AI巨头发债潮推高美债收益率，Oracle信用违约成本创纪录

The companies building AI are now some of the biggest borrowers on the planet. In the space of 12 months, Alphabet's total debt went from 41 billion to $112 billion. And Metas went from 49 billion to $112 billion. Amazon and Oracle went even further. They're spending more on chips, power, and data centers than their own operations bring in. And for a while, the AI buildout was paid for by some of the deepest pockets in corporate history. very healthy and astronomically large cash piles.

But now it's being paid for with borrowed money. And SpaceX just joined the rush as well, lining up a $40 billion package to buy Nvidia chips with $30 billion set to come from the bond market. And what's more, all of these corporate bonds are entering a market that's already being asked to absorb more than $40 trillion of US government debt. Now, in case you haven't noticed, that market is turning very ugly. Treasury yields are hitting levels not seen in decades.

So, borrowing costs are ripping higher and higher, and investors are demanding a painful price to keep swallowing all this debt. And now, the AI boom is dumping another mountain of borrowing on top of it all. So, how did this mess come about? Well, to find out, we'll have to look at how AI got onto the non-stop borrowing treadmill to begin with, why the bond market struggles to take it all, and how the whole chain is wired together.

My name is DC and this is the Coin Bureau. So, how did the richest companies in tech end up borrowing like this? Because for the longest time, they simply didn't have to. The big AI players funded their data centers out of profits, and those profits were enormous. But then the game changed when it became clear a monumental physical buildout was underway. Chips, power, and land all became way more expensive and more contested, and nobody wanted to be the company that was late to the party.

So over the last four quarters, Amazon spent $173 billion on capital expenditure against $161 billion of operating cash flow. Oracle spent 75 billion against $46 billion of operating cash flow. And Alphabet posted its first negative free cash flow quarter ever in Q2 2026. Then went straight to the bond market with a $25 billion deal with maturities stretching from two to 40 years. 40 years for a technology that didn't exist in its current form just 5 years ago.

So to summarize, the buildout outgrew the cash. But plain bonds were only the start because the biggest projects needed something way more creative. So just a year ago, Meta closed the financing for its Hyperion campus in Richland Parish, Louisiana. It did this through a special purpose vehicle or SPV, which is a separate company set up purely to own the asset and carry the debt. Funds managed by Blue Owl own 80% of it.

Meta owns the other 20% and the vehicle issued a 27.3 billion bond paying $6.58% that doesn't mature until 2049. Pimco bought around $18 billion of it. So in simple terms, Meta now leases the campus on terms where it pays no matter what happens, but the debt stays off Meta's own balance sheet because of that SPV structure. Nothing strange about that, like at all. And then the chips themselves became collateral. Back in March, Coreweave closed an $ 8.5 billion loan backed by Nvidia GPUs and a contract with Meta.

And Moody's gave it the first investment grade rating ever handed to GPUbacked debt. And this is where the AI trade became even more intertwined with credit markets. Now the market was essentially weighing up who would lend against the models, not just what the models were capable of from a software perspective. So, who's actually buying all of this debt? Well, the same entities who buy everything else. Pension funds, insurers, and giant asset managers like PIMCO and BlackRock.

That pool of money is huge. But any given moment, it's finite. Think of it as an auction room with roughly the same biders in it every single week. When more lots arrive, the bids don't just randomly get richer. Each lot just goes for a little less. And in the bond market, a lower price means a higher yield. the interest rate the borrower ends up paying. Now, the biggest seller in that room was already the US government.

The national debt is at $40 trillion. The deficit runs at 6% of GDP, and interest payments came to over $1 trillion in the first 11 months of fiscal 2026. But of course, it's not just Washington. Japanese, French, British, and German government bond yields are also at multi-year high. So, in other words, the room was crowded before AI walked in. Then the hyperscalers turned up with a boatload of debt. Amazon, Microsoft, Alphabet, Meta, and Oracle issued around $200 billion of investment grade debt in the first half of 2026, nearly double their total for all of 2025.

And now SpaceX is lining up another 30 billion of investment grade debt as part of that $40 billion package to buy Nvidia chips. Every single dollar PIMCO put into the Hyperion bond was a dollar that didn't go into a treasury, a mortgage bond, or another company's debt. Now, of course, the market wasn't built to take this much new supply this fast. And as I'm sure you've seen in the headlines recently, it's starting to show, but with so many moving parts to this madness that is the AI buildout, you can easily get lost in your analysis and miss the detail that's truly important.

So, why not let our research team do it for you? Because right here on YouTube, you can join the Coin Bureau Club live plan. This is where you get daily coverage across all major markets, trady and crypto. Our teams read on where capital is flowing and none of the noise. And the best part, if you join today, your first month is absolutely free. Just tap the join button below this video to get started. Okay, back to the bond market.

So, where exactly is the strange showing? Well, the first place is government debt. Exactly 1 year ago, from the time of this video, the US 10-year Treasury yield was at 4.18%. Right now, as I shoot this, it's at 5.31% and the 30-year has climbed from 4.76% to 5.66%. And sure, deficits, inflation, and oil price above $100 a barrel all feed into that. But corporate borrowers chasing the same buyers are part of the pile, too, and AI is now the largest of them.

The second place you can see stress in the system is the AI bonds themselves. Back in February, hyperscaler bond deals were drawing orders worth nearly five times what was an offer. By July, that had fallen below two times. So when Amazon came back for another $25 billion in the summer, it had to sweeten the price just to get the deal done. And as of late July, 78 of the 91 hyperscala bonds issued in 2026 were trading at higher yields than when they were sold.

So in other words, most of the people who bought them are already underwater. Not quite the safe haven they signed up for. The third place you can clearly see strain is Oracle itself. Its debt has climbed from $ 105 billion to 155 billion in a single year. So the cost of ensuring that debt against default measured through credit default swaps has gone from below 50 basis points to a record of more than 225 in late September.

Oracle still carries an investment grade rating, but the market is pricing it more like Chuck. So, the pressure isn't spread evenly. It's concentrated where the borrowing is heaviest. But Oracle isn't a standalone story. It's in the middle of a chain and every link in that chain heavily depends on the next one. Around 20 banks lent $18 billion for Project Jupiter, Oracle Stargate Campus in New Mexico. Oracle is the builder and it signs up the customer which is OpenAI with a contract worth about $300 billion over 5 years.

Open AAI then raises money from venture capital, SoftBank sovereign wealth funds and Nvidia to pay Oracle and Soft Bank itself sold 11.1 billion of junk bonds with yields of up to 9.75%. So even the investors are now borrowing and then there's the loop running behind it all. Nvidia invests in OpenAI