Oracle has $664 billion of AI contracts. Its data center debt is already trading below par

by

Oracle’s AI business has more contracted revenue than ever.

Its lenders are already putting a discount on the infrastructure needed to deliver it.

Oracle reported $664 billion of remaining performance obligations in September, up $209 billion from a year earlier. It booked more than $30 billion of additional AI cloud contracts in one quarter. Cloud infrastructure revenue jumped 121% to $7.4 billion.

That is the part everyone wants to show you.

Now look at the other side.

Oracle spent $28.5 billion on capital expenditures in one quarter.

Operating cash flow was $23.1 billion.

Free cash flow was negative $5.4 billion.

The company expects to spend another $90 billion to $95 billion during fiscal 2027.

So Oracle has contracts.

It has customers.

It has GPUs.

It has demand.

What it doesn’t have yet is all the physical infrastructure required to turn those contracts into operating cash.

That is where the financing machine comes in.

Project Jupiter in New Mexico is a perfect case study.

The facility is tied to Oracle’s OpenAI agreement and was financed with roughly $18 billion of bank loans. Those loans were supposed to be distributed to other investors after the banks originated them.

That distribution has run into trouble.

By September, the loans were being quoted around 89 to 91 cents on the dollar. Reuters reported that efforts to distribute the debt more broadly had stalled amid concerns about Oracle’s rising borrowing and weaker credit profile.

That’s not an AI demand forecast.

That’s a lender putting a price on the financing.

Then Oracle sent a force majeure notice to the Project Jupiter developer.

Oracle says the project remains on schedule. The notice does not mean Oracle is abandoning the site. It gives Oracle contractual protection to delay certain payments if the facility cannot meet its target for reasons covered by the agreement.

But look at the timing.

The project has faced repeated power and permitting problems. The natural-gas pipeline intended to serve the site has been pushed to February 2027 after regulatory setbacks.

On September 10, Oracle’s co-CEO Clay Magouyrk dismissed the idea that a large infrastructure project should depend on every single deadline being hit.

His words were unusually revealing:

Anyone whose plan depends on “100% achievement of every one of their deliverables,” he said, has what Oracle calls “a bad plan.”

Two weeks later, Oracle invoked force majeure on the same project.

That doesn’t prove the project is failing.

It does show something else.

The financial commitments are arriving before the physical assets are finished.

And that is where the AI buildout starts to resemble an older boom.

During the late-1990s telecom expansion, companies borrowed enormous amounts to build fiber networks based on expectations for explosive bandwidth demand. The physical networks were expensive and largely irreversible once built.

Then demand failed to arrive at the expected pace.

The result wasn’t simply lower stock prices.

Debt remained while the price of network capacity collapsed.

Congressional testimony later described the bust as excessive investment in communications capacity financed through capital loans, followed by a demand slowdown that left companies unable to support their debt.

One contemporary account of the fiber bust described the problem even more clearly: companies had raised billions to build networks, but demand didn’t materialize quickly enough, while excess capacity pushed prices down.

The AI version has a crucial difference.

The demand today is much more tangible.

Oracle isn’t building empty data centers and hoping someone eventually shows up.

It already has customers signing enormous contracts.

That is why the $664 billion backlog can be both real and misleading at the same time.

A contract isn’t a completed data center.

A contract doesn’t generate cash until Oracle can deliver the computing capacity.

And computing capacity doesn’t exist until someone finances the land, buildings, power systems, chips, cooling equipment and network connections.

That creates a chain:

OpenAI signs the contract.

Oracle promises the capacity.

Banks finance the data center.

Developers build it.

Oracle pays for the infrastructure.

The facility turns on.

Only then does the contracted revenue become operating cash.

The weak link right now isn’t necessarily the customer.

It’s the enormous amount of money that has to survive the middle of that chain.

And Oracle is increasingly relying on financing to bridge it.

In Q1 alone, Oracle raised $20 billion through an equity sale. Customer prepayments also helped its cash flow, with Oracle reporting $11.4 billion of customer prepayments during the quarter. Even after those inflows, gross capex exceeded operating cash flow.

This is why the credit market matters more here than another stock chart.

Equity investors can believe the $664 billion backlog will eventually become revenue.

Credit investors have a simpler question:

Who pays if the building is late, the power isn’t ready, the customer changes its plans, or the financing cannot be rolled over?

That question is already showing up in the price of the debt.

And Oracle isn’t alone.

Reuters reported this month that corporate bond buyers are becoming more selective about AI-related debt. Spreads on AI-linked bonds were around 115 basis points compared with 78 basis points for the broader market, while hyperscaler debt issuance is projected to rise sharply.

Meanwhile, SoftBank is preparing one of its largest junk-bond offerings to fund its OpenAI investment, with more than $11 billion of high-yield debt planned.

The financing machine is getting larger while the assets being financed are still under construction.

That is the part of this boom I would watch.

Not whether AI demand exists.

It clearly does.

Watch the distance between contracted demand and completed infrastructure.

If that gap keeps widening, companies have to keep borrowing to carry it.

If lenders start demanding higher returns, the economics change.

If projects slip, lease obligations and construction loans don’t magically disappear.

And if AI computing prices eventually fall faster than the debt costs supporting the infrastructure, the industry can have plenty of customers and still have a financing problem.

That’s what makes Oracle such a useful case study.

The $664 billion number tells you what customers have promised.

The 89-cent loan tells you what creditors are beginning to think about the machinery required to deliver it.

Those are two very different prices for the same AI boom.

Not financial advice.

Got a news tip or correction? Let us know

If you got something out of this, please chip in to keep this site running, or subscribe to go ad-free.

0 views