华尔街隐藏AI债务炸弹：GPU租赁价两年暴跌70%

an Nvidia H100 chip rented for roughly $8 an hour in early 2024. By late 2025, that same chip was going for $2 to $3. That's a 50 to 70% collapse in rental value in under 2 years. After 3 years, analysts estimate a GPU's resale value can fall to around half of what you originally paid. What's more, the debt being written against those chips runs 5, 10, 15 years. and one metal linked structure doesn't fully mature until 2049.

So, Wall Street is doing more than investing in AI. When it comes down to it, Wall Street is lending against AI. And the way it's lending right now, well, it looks a lot like the leadup to the last market blow up in 2006. So today, we're breaking down how this debt is structured, why the collateral underneath it doesn't hold its value, and why publicly listed Bitcoin miners are now positioned in the most leveraged pocket of the entire thing.

My name is Lewis, and this is the Coin Bureau. Morgan Stanley puts global data center CEX between 2025 and 2028 at roughly $2.9 trillion. Split it out and it's about $1.3 trillion of buildings and power plus $1.6 trillion of the actual computing hardware. Against that, the operating cash flow of the hyperscalers so Microsoft, Amazon, Alphabet, Meta cover around $1.4 trillion, which leaves a financing gap of roughly 1.5 trillion.

That's more than the entire market cap of most stock exchanges on Earth, and it has to be borrowed from somewhere. So, where is it coming from? Well, Morgan Stanley's own breakdown says about $800 billion from private equity and assetbased finance, roughly 200 billion from ordinary corporate bonds, and around 150 billion from securitized products with sovereign wealth funds and joint ventures mopping up the rest. Private credit, by the way, just means non-bank lenders. the Apollos, the Blue Owls, the Blackstones, the PIM codes lending directly rather than through a bank.

And the growth there is mind-blowing. Direct private credit lending to AI related companies has gone from effectively nothing a few years ago to over 200 billion outstanding. The five biggest hyperscalers issued $121 billion of US corporate bonds in 2025 alone against an annual average of about 28 billion across the previous four years, more than four times the run rate. And global AI related debt issuance is tracking towards nearly $570 billion in 2026, roughly double the year before.

So the money for the AI buildout is coming out of credit. And credits always has a structure attached to it, which is where things get really interesting. When a company borrows money, that debt normally lands on its balance sheet where rating agencies and shareholders can see it and price it. The whole game in AI finance right now is making sure that doesn't happen. Take Meta's Hyperion campus in Richland Parish, Louisiana. over 4 million square ft, up to 5.5 gawatt of capacity, and around $30 billion of total financing.

But Meta didn't borrow it. A bankruptcy remote vehicle called Big Net Investors LLC did. Bankruptcy remote, by the way, just means the vehicle is legally walled off. So if it fails, it fails alone. Now, Blue Owl managed funds own about 80% of it. Meta keeps around 20% and roughly 27 billion dollars of debt was issued out of that vehicle anchored by PIMCO to the tune of about 18 billion. SNP rated the paper A+ one notch below Meta at a fixed coupon of 6.58% roughly 225 basis points over treasuries and the debt doesn't get fully paid off until 2049 more than two decades from now.

Meta leases the site back on rolling four-year terms, which is how they keep the debt off Meta's books. Oracle is doing something similar, carrying something in the region of $248 billion of lease commitments through third party vehicles on take or pay master leases on top of more than $130 billion of actual debt. Both S&P and Moody's have Oracle on negative watch and its 5-year credit default swaps have gone to a 16-year high.

There's also a very large chipbacked financing structured for anthropic through an SPV that sells tranched notes to investors with Broadcom guaranteeing part of the senior slice. Again, deliberately arranged so the obligation doesn't sit on Anthropic's balance sheet. And a Nikkeay investigation this year put the five biggest tech firms at $1.35 trillion of reported debt plus $1.65 trillion of offbalance sheet AI obligations with the hidden portion growing eight-fold in just four years.

Debt, debt, and more debt. Wall Street is selling A+ rated bonds backed by lease payments from companies that aren't even investment grade themselves. And you know what? We saw this same trick back in 2006. You take a pool of weaker credit. You slice it into tranches. You give the senior slice first claim on the cash flows and the rating agency stamps it. Structured finance lawyers have already started mapping these AI vehicles onto 2008 era litigation, which tells you how familiar all of this is.

Now, we're obviously getting into the weeds here, and tracking a story that spans private credit and a chip depreciation is a full-time job, and it happens to be ours. So, instead of spending 16 hours a day researching, let us do it for you. Because right here on YouTube, you can join the Coin Bureau Club light plan. You'll get daily updates across both tradi, our team's read on what's actually moving the markets, and none of the noise.

We've made the light plan affordable, so it's accessible to all. And if you sign up now, your first month is absolutely free. Just tap the join button below this video to get started. Now, back to the debt and the collateral problem of the AI buildout because the structure is only as good as the thing it's secured against. And that's where this all gets very messy. Operators are depreciating GPUs over roughly 4 to 6 years in their accounts.

Amazon has already cut the assumed useful life of some server fleets, but the market data says the real economic decay is closer to 50% in 3 years. And the rental collapse I mentioned earlier is the evidence. So you have an asset with a useful commercial life measured in 3 or 4 years that has been funded with paper that matures in 5, 10, or in Meta's case 24 years. That mismatch is a big problem. But to be fair, there's an argument against this concern.

So let's consider it. The hyperscalers are enormously cash generative at the parents level. Cororeweave, the biggest of the AI cloud providers, raised an 8.5 billion delayed draw term loan on the 31st of March 2026 that Moody's rated A3 and Fitch rated A minus. So this is investment grade debt secured against high performance computing. But let's look at why it got that rating. Crucially, that facility is anchored by a meta contract.

Now look at what happened seven weeks later. On the 18th of May, Cororeweave came back with a 3.1 billion facility secured against assets tied to two customers who are not investment grade. Moody's gave that one BA2. Fitch gave it BB+. So the rating moved because the tenant changed. Then in August, a further $2.6 billion facility priced 100 basis points wider still against underlying customer contracts averaging about three years on paper running roughly five.

So the ratings are tracking the counterparty, not the asset, which raises the question about where the counterparties are getting their money. Well, increasingly it's from the chip vendor. On the 10th of August 2026, Nvidia unveiled a platform to mobilize over $500 billion of thirdparty capital for AI infrastructure alongside Apollo, Black Rockck, Blackstone, Brookfield, Goldman Sachs, and KKR. And Black Rockck's Larry Frink compared it to the creation of mortgagebacked securities in the 1970s, calling it the start of the next future for financial engineering.

That's the world's largest asset manager reaching for the mortgage comparison himself. But many have responded with the only point that's really important. For investors to pour half a trillion dollars into computing infrastructure, those chips need to h