Monologue: Anthropic's $413bn Burden

Ed Zitron (0:01)

CAUZO MEDIA Hello and welcome to this week's Better Offline monologue. I'm your host, Ed Zitron. [♪ music playing, [♪ music playing, [♪ music playing, So Reuters has continued to drip, drip, drip Anthropic's IPO prospectus in a way that's very frustrating for anyone trying to follow this story, but in doing so have revealed a deeply worrying truth, that Anthropic has $413 billion worth of non-cancellable compute contracts and chip leases across Google, Amazon, Microsoft, and Broadcom. $252 billion of that is dedicated to the hyperscalers and $161.2 billion to Broadcom. Broadcom is, in case you didn't know, the company that makes Google's TPU AI chips, of which Anthropic is the largest customer by country mile, as far as rentals go, and now as a result of this deal, it's the largest customer purchasing them. And by the way, yeah, Anthropic buys them, puts them in a Google data center, then Google, they pay Google to rent them back. It's not circular financing though, stop calling it that.

Anyway, in a story that came out the morning of October 1st, Reuters also revealed that Broadcom had agreed to lend Anthropic $42 billion to finance purchases of these TPU chips, though they called them leases, their purchases, which is the first example we've seen in the AI bubble of .com style vendor financing, where a company lends another company the money to pay them. The contract, per Reuters, allows Broadcom to also convert the debt into Anthropic shares, an interesting escape hatch for the inevitable point at which the AI lab can't actually afford its bills. Also Broadcom really can't afford this debt. They were apparently raising $60 to $100 billion a day. Is this part of this? Because if it's not, I don't know, doesn't really make sense. In any case, that $42 billion would only scratch about a third of what Anthropic needs with Broadcom, which means it will have to find the cash elsewhere. As a reminder from earlier in the week, Anthropic lost $8 billion on $4.6 billion of revenue in 2025, which means that this company is light years away from funding any of these ridiculous costs out of cash flow. And really that begs the question, how will Anthropic pay for all these bills?

Let's assume that this horrible oaf of a company is allowed to go public, and that the ratings agencies allow it to become investment-grade, which I believe should have people put in prison, which I know sounds dramatic, but it actively harms investors, and it will be historically awful for them. Even if that were the case, the company would have to pull in somewhere between $50 billion and $100 billion a year. And even at investment-grade, they'd still be borrowing at 7% to 10%, because their negative cash flows and their opaque path to profitability mean that investors would want to be compensated for that risk. And let's be clear that Anthropic would need to raise debt every single year without fail to keep its business afloat. And every time it did so, investors would likely demand more for their money. And that would be if the bonds didn't sell off, which they absolutely would because of any negative story about this company.

Now, this all assumes that Anthropic would get an investment-grade rating, because if it got a junk rating, it would have to borrow at rates somewhere between 10% and 13%, which would be effectively lethal, basically a personal loan at this point. While it could do convertible debt, meaning that the loan would convert into shares if they hit a certain price, and would otherwise be a low interest rate, there are limits to how much one can borrow, and Anthropic needs tens of billions of dollars multiple times a year without fail to pay the bloody cloud bills. I should also add, yes, Nvidia would likely give it a couple billion dollars. There would be some venture investment, but we are talking tens of billions. I haven't even discussed Anthropic's putrid cousin OpenAI, who also wants to go public sometime in the next year, and will also need all this debt, and projects to spend $750 billion on cloud compute through 2030 All those Oracle data centers, including the one in New Mexico, where Oracle just declared a force majeure, depend on OpenAI's ability to pay, which will depend on its ability to raise money, which all comes down to whether venture capitalists or debt markets can keep it alive. And the more money that everyone borrows, by which I mean AI data center companies, AI labs and hyperscalers, the more expensive it becomes for everyone to borrow, as I discussed a few weeks ago with Paul Kudrowski.

All that AI debt is creating meaningful competition for US treasuries, at a time when the war in Iran and Ukraine is causing a sell-off in the bond market, and when bonds on the market sell off, their effective interest rate goes up, which means that any future government debt, and indeed, any future debt from anyone, will have to price higher. To explain this really simply, imagine a bond, a treasury bond, is paying 5% and it cost $100 to buy. It's technically a little different, but just for the example. Now, if that bond sells off and someone can buy it for $90 instead of $100, the effective interest rate is actually higher, because you're getting more bank for your buck. That doesn't make sense. Email me, I'll happily explain. Nevertheless, this is going to make everything more expensive. And I hear no compelling counter-argument to any of this outside of people saying it'll all work out, and that these are the fastest growing companies in the world. Ed, check out Anthropic's annualized run rate, which according to Insane Analysis, was calculated by an investor by multiplying July 31st, so approximately $178 million, by 365 If that's what's happening, I question every single run rate number we've ever heard. I think they were bollocks before, and they're horse shit now. And this childlike thought process will mean that the media and the markets are utterly unprepared for when things begin to collapse. The information just reported that SMBC and MUFJ, the two largest Japanese banks and major backers of effectively every AI data center, including Stargate's, I mean, actually all of them, I think one or both of them, someway in every single one, they're becoming more selective with their investments in AI.