英伟达500亿美元融资引市场担忧，AI资本开支债务化风险加剧

Nvidia has just lined up more than $500 billion of other people's money to finance the AI buildout. Six firms have signed on, Apollo, Black Rockck, Blackstone, Brookfield, Goldman Sachs, and KKR in one of the largest financing packages the AI industry has ever put together. And on announcement day, Nvidia's stock fell roughly 2.6%, 6% erasing about $130 billion in market value. One of the biggest funding wins in the history of the sector and the market treated it like bad news.

But why? Well, because that announcement wasn't really about growth at all. No, this record deal was all about credit. And as anyone who's taken even a cursory glance at the history of credit cycles will tell you, these things don't stay confined to one sector for long. So today, we're breaking down why the world's most profitable companies stopped generating cash, where the money for all this is really coming from, and how a repricing in AI reaches into things you own that have nothing to do with AI at all.

My name is Guy and this is the coin bureau. Now let's start with the announcement that changed the whole dynamic of the AI trade because for about 2 years the market rewarded every dollar of AI spending. Combined hyperscala capital expenditure for 2026 is running somewhere between $690 billion and $800 billion with Moody's putting its six company estimate at 785 billion and approaching a trillion next year. That's growth of roughly 67 to 80% in a single year, which is pretty remarkable.

But the spending isn't what's actually interesting here. The interesting part is what it's doing to cash flows. In the most recent reported quarter, Amazon spent $54.2 billion on capex and posted negative $8.8 billion of free cash flow, which is the cash left over once the bills and the buildings are paid for. Alphabet spent $44.9 billion and printed negative 5.9 billion, the first negative free cash flow quarter in its history as a public company.

Meanwhile, Meta spent over $30 billion and scraped out $1.7 billion positive. Oracle spent $16.5 billion and came in at negative 1.9 billion. Only Microsoft is still behaving like the big tech we used to know, spending $ 35.8 billion and still generating 19.6 billion of positive free cash flow. PIMCO has capex eating 94% of hyperscala operating cash flow this year, up from under half just 2 years ago. So the cash machines have stopped being cash machines.

But then if the profits aren't covering it, what is? Well, push tech fundamentals aside and say hello to the bond market. Tech and adjacent issuers put out roughly $225 billion of bonds through the first half of this year. For context, LSEG data has hyperscala specific issuance going from $16.7 billion in 2024 to $193 billion by the middle of 2026, close to a tenfold jump. Tech is now roughly 15% of all US corporate bond issuance.

And this is just the warm-up because Morgan Stanley and JP Morgan both see around $1.5 trillion of new tech and data center debt needed through 2028. Goldman's credit team puts over a third of this year's AI infrastructure spend as debt funded. Look at the individual firms and things become even clearer. Last quarter, Alphabet issued $21.1 billion of net debt. Meta issued 25.9 billion. Amazon issued 10.2 billion. Amazon's long-term debt is up 81% to over $119 billion.

Oracle is carrying 156 billion of total debt against negative free cash flow. And in July, S&P cut it from BBB to BBB minus. That's one notch above junk. And S&P noted that nearly half of Oracle's roughly $638 billion of remaining performance obligations are tied to a single customer, OpenAI. Now, the debt you can see is only part of it. Moody's tracks data center lease commitments across six major players at $1.2 2 trillion with over $820 billion of that attached to facilities that haven't even been built yet.

Morgan Stanley splits its own $1.8 trillion offbalance sheet estimate into roughly a trillion of purchase commitments, $800 billion of unccommenced leases, and about 110 billion of payables financing. Meta has around $279 billion of future lease agreements for AI data centers, a 53% jump in one quarter. Now, that was a lot of numbers, a lot of big numbers. But what's arguably more important is how all of this is structured.

And to communicate this clearly, I'll use an analogy. So, imagine a car dealership that also arranges your loan. You can't afford the car outright, so the dealer introduces you to a lender, guarantees a slice of your repayments, and then books the sale as revenue the moment you drive off. The dealer looks like it's growing. What it's actually doing, though, is manufacturing its own customer. That's called vendor financing, and it's something the Bank for International Settlements flagged in its annual report this year, naming circular AI financing as one of the top three risks to global financial stability, right alongside sovereign debt fragility.

That's a fairly stark assessment coming from the central bank for central banks. Now, to be fair, the new structures are a bit cleverer than versions we've seen in the past. The risk gets pushed off Nvidia's balance sheet into specialurpose vehicles, private credit funds, and leasebacked paper instead. Jensen Hang has said Nvidia may backs stop up to 25%, which means up to 125 billion of exposure. But, as many are pointing out, there's a bit of a mismatch here.

GPUs lose value quickly and the debt used to finance them can run for decades. In other words, you're financing hardware that ages like a laptop with debt designed for something more like a toll road. The bear's favorite bear, Michael Bur, has argued that five and six-year depreciation schedules on chips with a 2 to threeyear competitive life understates industry depreciation by around $176 billion across 2026 to 2028.

By his estimate, that overstates Oracle's 2028 profits by roughly 27% and Metas by about 21%. Nvidia's 5-year credit default swaps, the market's insurance price on it defaulting, doubled over a two-month stretch. And now, this is where things branch into something much more systemic because this actually impacts everything from private credit markets to sovereign debt. And by the way, if you enjoy deep dives into topics like this, then you should definitely subscribe to the Finance Bureau.

That's our dedicated channel for everything macro, markets, politics, and anything else that impacts your portfolio. And you can find the Finance Bureau by simply following the link in the description or by scanning the QR code on screen. Okay, let's get back to the madness of circular financing. And by this point, you might be thinking, well, sure there's debt and risk, but these are the most creditworthy companies on the planet.

This isn't 1999. What's the big deal? And that is a reasonable objection. After all, Microsoft, Alphabet, Amazon, and Meta are hugely profitable businesses with investment grade ratings and diversified revenue from search, ads, and cloud that doesn't depend on AI monetization to survive. They bring in money from outside investors rather than Nvidia funding the deals itself. that arguably gives Nvidia more protection than the vendor financing of the dotcom era when the vendor was often the one left holding the risk.

Jensen Hang says this is the first time compute has become an investable asset class describing chips as fungeible and broadly adopted. Black Rockck's Larry Frink compared the whole thing to the birth of mortgagebacked securities in the 1970s, noting each gawatt of capacity costs 50 to 60 billion to build and the US needs over 70 gawatt. John Gray says compute will be underwritten like mortgages and said that AI usage across Blackstone's portfolio companies rose sevenfold this year.

On the depreciation issue, meanwhile, Nvidia circulated a seven-page memo to analysts rejecting the Enron comparisons arguing that four to six years is realistic because chips cascade from training into inference and enterprise workloads. Now,