while Musk is the CEO of AI companies and space companies that are competing for capital and chips.
Nevertheless, the pairing is interesting. Scale and robust balance sheets determine who survives the shakeout, a separate major investor in power and one of the largest GPUs buyers in the world is saying the same thing via a different angle.
