Zuckerberg signed a $100 billion check blind and Nvidia backs $105 billion of leases

On a podcast published October 12, Mark Zuckerberg said Meta committed somewhere between $100 billion and $200 billion to AI compute late last year, before anyone had seen the model that money was meant to train. A day later Semafor's Liz Hoffman said Nvidia is too big to fail and compared it to AIG. Nvidia's own filing shows why she can say that with a straight face. The company has guaranteed up to $105 billion of data center leases for an OpenAI affiliate. Everyone knows the buildout is expensive. This week showed whose name is on the paper.

Zuckerberg signed before he saw the model

Zuckerberg and Alexandr Wang, Meta's chief AI officer, were on the show to launch a model called Muse Wonder. The useful part came nearly an hour in, when the host asked how you manage a product whose feedback loop is longer than a year. Zuckerberg described the budget meeting at the end of 2025.

we're going to write this check for, like, I don't know, like $100 billion, $200 billion. I'm like, Alex, are you sure this is going to work?

Mark Zuckerberg, October 12, 2026, source

Wang's answer, by his own account, was a "vibe report" and some plots. The two then retold a warning from "Susan" at an executive offsite in Hawaii, which I take to be CFO Susan Li. The transcript's speaker labels blur at this point, so I will not pin the retelling on either man, but the warning was that if the model is bad, "it's really, really, really bad" in financial terms.

Zuckerberg also gave the unit price. Running a one gigawatt cluster costs about $15 billion a year, he said, and building one costs more.

If you're talking about building the CapEx, it's probably closer to, you know, I mean, you pay for it upfront, so it's probably, you know, $50 or $60 billion.

Mark Zuckerberg, October 12, 2026, source

Meta's Hyperion site in Louisiana is planned for 5 gigawatts, with 2 gigawatts due by 2030. Multiply his number by five and Hyperion alone is $250 to $300 billion. That multiplication is mine. Meta can carry it, because Meta has an ad business behind it. Most of the companies buying the same chips cannot.

Hoffman says Nvidia now looks like AIG

That gap is where Nvidia stepped in. Hoffman went on Prof G Markets on October 13 to defend a newsletter in which she called Nvidia too big to fail. Her test is whether Washington would rescue it. She thinks it would not rescue Meta or Google and would rescue Nvidia. She described two old financing tricks done at a new size. One is vendor finance, the way Ford lends you money to buy a Ford. The other is lending your credit rating to a customer who has none, which she likened to "your parents co-signing on your first apartment."

The co-signing is on file. Nvidia's 10-Q for the quarter ended July 26 says that in August it guaranteed leases for an affiliate of OpenAI at SB Energy's campus in Pike County, Ohio. The guarantees cover about 4.25 gigawatts and are capped at $105 billion. Separate guarantees for AI cloud partners add a maximum of $3.5 billion. Then there are the equity stakes. On October 10 the Financial Times reported that Nvidia, which has already put $800 million into Reflection AI, is in early talks to buy the startup outright, according to Reuters.

we were comparing Nvidia to the big banks, but they actually look in some ways a lot more like AIG in the role that AIG was playing. They were not a bank, but they were providing insurance for all of these things.

Liz Hoffman, October 13, 2026, source

She was careful to say Nvidia's balance sheet is strong and that this is a thought experiment. I think the comparison is half right. AIG wrote insurance on assets it did not understand. Nvidia is guaranteeing rent for its largest customer, in buildings that hold its own product. It knows the asset. What it cannot know is whether OpenAI's revenue arrives on the lease schedule.

Gray reads the delayed IPOs as discipline

Jonathan Gray, Blackstone's president, is on the other side of these leases as a landlord. At the Bloomberg New Economy Forum in New Delhi on October 12 he told a panel that AI's "demand for capital" is part of what is pushing bond yields up, and that Blackstone will stay "very long electricity". The next day a Bloomberg Television anchor asked him whether cost of capital could choke the buildout.

But in aggregate, when I look back at this versus other periods of excess, like the telecom boom where people built supply ahead of demand and where there was a ton of leverage, this looks and feels very different.

Jonathan Gray, October 13, 2026, Bloomberg Television

His evidence was that several AI public offerings have been postponed. "Investors are cautious. And so to me, that's a healthy sign." He also said most of Blackstone's exposure is long leases to the biggest companies in the world. That is true and it is also Hoffman's point seen from the landlord's chair. A lease is only as good as the tenant, and when the tenant is a lab that loses money, the landlord wants a bigger name behind it. Nvidia is that name.

Porat calls the capital destruction case illogical

Ruth Porat, Alphabet's president and chief investment officer, got the bear case read to her on stage at a Fortune event on October 12. Jefferies has argued since July that the likeliest end of the boom is "massive capital destruction in the US" as cheap open models pull users away from the labs the data centers were built for. Porat, who advised the Treasury in 2008, did not engage with the debt question at first. Pressed, she said this.

It's illogical to say that we're in a capacity-constrained position because there's such strong demand for the insight and the ability to improve business and government, and yet you can't drive a return on it.

Ruth Porat, October 12, 2026, source

She repeated that Alphabet will spend about $200 billion this year and "significantly" more next year. Her answer holds for Alphabet, which sells its own capacity to its own customers. It says nothing about a lab renting 4.25 gigawatts on someone else's guarantee. When the interviewer suggested some people are building recklessly, Porat's reply was "There always are", followed by a hope that the market shakes them out. From someone who advised Hank Paulson through 2008, I found that thin.

Su's numbers say the demand is real today

The strongest argument for the optimists came from Lisa Su at the Open Compute Project summit on October 13. AMD's chief executive said inference is now about 60% of the AI compute market and that agent traffic on OpenRouter grew fourteenfold this year.

We're now at some number like almost 40 quadrillion inference tokens a month.

Lisa Su, October 13, 2026, source

She said AMD has raised its forecast for server CPUs three times in nine months, to more than $200 billion by 2030, and that its Helios rack began shipping in September. The first customers she named were OpenAI, Meta, Microsoft, Oracle and Anthropic. Read that list against the 10-Q. One of those five buyers is the tenant whose rent Nvidia has promised to cover.

So who is left holding the lease?

Nvidia, mostly, and by extension everyone who owns an index fund. Meta and Alphabet are betting their own cash, and if Zuckerberg's check was a mistake his shareholders eat it and nobody else does. The labs are different. Their buildings are financed on leases, the leases are backed by a chip supplier, and that supplier's largest customers are the same labs. Gray is right that this carries less leverage than telecom did in 2000. Hoffman is right that the risk has collected in one place. I would watch one line in Nvidia's next 10-Q, the guarantee cap. If $105 billion grows toward the 3.8 additional gigawatts the filing already leaves room for, the AIG comparison gets less theoretical.