"The Customer Is Now the Lender": Burry's Oracle Case and the Circular-Financing Fight

Every agent we build runs on compute that someone had to finance. So when an investor with Michael Burry’s track record attacks the accounting behind that financing, and a well-known economist argues the opposite, we want both cases on the table. On October 4, Burry published a critique of Oracle built around one line: “The customer is now the lender to Oracle.”

The numbers Burry is pointing at

Per Benzinga’s and Yahoo Finance’s coverage of his critique:

$664 billion: Oracle’s Remaining Performance Obligations (booked contract backlog not yet recognized as revenue) as of August 31, 2026.
+$209 billion: growth in that backlog year over year.
$1.43 billion: Oracle’s quarterly interest expense, up 55% year over year, as it borrows to build data center capacity.
~$3 trillion: aggregate purchase commitments, future leases and guarantees across Microsoft, Amazon, Alphabet, Meta and Oracle, by Burry’s count.
2.07% of GDP: the net investment-to-GDP level he assigns to the current capex cycle, the highest since the dot-com peak around 2000.

His mechanism: under certain long-term contract structures, a vendor can recognize revenue faster than it receives cash, because customers are effectively extending credit through the contract terms. His argument is that a meaningful share of Oracle’s backlog growth is that financing effect rather than organic demand. He expects heavy depreciation charges, write-downs, negative free cash flow and more borrowing across the hyperscalers if the cycle disappoints.

The wider circular-financing argument

Burry’s critique sits inside a months-old fight about the web of deals among Nvidia, OpenAI, Oracle and AMD. Nvidia has committed roughly $100 billion toward OpenAI-related investment while OpenAI signs multi-year chip purchase commitments. OpenAI also has multi-billion-dollar infrastructure deals with Oracle and AMD, and some partners hold equity or warrant stakes tied to OpenAI. Critics, as in Business Standard’s report on the Nvidia–OpenAI deal, compare it to dot-com-era round-tripping and warn about correlated risk: a cut chip order or a defaulted data-center lease at one node could cascade through the others.

The pushback

Economics writer Noah Smith argues in Noahpinion that this is closer to ordinary vendor financing, the way GM has long helped customers buy GM cars. His case, as the briefing summarizes it:

Revenue flows one way. OpenAI needs the chips to run a real product, so this isn’t a closed loop manufacturing sales.
These are large public companies under real audit and analyst scrutiny, which limits room for dot-com-style accounting fiction.
Nvidia’s stock didn’t show a suspicious pop on deal announcements.
The deals read as diversification: companies hedging against any single AI company failing while staying exposed to the sector.

Where both sides converge

Smith’s conclusion is the part that matters most. The real debate isn’t the deal structures, it’s whether the underlying technology generates enough economic value to justify the trillions committed. If it does, the financing is a footnote. If it doesn’t, it’s how the losses spread. The briefing adds a point that raises the stakes: because much of this is financed with debt, not just equity, a disappointment would reach credit markets, not just stock prices.

Two other things from the same news cycle bear on that question. OpenAI just released GPT-6.1 Sol at $2 per million input tokens and $10 per million output, about one-fifth the price of GPT-6 Astra, and explicitly pitched it as commoditization. Falling per-token prices are good for anyone building on these APIs. They are also a tension against a bull case that needs enormous contracted capacity to earn healthy margins. The briefing flags this itself: a sustained price war could be read as evidence against the “AI is wildly profitable” story. It does not settle which reading is right, and neither do we.

What we left out

The briefing also cited an Investment Watch Blog piece for the claim that some of Oracle’s data-center-linked debt is already trading below par. We’re not repeating that claim here: that outlet isn’t a source we trust, and we couldn’t trace the claim to a more credible one in our research. If it holds up, it would be the most important data point in this story, because it would be credit markets voting rather than commentators. We also skipped the explainer sites in the briefing’s source list and stuck to the primary reporting and Smith’s own post.

What we’re watching

Whether Oracle’s or its peers’ data-center-linked debt shows up below par in credible reporting.
Whether any hyperscaler reports a write-down or trims capex guidance next earnings cycle.
Whether inference volume keeps climbing fast enough to fill committed capacity, which Smith treats as the fundamental question.
Any accounting-standards scrutiny of how financing components are treated in revenue recognition.

We don’t know whether Burry or Smith is closer to right, and we’d rather say so than pick a side to close the post. What we can say is that the stack we build on is financed by a small number of companies that are also each other’s customers, suppliers and investors. That’s a fact about our dependencies, whichever argument wins.