November 21, 2026
Heard in AI

Neocloud slide: Moonshots panel splits over what could pop AI bubble

As neocloud stocks fell 30% to 40% in October, the Moonshots panel split over how an AI correction would come: costly loans, Alexander Wissner-Gross's demand-shock scenario, or Elon Musk's cheaper credit squeezing rivals.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on 2 podcast episodes published October 24–28, 2026

In October, shares of the neoclouds "got hammered," in Peter Diamandis's words. Neoclouds are companies that borrowed money to buy NVIDIA chips and rent the computing out to AI developers. By Diamandis's count, CoreWeave and Nebius were down 30% for the month and IREN was down 40%. On an episode of Moonshots recorded October 27, 2026, and published the next day, Diamandis, founder of XPRIZE and Singularity University, asked his co-hosts whether this was the AI bubble starting to pop.

They did not agree on how a correction would actually happen. The explanations on offer included bad loan terms, a sudden collapse in demand for computing, a shock from outside the industry, and Elon Musk using his better access to credit to squeeze rivals out.

What set off the sell-off

Diamandis said the trigger was CoreWeave's "$51 billion in debt" and its obligations if it defaulted. That figure matches CoreWeave's own accounts only when two categories are added together. The company's second-quarter results list about $35.1 billion in debt and $16.3 billion in operating-lease liabilities as of June 30, 2026. The same report puts CoreWeave's revenue backlog at about $104 billion. The backlog counts contracted future payments that still depend on CoreWeave delivering the service.

Diamandis read out a post from investor Chamath Palihapitiya predicting "a lot of partially stranded compute and compute sitting in inventory in 2027," machines that are "bought and paid for" but "effectively under earning or collecting dust." Palihapitiya expected 2027 to be "a turbulent year" that would "clear out a bunch of bad bets." Musk replied that debt markets had almost entirely shut off money for any company without substantial revenue independent of AI data centers, and that "the 200 neoclouds" faced "guppy-level death and consolidation." IT Pro reported the same forecast of failures and consolidation among more than 200 specialized AI-cloud providers.

Diamandis compared the moment to the fiber bust around 2000, when companies laid far more internet cable than the boom needed. That spare "dark fiber" was eventually used by services like YouTube and Netflix, he said, "but there were a lot of bankruptcies along the way."

Dave Blundin: it depends on who signed bad loans

Dave Blundin, founder and general partner of Link Ventures, accepted the fiber comparison and offered a theory of how such booms go wrong. The internet thrived, he said, while the people building it were about 80% engineers and 20% business school graduates. When that ratio flipped, the overbuilding began. In AI, too, people race in "not knowing anything about AI" because they want a piece of the boom, and the easiest way in is a big financing deal. He estimated about $100 billion of data-center debt beyond Musk's $40 billion, "tiny by any global standard" but large within one market.

Blundin didn't think the hardware itself was at risk. "All compute is sold out for years," he said, so anyone with a working data center should do very well. The danger, in his view, lies in the financing. Musk borrows at 8%, he said, while the others are "apparently" borrowing at 15%, "and that can add up." A technical breakthrough that made today's chips, fiber or power investments obsolete could turn that cost into a crisis, and at the pace technology is changing, "it wouldn't take much." He joked that Wissner-Gross could "vibe up something in a couple of weeks" that made $10 billion or $20 billion of 15% borrowing "look really, really shaky."

His conclusion was that the question has no technical answer: "It's did you sign a bad deal with a bad lender? You know, it's pure business."

Alexander Wissner-Gross: software could outrun the hardware

Alexander Wissner-Gross, a computer scientist and founder of Reified, said the elephant in the room was "dollars in versus dollars out," meaning the return on money spent on computing. He offered his answer as one possible scenario, looking, as he put it, into his "non-existent crystal ball."

The scenario starts with recursive self-improvement, in which AI systems help improve the algorithms, models and training methods used to build better AI. So far, he said, every such gain has worked like "microeconomics 101": better software makes hardware more useful, so the industry wants even more of it. But if, perhaps in the next year or two, researchers found something like "the perfect model," so much better than today's that it needed far less computing, "software could potentially run away from the hardware." That would cause a demand shock, something he said the industry has never really experienced, apart from a smaller and mostly speculative one after an earlier DeepSeek model. Because infrastructure is "arguably" over-leveraged right now, he said, such a shock "could lead to a major correction." He added: "It's not financial advice, but it could happen."

Later he named what he saw as the likeliest trigger if there is a bubble at all. He separated stock prices from the credit markets that lend to builders, and said he would bet on "a demand surprise." New computing capacity would come online, but "the killer applications don't arrive at exactly the right time to motivate all that spend." His example was OpenAI, which he said assumed, wrongly as it appears, that all its consumers would want reasoning models and the extra computing they consume, without a great use case to justify the spending.

How big the debt really is

The discussion then put the debt in a larger frame. Counting direct data-center debt at about 500 billion, and roughly double that once special-purpose vehicles and vendor financing are included, the argument went that the total is approaching the scale of the subprime mortgage market, put at 1.3 trillion. That estimate is several times Blundin's.

The same argument tied the risk to Wissner-Gross's point about efficiency. Once models are good enough, engineers optimize them and the hardware they run on. The example offered was a model with the performance people were excited about a year ago now running on a smartphone: Bonsai. PrismML's announcement of Bonsai 27B describes compressed versions of an open Qwen model, with the smallest at 3.9 GB, sized for phone memory. On the company's own 15-benchmark suite, its two variants kept roughly 90% and 95% of the full model's average score. The phone demonstration used image context that had been cached and loaded in advance. From this came a warning, framed as "not financial advice," to stay "nowhere near" memory-chip stocks such as Micron right now.

The comparison with Musk came next. His return on chips was described as "like nine months," because he sells them at around three times the market price, so he will have paid off the chips before the bubble shifts. "That doesn't hold for the core weaves and others of the world," the argument concluded.

Other ways it could pop

Diamandis proposed a different trigger: a major hack that brings down a power center or a stock market, with the fear that follows. The reply noted a recent wave of targeted attacks on PHP, a widely used web programming language, and a chart of critical vulnerabilities that "looks exponential." But it argued that a pop would come less from something happening with AI than from investors asking where the value sits in the chain, which made Anthropic's coming stock market debut a test. Wissner-Gross said a poorly performing debut would not necessarily create a crisis, and drew a line between the stock market and the credit markets.

Salim Ismail, founder of Open ExO, argued that demand made this boom different from the fiber and internet bubbles. He cited NVIDIA CEO Jensen Huang saying its chips cannot be bought for two years because the order book is sold out. Companies may fail because of the deals they made, Ismail said, "but the technology is going to win. The capital structure may still lose in that."

Blundin agreed, noting that in 2001, headlines declared that the internet had imploded just as Google was starting to take off. NVIDIA runs at 80% gross margins and is sold out for years, he said, so the real financial question is whether that is already reflected in the share price. Chip companies are now worth one to five trillion dollars each, he said, and headlines about overfunded data centers describe individual deals becoming overpriced. "That's financial news," he said. "You've got to kind of tune that out if you want to look at what's actually happening in the world."

Why SpaceX can borrow when neoclouds can't

In the same post, Musk said SpaceX had closed $40 billion of debt to buy GPUs. Diamandis asked how that was possible in the same month lenders were shutting out neoclouds. Musk's explanation, as Diamandis quoted it: "we're investment grade and have Starlink and launch revenue." IT Pro's account reports the same reasoning. Investment grade means a credit rating that marks a borrower as unlikely to default.

Diamandis spelled out the difference. A neocloud's only income is renting out chips, so if rental prices fall it cannot pay its loans. SpaceX, he said, has 13.6 million Starlink subscribers paying every month and a launch business with "zero real competitors." It also sells computing itself, which makes it a direct competitor. "Guppy-level death," in Diamandis's reading, describes a market where "the big fish get bigger." He pointed to Anthropic agreeing to buy $120.5 billion of computing from SpaceX through 2029. Anthropic's S-1 filing describes those agreements as worth "up to" that amount. After initial periods ending in January and March 2027, either side can cancel with 90 days' notice.

A similar consolidation view came a few days earlier from inside the market. Evan Conrad, CEO of the GPU marketplace SF Compute, told The Cognitive Revolution that credit, not chips or power, was holding back the build-out. He said new customers are very hard to sign unless they are investment grade, and he predicted a wave of consolidation among AI labs and inference companies.

Asked whether the build-out would end up with three or four giants, Wissner-Gross said he doubted it, and that he wasn't sure the market would be recognizable in five to ten years. He said Musk appeared to be "trying very intentionally" to "suck the oxygen out of the room in public credit markets," and was "talking his own book" because "he wants all the loans for his own entities." He thought Western credit markets had real depth, and pointed to larger pools of money, including insurance markets, that could also finance infrastructure. As credit markets are "sucked dry," he expected "increasingly adventurous and aggressive financial engineering" from neocloud CEOs seeking capital "in exotic ways."

A license to print money

Blundin seized on the "talking his own book" point and compared Musk to Michael Saylor. When Saylor was in hypergrowth with Bitcoin a couple of years ago, Blundin said, he could buy Bitcoin and sell shares with a spread of about 30% between the two, which "manufactured basically 100 billion dollars of wealth" in a couple of years. Musk, he said, is in the same position: he can borrow at 8%, buy computing, sell it to Anthropic and, using the nine-month figure raised earlier, get back above breakeven within that time. "That's a license to print infinite money," Blundin said.

He took the idea further. Palihapitiya, he said, argues that Musk's 8% debt beats buying sovereign debt, and in that view the AI debt market becomes "much bigger than sovereign debt" if it stays on its current path. Musk, Blundin said, would then be "bigger than a country," able to borrow hundreds of billions. "Better than 90% of the countries out there," Diamandis said. "And much more reliable," Ismail added. Blundin finished the thought: "bigger than a large country."

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03
  4. 04

Connected ideas and articles

From the conversation

Podcast episodes

Moonshots with Peter Diamandis

Anthropic’s $2T IPO, Elon Predicts Neocloud Death, and China’s Kimi Catches Up | #306

Episode published (recorded )This article draws on 15:29–23:47 and 23:48–31:35 (approximate times)

Article history

Updates to this article

Tags