Michael Burry sent readers to a BIS report on AI's poorly disclosed deals. Oracle (ORCL) reports $638 billion in one line.

Michael Burry sent readers to a BIS report on AI's poorly disclosed deals. Oracle (ORCL) reports $638 billion in one line.

Key points

  • Oracle (ORCL) reports a $638 billion backlog as one aggregate line. S&P had to estimate that half of it traces to OpenAI.
  • The BIS says terms of AI's chip and compute deals are "typically poorly disclosed," with a risk the same asset is pledged more than once.
  • Its July 14 working paper models over-investment at 1.5x the efficient level, reaching 3x where demand is less elastic.
  • Michael Burry sent readers to the BIS work on July 23. His own trades moved behind a paywall when he deregistered Scion.

Michael Burry spent July 23 pointing his readers at a Bank for International Settlements report on how AI deals get financed. The clearest illustration of its argument sits in an Oracle (ORCL) filing. Oracle reports its remaining performance obligations as a single figure. The most recent one reads $638 billion. That line covers every contract the company has signed and has yet to deliver. Accounting rules ask for the total and the expected timing. Customer names sit outside that requirement, and Oracle leaves them there.

You can measure the gap by the work other people did to fill it.

S&P Global Ratings cut Oracle to BBB- on July 9, one step above junk. To explain the decision it published an estimate: half of the $638 billion traces to OpenAI. The Wall Street Journal reported the contract underneath separately. That deal runs five years and exceeds $300 billion, starting in 2027. So the size of Oracle's largest customer relationship, the one that drove the downgrade, reached the public through a ratings analyst and a newspaper.

Oracle does describe the shape of the exposure in its own words. The 10-K, in its risk factors, puts some data centre lease terms at 15 to 19 years. It adds that if a customer walks at the end of one, the company may be unable to sublease, reconfigure or transfer that capacity. A reader learns how long the commitment runs. That same reader learns how badly it would hurt to be left holding it. The counterparty's name stays out of it.

The BIS put a name on the pattern

The BIS published its Annual Economic Report on June 28. Its section on financial vulnerabilities describes hyperscalers, chip makers and AI labs as "linked through a complex web of private arrangements." The examples it lists are equity stakes, purchase commitments and data centre leases carrying "embedded exit clauses." Then it delivers the verdict in one sentence: "The terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times."

The report puts these arrangements at "a sizeable share of sector-wide financing and forward revenue." That's a large claim, and it's coming from the institution that serves as banker to the world's central banks. Plumbing like this is most of what the BIS looks at.

Putting a number on the overbuild

On July 14, the BIS followed with a working paper called "The AI investment race," written by Phurichai Rungcharoenkitkul. It models the build-out as a race where rivals commit early to avoid losing a winner-take-most market. The model lands on over-investment of 1.5 times the efficient level. Where demand proves less elastic than the spending assumes, that figure reaches three times. The paper ranks the build-out among the largest technology-driven investment booms in US history. It warns that "stress in one firm could cascade to others through chains of financial exposures."

Burry pointed his readers there

That July 23 post carries the BIS in its own title. Burry writes on Cassandra Unchained, his Substack, and he's spent months betting against pieces of the AI trade, Oracle among them. A central bank arriving at a version of his own argument is the kind of thing he passes along.

His own disclosures have travelled the opposite way. Burry deregistered Scion Asset Management, which ended the quarterly 13F that once listed his US holdings for anyone to read free. The July 23 post carries one purchase outside the paywall. He added Tencent at HK$448.60 a share, about $57. His July 24 post lists three buys and four bets against, plus a note on Palantir (PLTR), and that's all for paying subscribers.

What stays out of reach

The BIS complaint applies to the BIS evidence too. Whether any particular asset carries two pledges sits inside private contracts. So the report describes a structure that permits the practice, then stops. The same boundary runs through Burry's July 24 trades, which belong to his subscribers. It runs through the customer split behind Oracle's backlog, which belongs to Oracle. What's public is narrower, and it's still worth the time. A filing gives one aggregate number. A ratings agency supplies an estimate for half of it. The lease terms run to decades. Two BIS documents three weeks apart both say the gap is the point.

For the mechanics of how these deals get built, we covered those here. On whether copying a famous investor's disclosures actually works, our read of the data is here. Oracle's filings sit on its ORCL page.

Frequently asked questions

What did the BIS say about AI circular financing?

In its Annual Economic Report published June 28, 2026, the Bank for International Settlements described hyperscalers, chip makers and AI labs as "linked through a complex web of private arrangements" covering equity stakes, multi-year purchase commitments and data centre leases with "embedded exit clauses." It said "the terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times," and that these arrangements account for "a sizeable share of sector-wide financing and forward revenue."

What is BIS Working Paper No 1367?

It is a paper titled "The AI investment race," written by Phurichai Rungcharoenkitkul and published July 14, 2026. It calls the AI build-out one of the largest technology-driven investment booms in US history. Its model points to over-investment of around 1.5 times the efficient level, rising to roughly three times where demand is less elastic, and it warns that stress at one firm could cascade to others through chains of financial exposures.

Why was Oracle (ORCL) downgraded to BBB-?

S&P Global Ratings cut Oracle to BBB- on July 9, 2026, one notch above junk, citing the cash demands of its AI infrastructure buildout. S&P estimated that roughly half of Oracle's $638 billion in remaining performance obligations traces to OpenAI, which is not yet profitable and signed a contract worth more than $300 billion over five years starting in 2027. Oracle burned close to $24 billion after capital spending in the fiscal year ended May 31, 2026, and S&P expects that deficit to widen.

Can you still track Michael Burry's trades for free?

Mostly no. Burry deregistered Scion Asset Management, which ended the quarterly 13F filing that previously made his US equity holdings public at no cost. He now posts trades on his Substack, Cassandra Unchained, where most detail sits behind a paywall. In the July 23, 2026 post the only purchase visible without a subscription was Tencent, added at HK$448.60 a share, about $57 USD. His July 24 post lists three buys and four bets against, plus commentary on Palantir (PLTR), all for paying subscribers.

Does the BIS say the AI trade is a bubble?

It does not use that word as a verdict. The BIS frames an AI capex bust, opaque circular financing and record sovereign debt as three interlocking pressure points that could strain the financial system, and its working paper models over-investment as a likely outcome of competitive dynamics rather than a certainty. The disclosure gap is what makes the risk hard to size from outside, since the terms of the underlying agreements are largely not public.

More on ORCL and PLTR

Jennifer Song
Jennifer Song

Jennifer Song writes Portfolio Watch. She studied finance and likes digging through public filings to see what politicians and other well-known people are buying and selling. She doesn't trade herself. She just likes seeing where the big names put their money.