Senator Everett Dirksen is famously alleged to have remarked “a billion here, a billion there, pretty soon you’re talking real money”.

Oracle is talking real money; they’re borrowing $1.64B each working day. Mr. Market is skeptical that the real money is going to be repaid, as Caleb Mutua reports in Morgan Stanley Warns Oracle Credit Protection Nearing Record High:

A gauge of risk on Oracle Corp.’s (ORCL) debt reached a three-year high in November, and things are only going to get worse in 2026 unless the database giant is able to assuage investor anxiety about a massive artificial intelligence spending spree, according to Morgan Stanley.

A funding gap, swelling balance sheet and obsolescence risk are just some of the hazards Oracle is facing, according to Lindsay Tyler and David Hamburger, credit analysts at the brokerage. The cost of insuring Oracle Corp.’s debt against default over the next five years rose to 1.25 percentage point a year on Tuesday, according to ICE Data Services.

Mutua reports that:

The company borrowed $18 billion in the US high-grade market in September. Then in early November, a group of about 20 banks arranged a roughly $18 billion project finance loan to construct a data center campus in New Mexico, which Oracle will take over as tenant.

Banks are also providing a separate $38 billion loan package to help finance the construction of data centers in Texas and Wisconsin developed by Vantage Data Centers,

But notice that only $18B of this debt appears on Oracle’s balance sheet. Despite that, their credit default swaps spiked and the stock dropped 29% in the last month.

Below the fold I look into why Oracle and other hyperscalers desperate efforts to keep the vast sums they’re borrowing off their books aren’t working.

Part of the reason the market is unhappy started in mid-September with The Economist’s The $4trn accounting puzzle at the heart of the AI cloud. It raised the issue that I covered in Depreciation, that the hardware that represents about 60% of the cost of a new AI data center doesn’t last long. It took a while for the financial press to focus on the issuea, but now they have.

The most recent one I’ve seen was triggered by the outage at the CME (caused by overheating in Chicago in November!). In AI Can Cook the Entire Market Now Tracy Alloway posted part of the transcript of an Odd Lots podcast with Paul Kedrosky pointing out a reason I didn’t cover why the GPUs in AI data centers depreciate quickly:

When you run using the latest, say, an Nvidia chip for training a model, those things are being run flat out, 24 hours a day, seven days a week, which is why they’re liquid-cooled, they’re inside of these giant centers where one of your primary problems is keeping them all cool. It’s like saying ‘I bought a used car and I don’t care what it was used for.’ Well, if it turns out it was used by someone who was doing like Le Mans 24 hours of endurance with it, that’s very different even if the mileage is the same as someone who only drove to church on Sundays.

These are very different consequences with respect to what’s called the thermal degradation of the chip. The chip’s been run hot and flat out, so probably its useful lifespan might be on the order of two years, maybe even 18 months. There’s a huge difference in terms of how the chip was used, leaving aside whether or not there’s a new generation of what’s come along. So it takes us back to these depreciation schedules.

There was a similar problem after the Ethereum merge:

73% of Ethereum miners have just given up: “About 10.6 million RTX 3070 equivalents have stopped mining since the merge.”

We strongly recommend that you do not hit eBay for a cheap video card, despite the listings reassuring you that this card was only used by a little old lady to play Minecraft on Sundays and totally not for crypto mining, and that you should ignore the burnt odor and the charred RAM. Unless you’re poor, and the card’s so incredibly cheap that you’re willing to play NVidia Roulette.

How well do miners treat their precious babies? “GPU crypto miners in Vietnam appear to be jet washing their old mining kit before putting the components up for sale.” There are real cleaning methods that involve doing something like this with liquid fluorocarbons — but the crypto miners seem to be using just water.

But this depreciation problem is only one part of why the market is skeptical of the hyperscalers technique for financing their AI data centers. The technique is called Conduit Debt Financing, and Les Barclays’ Unpacking the Mechanics of Conduit Debt Financing provides an accessible explanation of how it works:

Conduit debt financing is a structure where an intermediary entity (the “conduit”) issues debt securities to investors and passes the proceeds through to an end borrower. The key feature distinguishing conduit debt from regular corporate bonds is that the conduit issuer has no substantial operations or assets beyond the financing transaction itself. The conduit is purely a pass-through vehicle, the debt repayment relies entirely on revenues or assets from the ultimate borrower.

Think of it this way: Company A wants to borrow money but doesn’t want that debt appearing on its balance sheet or affecting its credit rating. So it works with a conduit entity, Company B, which issues bonds to investors. Company B takes that capital and uses it to build infrastructure or acquire assets that Company A needs. Company A then enters into long-term lease or service agreements with Company B, and those payments service the debt. On paper, Company A is just a customer making payments, not a debtor owing bondholders.

The structure creates separation. The conduit issuer’s creditworthiness depends on the revenue stream from the end user, not on the conduit’s own balance sheet (because there isn’t really one). This is why conduit debt is often referred to as “pass-through” financing, the economics flow through the conduit structure to reach the underlying obligor.

The article continues to examine Meta’s deal in great detail, and notes some of the legal risks of this technique:

Legal risks when things break: Substantive consolidation (court merges conduit with sponsor), recharacterization (lease treated as secured financing), and fraudulent transfer challenges. The structures haven’t been stress-tested yet because hyperscalers are wildly profitable. But if AI monetization disappoints or custom silicon undercuts demand, we’ll discover whether bondholders have secured claims on essential infrastructure or are functionally unsecured creditors of overleveraged single-purpose entities.

The article asks the big question:

Why would Meta finance this via the project finance markets? And why does it cost $6.5 billion more?

That’s how much more Meta is paying to finance this new AI data center using the project finance market versus what they could have paid had they used traditional corporate debt. So why on earth is this being called a win? And even crazier, why are other AI giants like Oracle and xAI looking to copy it?

The $6.5B is the total of the 1% extra interest above Meta’s corporate bond rate over the 20 years.

If Counduit Debt Financing is a standard tool of project finance, why is Mr. Market unhappy with the hyperscalers’ use of it? Jonathan Weil’s somewhat less detailed look at Meta’s $27B deal in AI Meets Aggressive Accounting at Meta’s Gigantic New Data Center reveals how they are pushing the envelope of GAAP (Generally Accepted Accounting Principles):

Construction on the project was well under way when Meta announced a new financing deal last month. Meta moved the project, called Hyperion, off its books into a new joint venture with investment manager Blue Owl Capital. Meta owns 20%, and funds managed by Blue Owl own the other 80%. Last month, a holding company called Beignet Investor, which owns the Blue Owl portion, sold a then-record $27.3 billion of bonds to investors, mostly to Pimco.

Meta said it won’t be consolidating the joint venture, meaning the venture’s assets and liabilities will remain off Meta’s balance sheet. Instead Meta will rent the data center for as long as 20 years, beginning in 2029. But it will start with a four-year lease term, with options to renew every four years.

This lease structure minimizes the lease liabilities and related assets Meta will recognize, and enables Meta to use “operating lease,” rather than “finance lease,” treatment. If Meta used the latter, it would look more like Meta owns the asset and is financing it with debt.

Under GAAP, when would Meta be required to treat it as a finance lease?

The joint venture is what is known in accounting parlance as a variable interest entity, or VIE for short. That term means the ownership doesn’t necessarily reflect which company controls it or has the most economic exposure. If Meta is the venture’s “primary beneficiary”—which is another accounting term of art—Meta is required to consolidate it.

Under the accounting rules, Meta is the primary beneficiary if two things are true. First, it must have “the power to direct the activities that most significantly impact the VIE’s economic performance.” Second, it must have the obligation to absorb significant losses of the VIE, or the right to receive significant benefits from it.

Does Meta have “the power to direct the activities“ at the data center it will operate?:

Blue Owl has control over the venture’s board. But voting rights and legal form aren’t determinative for these purposes. What counts under the accounting rules is Meta’s substantive power and economic influence. Meta in its disclosures said “we do not direct the activities that most significantly impact the venture’s economic performance.” But the test under the accounting rules is whether Meta has the power to do so.

Does Meta receive “significant benefits”? Is it required to “absorb losses”?:

The second test—whether Meta has skin in the game economically—has an even clearer answer. Meta has operational control over the data center and its construction. It bears the risks of cost overruns and construction delays. Meta also has provided what is called a residual-value guarantee to cover bondholders for the full amount owed if Meta doesn’t renew its lease or terminates early.

The lease is notionally for 20 years but Meta can get out every four years. Is Meta likely to terminate early? In other words, how likely in 2041 is Meta to need an enormous 16-year old data center? Assuming that the hardware has an economic life of 2 years, the kit representing about 60% of the initial cost would be 8 generations behind the state of the art. In fact 60% of the cost is likely to be obsolete by the first renewal deadline, even if we assume Nvidia won’t actually be on the one-year cadence it hasa announced.

But what about the other 40%? It has a longer life, but not that long. The reason everyone builds new data centers is that the older ones can’t deliver the power and cooling current Nvidia systems need. 80% of recent data centers in China are empty because they were built for old systems.

But the new ones will be obsolete soon:

Today, Nvidia’s rack systems are hovering around 140kW in compute capacity. But we’ve yet to reach a limit. By 2027, Nvidia plans to launch 600kW racks which pack 576 GPU dies into the space one occupied by just 32.

Current data centers won’t handle these systems - indeed how to build data centers that do is a research problem:

To get ahead of this trend toward denser AI deployments, Digital Realty announced a research center in collaboration with Nvidia in October.

The facility, located in Manassas, Virginia, aims to develop a new kind of datacenter, which Nvidia CEO Jensen Haung has taken to calling AI factories, that consumes power and churn out tokens in return.

If the design of data centers for Nvidia’s 2027 systems is only now being researched, how likely is it that Meta will renew the lease on a data center built for Nvidia’s 2025 systems in 2041? So while the risk that Meta will terminate the lease in 2029 is low, termination before 2041 is certain. And thus so are residual-value guarantee payments.

How does the risk of non-renewal play out under GAAP?

Another judgment call: Under the accounting rules, Meta would have to include the residual-value guarantee in its lease liabilities if the payments owed are “probable.” That could be in tension with Meta’s assumption that the lease renewal isn’t “reasonably certain.”

If renewal is uncertain, the guarantee is more likely to be triggered. But if the guarantee is triggered, Meta would have to recognize the liability.

Weil sums it up concisely:

Ultimately, the fact pattern Meta relies on to meet its conflicting objectives strains credibility. To believe Meta’s books, one must accept that Meta lacks the power to call the shots that matter most, that there’s reasonable doubt it will stay beyond four years, and that it probably won’t have to honor its guarantee—all at the same time.

These accounting shenanigans explain why Sam Altman said the queit part out loud recently and then had to walk it back. Jose Antonio Lanz reports this in OpenAI Sought Government Loan Guarantees Days Before Sam Altman’s Denial (my emphasis):

OpenAI explicitly requested federal loan guarantees for AI infrastructure in an October 27 letter to the White House—which kindly refused the offer, with AI czar David Sacks saying that at least 5 other companies could take OpenAI’s place—directly contradicting CEO Sam Altman’s public statements claiming the company doesn’t want government support.

The 11-page letter, submitted to the Office of Science and Technology Policy, called for expanding tax credits and deploying “grants, cost-sharing agreements, loans, or loan guarantees to expand industrial base capacity” for AI data centers and grid components. The letter detailed how “direct funding could also help shorten lead times for critical grid components—transformers, HVDC converters, switchgear, and cables—from years to months.”

After this PR faux pas some less obvious way taxpayer dollars could keep the AI bubble inflating had to be found. Just over two weeks later Thomas Beaumont reported that Trump signs executive order for AI project called Genesis Mission to boost scientific discoveries:

Trump unveiled the “Genesis Mission” as part of an executive order he signed Monday that directs the Department of Energy and national labs to build a digital platform to concentrate the nation’s scientific data in one place.

It solicits private sector and university partners to use their AI capability to help the government solve engineering, energy and national security problems, including streamlining the nation’s electric grid, according to White House officials who spoke to reporters on condition of anonymity to describe the order before it was signed.

This appears to be a project of David Sacks, the White House AI advisor and a prominent member of the “PayPal Mafia”. Sacks was the subject of a massive, 5-author New York Times profile entitled Silicon Valley’s Man in the White House Is Benefiting Himself and His Friends:

• Mr. Sacks has offered astonishing White House access to his tech industry compatriots and pushed to eliminate government obstacles facing A.I. companies. That has set up giants like Nvidia to reap an estimate of as much as $200 billion in new sales.
• Mr. Sacks has recommended A.I. policies that have sometimes run counter to national security recommendations, alarming some of his White House colleagues and raising questions about his priorities.
• Mr. Sacks has positioned himself to personally benefit. He has 708 tech investments, including at least 449 stakes in companies with ties to artificial intelligence that could be aided directly or indirectly by his policies, according to a New York Times analysis of his financial disclosures.
• His public filings designate 438 of his tech investments as software or hardware companies, even though the firms promote themselves as A.I. enterprises, offer A.I. services or have A.I. in their names, The Times found.
• Mr. Sacks has raised the profile of his weekly podcast, “All-In,” through his government role, and expanded its business.

The article quotes Steve Bannon:

Steve Bannon, a former adviser to Mr. Trump and a critic of Silicon Valley billionaires, said Mr. Sacks was a quintessential example of ethical conflicts in an administration where “the tech bros are out of control.”

“They are leading the White House down the road to perdition with this ascendant technocratic oligarchy,” he said.

Gary Marcus asked Has the bailout of generative AI already begun?:

“The way this works”, said an investor friend to me this morning: “is that when Nvidia is about to miss their quarter, Jen Hsun calls David Sacks, who then gets this government initiative to place a giant order for chips that go into a warehouse.”

I obviously can’t confirm or deny that actually happened. My friend might or might not have been kidding. But either way the White House’s new Science and AI program, Genesis, announced by Executive Order on Monday, does seem to involve the government buying a lot of chips from a lot of AI companies, many of which are losing money.

And David Sack’s turnaround from “read my lips, no AI bailout” (November 6) to “we can’t afford to [let this all crash]” tweet (November 24) came just hours before the Genesis announcement.

I think the six companies Sacks was talking about are divided into two groups:

• OpenAI, Anthropic and xAI, none of whom have a viable business model.
• Meta, Google and Microsoft, all of whom are pouring the cash from their viable business models into this non-viable business,

This is the reason why the hyperscalers are taking desperate financial measures. They are driven by FOMO but they all see the probability that the debt won’t be paid back. Where is the revenue to pay them back going to come from? It isn’t going to come from consumers, because edge inference is good enough for almost all consumers (which is why 92% of OpenAI’s customers pay $0). It isn’t going to come from companies laying off hordes of low-paid workers, because they’re low-paid.

So before they need to replace the 60% of the loan’s value with the next generation of hardware in 2027 they need to find enterprise generative AI applications that are so wildly protiftable for their customers that they will pay enough over the cost of running the applications to cover not just the payments on the loans but also another 30% of the loan value every year. For Meta alone this is around $30B a year!

And they need to be aware that the Chinese are going to kill their margins. Thanks to their massive investments in the “hoax” of renewable energy, power is so much cheaper in China that systems built with their less efficient chips are cost-competitive with Nvida’s in operation. Not to mention that the Chinese chip makers operate on much lower margins than Nvidia. Nvidia’s chips will get better, and so will the Chinese chips. But power in the US will get more expensive, in part because of the AI buildout, and in China it will get cheaper.

This won’t end well