The $334 Billion Question: Who's Going to Lend for the Next Data Center?

The $334 Billion Question: Who's Going to Lend for the Next Data Center?

Let me tell you something I've been watching all week, and every new headline makes me lean forward a little more in my chair.

On Tuesday, Bloomberg reported that Blackstone — the world's largest alternative asset manager with $1.3 trillion under management — is trying to tie up a $3 billion loan for a single data center in Sydney. AirTrunk's SYD3 project, 400-plus megawatts, one facility. The same week, Goldman Sachs released a note warning that hyperscaler leverage ratios have doubled to 1.8x in six months. The same week, Nikkei revealed that five US tech giants are carrying $1.65 trillion in off-balance-sheet debt — more than the debt that actually shows up on their books. The same week Intel confirmed layoffs in its data center group despite that unit growing 22% year-over-year.

These are not separate stories. They are different angles on the same question, and that question is: who's going to finance the next wave of AI infrastructure, and at what cost?

The Number That Should Stop You Cold — $334.5 Billion in Six Months

Let's start with the number that Bloomberg dropped in that AirTrunk piece, because it's the kind of number that changes how you think about an entire industry. At least $334.5 billion worth of bonds and loans have been issued so far in 2026 to fund AI infrastructure projects. That's not a projection. That's not a forecast. That's what's already been borrowed in six months.

To put that in perspective: for the entire year of 2025, the total was $185.5 billion. We have nearly doubled the full-year total in half the time. The six largest AI spenders — Microsoft, Google, Amazon, Meta, Oracle, and the usual crowd — issued $244 billion in bonds this year alone. That is fourteen times what they issued in all of 2024.

Fourteen times. In one year.

Now, if you've been following my writing this month — the Oracle debt crisis, the CoreWeave convergence, the Apple cash pile analysis — you know where I'm going with this. When borrowing accelerates this fast, something breaks. The only question is what, and when.

Force 1 — Blackstone's $3 Billion AirTrunk Loan and What It Reveals

The immediate headline is that Blackstone is seeking a $3 billion loan for AirTrunk's SYD3 data center in Sydney. A single facility. Four hundred megawatts. Three billion dollars.

But the context is what matters. AirTrunk was acquired by Blackstone in September 2024 for A$24 billion (roughly $16 billion US). Since then, the company has been on a borrowing spree that would make a hedge fund blush: $1.2 billion green loan for a Tokyo data center from 12 banks, $8 billion total committed to Japan across Tokyo and Osaka facilities representing 530 megawatts, and now $3 billion for the Sydney buildout.

And here's the interesting part. The Wall Street banks that initially underwrote these loans are quietly trying to offload them. Multiple reports this year have documented banks syndicating data center debt to private credit funds and institutional investors. They're reducing their balance sheet exposure to a sector they helped create. When the banks that wrote the checks start looking for ways to pass the paper to someone else, that's a signal. It doesn't mean the whole thing collapses tomorrow. But it means the people closest to the numbers are getting nervous.

Force 2 — Goldman Sachs Wakes Up and Smells the Debt

Goldman Sachs dropped a research note on July 19 that should be required reading for anyone in this industry. The headline finding: hyperscaler leverage ratios have doubled to 1.8x in just six months. That's debt-to-EBITDA, for those keeping score at home. When that number doubles in half a year, you're not in a normal capital cycle anymore. You're in a structural shift.

The six largest AI spenders issued $244 billion in bonds this year. That's 14 times 2024 levels. Fourteen. Times. And Goldman projects $5.8 trillion in cumulative AI capital expenditure through 2030.

Now, I want to be fair here. The bulls will tell you these companies have the balance sheets to handle it. Microsoft has a AAA rating. Amazon can tap $62 billion in orders for a $25 billion bond sale. The demand exists — for now.

But leverage ratios don't double in six months during a healthy cycle. They double when the buildout is running ahead of the revenue. And the revenue from AI infrastructure is not materializing at the same pace as the debt. Everyone knows this. The question is how long the music keeps playing.

Force 3 — The $1.65 Trillion Phantom Debt Nobody's Talking About

This is the one that keeps me up at night. On July 21, Nikkei Asia published an analysis showing that five US tech giants — the usual hyperscaler crowd — are carrying an estimated $1.65 trillion in off-balance-sheet obligations. That is hidden debt. Debt that does not show up on the balance sheet the way a bond or a bank loan does.

How does this work? Think data center leases, equipment financing, purchase commitments, structured joint ventures. These are obligations that function exactly like debt — they require future cash payments — but they don't count as debt in the leverage ratios that credit rating agencies use. Meta's hidden debt alone is estimated at $420 billion, nearly triple its transparent, on-balance-sheet debt.

I wrote about this concept in the Apple versus hyperscalers piece two days ago — the circular financing problem where data center operators borrow against future leases that depend on demand that hasn't materialized yet. What Nikkei's analysis reveals is that the hidden layer is not a rounding error. It's now larger than the visible layer. The $1.65 trillion in off-balance-sheet obligations exceeds the combined official debt of these five companies, which sits at approximately $1.35 trillion.

That is not a healthy ratio. That is the kind of ratio that creates nasty surprises when the credit cycle turns.

The Secondary Bottleneck Nobody's Talking About — Even Intel's Data Center Group Is Getting Cut

Here's the piece that connects all of this to the actual hardware you might need to run your business.

On the same day the AirTrunk loan news broke, Intel confirmed layoffs in its Data Center and AI Group. Not a struggling division — the unit that reported $5.05 billion in Q1 revenue, up 22% year-over-year. The unit that makes the server CPUs and data center architectures that power a significant portion of the world's non-Nvidia infrastructure. And Intel is cutting it.

Intel has reduced its headcount by 35,000 employees since 2024. CEO Lip-Bu Tan has been clear about "tough decisions." But cutting the one division that's actually growing — the division best positioned to capitalize on the AI buildout — tells you something about the financial pressure at the company level, not just the hyperscaler level.

If Intel, which has shed 35,000 jobs in two years, cannot afford to keep its data center group fully staffed at a time when AI infrastructure spending is supposedly booming, then the boom is not as healthy as the headlines suggest. Supply chains for server hardware, alternative chips, and enterprise infrastructure are going to feel that tightening, whether the hyperscalers acknowledge it or not.

What This Actually Means for Independent Hosting Providers

Let me be direct about what this convergence of credit market signals means for someone running actual hosting infrastructure.

First — the cost of capital for data center construction is going up, and it's going up unevenly. The PIMCO "two markets" divergence I wrote about in the Oracle piece is not theoretical. When $334.5 billion of debt enters the market in six months, the next $10 billion is going to cost more than the first. If you are planning a colocation expansion or a new facility, your financing terms are going to be worse six months from now than they are today. Lock in what you can, when you can.

Second — watch the off-balance-sheet story. When the hidden debt layers of the hyperscalers start getting priced into credit default swaps, the ripple effects will hit equipment pricing, supplier terms, and hardware availability. The Nikkei analysis is not financial journalism for entertainment. It's a leading indicator. If you rely on hyperscaler demand to set the market price for your hardware, expect that demand to soften at the margins as debt costs rise.

Third — the Intel layoffs tell you something about server hardware availability. If the second-largest server CPU manufacturer in the world is cutting the team that designs and supports those chips, you should expect longer lead times, less engineering support, and slower iteration cycles on non-hyperscaler hardware. If you run a hosting operation that depends on Intel Xeon-class processors, now is the time to evaluate your supply chain and your alternative options.

Fourth — position yourself as the capital-light alternative. Every percentage point increase in data center financing costs makes the hyperscaler buildout model more expensive and less viable over time. Independent hosting, colocation, and managed infrastructure become more attractive relative to the hyperscaler offering not because the technology is better, but because the financial model is more sustainable. When your competitor is borrowing at 10% spreads and you're already built and paid for, you win on price without needing to compete on technology.

The Structural Reality — The Credit Door Is Closing

Here's the thing about a $334.5 billion borrowing spree. It doesn't stop gently. It stops when the next lender says no, or when the next credit rating agency downgrades the next tranche of data center debt, or when the next bank decides it has enough data center exposure and starts declining the next AirTrunk-sized deal.

We are not at the stop yet. The market is still functioning. Blackstone is likely to get its $3 billion. Amazon's $25 billion bond sale was oversubscribed. Meta's $25 billion offering drew $96 billion in orders. The music is still playing.

But the signals are multiplying. Leverage doubled in six months. Hidden debt exceeds visible debt. Banks are offloading data center loans. The industry's own chip suppliers are cutting the teams that build for the boom. These are not coincidences. They are the same story told from different angles, and the story is that the AI infrastructure buildout is running on borrowed time — literally.

The Bottom Line

I've been running infrastructure for over a decade, and I've seen this pattern before. Not the scale — the scale is unprecedented. But the pattern, yes. A boom fueled by easy credit. A narrative that justifies any valuation, any leverage ratio, any off-balance-sheet structure. A moment when everyone is making money and nobody wants to ask how long it can last.

The difference this time is that the numbers are too big to ignore. $334.5 billion in six months. $1.65 trillion in hidden debt. 14x annual bond issuance. When the turning point comes — and it will come — it will reshape not just the balance sheets of the hyperscalers, but the entire economics of the infrastructure layer they run on.

The question for independent operators is not whether the turning point will arrive. The question is whether you'll have positioned yourself to survive and thrive on the other side.

Build for that world. Not this one.

-- Allan Ali, Founder