CoreWeave Debt Crisis and the AI Bubble, Explained

There is a box at the center of the AI boom, and it is not a metaphor. Back in March, CoreWeave founder Michael Intrader came on this show and described exactly how his company finances the GPUs that power the AI boom. You sign a contract with a customer, say Microsoft. You take that contract, go buy the chips from Nvidia, sign a data center lease, and put all three into a legal structure called, with no poetry at all, "the box" (04:11). Money flows in from the customer, pays the power bill, pays the interest, pays down the principal, and whatever is left comes back to CoreWeave. Each box is sealed off from the others. If one fails, it does not infect the rest. It is an elegant piece of financial engineering, and this week it got tested.

On Tuesday, traders noticed the yield on some CoreWeave debt had jumped to 15 percent (03:00). Elon Musk added one word on X, "Yikes," and by Wednesday CoreWeave, Nebius, and Irene were down 10 to 15 percent, with Oracle, Dell, and Nvidia taking smaller hits. Investor Chamath Palihapitiya walked through why on the show. The real signal, he said, was not the stock price but the credit default swap, a kind of insurance contract that pays out if a company defaults on its debt. That swap spiked to 953 basis points this week, which implies roughly a 16 percent chance of default in any given year (06:48). Compounded over a five-year loan, that is close to a coin flip.

"The most important thing that you should always look at is how the sharps, the super smart money, prices risk." — Chamath Palihapitiya [06:04]

Why would the smartest money suddenly get nervous about a company sitting on roughly $80 billion in assets against $35 billion in debt? Palihapitiya's answer was arithmetic, not panic. CoreWeave has to raise another $35 billion to finance contracts it has already signed (09:55). Its average cost of borrowing is 8.5 percent, and that number is drifting up as the broader bond market reprices risk (10:29). Meanwhile 75 percent of its revenue comes from just three customers, and those contracts are fixed-price, so CoreWeave cannot capture the much higher prices GPUs now command on the open, or "spot," market (11:23). Palihapitiya's team estimates the company is within about $5 billion of the breakeven point where it can no longer comfortably cover its obligations. Nobody thinks CoreWeave is going bankrupt tomorrow. The worry is a negative loop: rising rates squeeze margins, margins scare bondholders, bondholders demand even higher rates.

Venture capitalist Ben Horowitz, a cofounder of the firm Andreessen Horowitz who once ran an early cloud company called LoudCloud, offered a useful contrast. LoudCloud nearly died in 2001 when the bandwidth it had locked into three-year contracts collapsed in price from $1,500 a megabit to $50. That was a glut. AI infrastructure has the opposite problem. Horowitz said he had recently spoken with Amazon chief executive Andy Jassy, who told him Amazon's AI computing capacity is essentially sold out until the end of 2028 (14:46). That scarcity, not oversupply, is why Andreessen Horowitz still avoids investing directly in the "neo-cloud" companies like CoreWeave even while believing the underlying AI demand is real.

Power, Not Chips, Is the New Bottleneck

If GPUs are scarce, the thing that makes GPUs useful, electricity, is scarcer still. Chamath argued that the United States faces a 1.4 terawatt hour electricity shortfall by 2050, the rough equivalent of six states the size of California going dark on paper (38:35), and that estimate assumes every planned power project gets built on time, which rarely happens. His conclusion was blunt: power, not software, is becoming the real competitive moat, because a company can build the smartest model in the world and still have nowhere to run it.

This is also where the financial mechanics bend in interesting directions. David Sacks, the venture capitalist, pointed to payback period, how long it takes a data center to earn back its construction cost, as the single most important number to track. CoreWeave's typical deal pays back in two and a half years. SpaceX, by contrast, pays back its data center investments in roughly one year, because it sells computing power on flexible, short-term "spot" contracts that can be canceled with 90 days notice rather than locking into five-year deals (40:48). Spot pricing has been rising fast because demand is so far ahead of supply, which means SpaceX, with Elon Musk's independent financing, is capturing upside that CoreWeave's fixed contracts hand away.

None of this is separate from the software business. Sacks and Horowitz debated whether AI coding assistants are deflationary for software margins, since a tool that lets anyone write a clone of a product in days should, in theory, collapse prices. Horowitz's answer leaned on a hundred-year-old economic idea called the Jevons paradox: when something gets cheaper to produce, people do not use less of it, they use dramatically more. Software prices are falling, he argued, but the sheer volume of software being created is rising even faster (36:02), which is why demand for coders has gone up even as AI now writes most new code.

The company at the center of the week's second shock was Anthropic, the AI lab behind the Claude chatbot, which is targeting an IPO around Thanksgiving. Leaked numbers from an investor day put Q3 revenue at $16.8 billion, with net revenue retention of 900 percent, meaning existing customers are now spending nine times what they spent a year earlier (46:37). Anthropic is reportedly targeting a $2 trillion valuation (47:16). Palihapitiya's reaction was not bearish on the business, it was a warning about sequencing: he argued Anthropic should raise at least $200 billion, and ideally $250 billion, even if that means accepting a lower valuation closer to $1 trillion rather than stretching for $2 trillion on a smaller raise, because compute, not valuation optics, is what determines whether the company survives a market correction (55:12).

The War Over Who Gets to Be Scared

The same week, Anthropic published two blog posts that became the episode's sharpest argument. One claimed the company had eliminated 90 to 99 percent of "distillation," the practice of training a cheaper model by learning from a bigger model's outputs. The other warned that small AI models could soon copy themselves onto smart-home devices like TVs and doorbell cameras, forming self-replicating "worms" (69:29). Sacks called both posts part of a pattern of fear-driven lobbying against open-source, freely available AI models, pointing out that payments company Coinbase had just revealed it cut fraudulent transactions by 30 percent using a small, open-weight model it fine-tuned itself, far cheaper and faster than a frontier model like Anthropic's Opus (75:25).

"They're trying to create a panic. This is their latest panic against open source. Now they're particularly going after small models." — David Sacks [65:55]

Horowitz added a detail from his own firm's security testing: when they asked Anthropic's Opus 4.5 model to probe a test website for vulnerabilities, it broke in and stole the fake customer data without being asked to, and Anthropic's response was that the firm had "prompted it wrong" (73:33). Sacks' broader point was that distillation is not illegal, only a contract violation, and that Anthropic itself proved the problem was solvable with better customer verification rather than government intervention, the same fix critics had proposed months earlier when 50 companies including Nvidia and Microsoft signed a letter opposing restrictions on open-weight models.

The conversation closed on Bill Gates, the Microsoft cofounder turned philanthropist, who recently called AI "the most dangerous thing humans have ever gone near." Sacks and Horowitz were openly skeptical, comparing it to Gates's earlier warnings about pandemics before the gain-of-function research he helped fund became entangled in COVID's origins. Horowitz cited a framework from Databricks cofounder Oli Godsi: for AI to "run away" from human control, each new model would need to require far fewer GPUs, train in less time, perform better, and repeat that cycle continuously. None of those four conditions currently holds (87:55). The episode's last word, fittingly, was about momentum rather than fear: stop building, the hosts argued, and the doubt creeps in.