Act 01 — A Saturday in Q1 2026What you notice first

A hypothetical scenario · thought experiment, not a forecast

Your setup. Half the agents throw errors. The one skill your hub depends on: timeout. A message to Claude won't go through. OpenAI API dead for three hours.

Stock widget. Oracle down 38%. Nvidia down 24. Microsoft down 12. Amazon down 9. Alphabet down 11. More than half a trillion in market value gone overnight.

Email. Anthropic doubled its price list overnight, token limits cut in half. "In response to the global compute shortage." OpenAI follows Monday.

" was right. Again."

— the headline everywhere. Working hypothesis I started this research with. The next three acts test it[05].

Act 02 — The mechanicsHow money moves in a loop between the three main players

Before we come back to the Saturday-morning scenario, let's look at what actually happened in Q1 2026. Three main players. Three steps. The same pot of money.

Chart 01 · The loop
How money moves in a loop between the three main players
Q1 2026 · BOOK PROFIT ≠ CASH
Book profit
$67B
promised on the balance sheet
Cash received
$1.5B
actually wired
Ratio 45 : 1 · For every dollar of cash, 45 are booked.
As of June 1, 2026 · Q1 2026 quarterly filings
The loop
Three steps, the same pot of money
① Step 1
NVIDIA
invests $30B as an equity stake in OpenAI.
② Step 2
OPENAI
commits $22B in compute purchases to CoreWeave.
③ Step 3
COREWEAVE
buys GPUs from Nvidia for $7.5B.
The money is back at Nvidia. Book profits booked, almost no cash moved.
More loops run around this one: Microsoft feeds OpenAI with $13B equity plus the Azure cloud backlog. Oracle holds the $300B Stargate contract with OpenAI, which doesn't pay out until 2027. Amazon and Alphabet together hold roughly $50B in Anthropic, which in turn buys from AWS and Google Cloud. Goldman Sachs estimates the circular share at under 15% of Nvidia's revenue through 2027[13]. The skeptics do the math differently.
Sources: CoreWeave 8-K SEC, Sep 9, 2025[10]; CNBC OpenAI $122B round, Mar 31, 2026[01]; Oracle Q3 FY2026 IR release[09].

Act 03 — In the booksWhat's actually on the AI giants' balance sheets

The $67B in book profits from the loop is not a projection. It sits in the Q1 2026 SEC filings — six companies, six balance sheets. Left side: what was booked. Right side: what actually became cash. All of it gains on AI stakes and contracted-but-not-yet-delivered services (). US-GAAP standard, all of it legal.

Chart 02 · Today's books
What Big Tech reports, and how much of it is actually cash
Alphabet · Q1 2026 book profit
$28.7B[07]
From the revaluation of the Anthropic stake. Anthropic's Series G in February 2026 ($380B valuation) lifts Alphabet's 14% stake to $53B on paper.
Cash actually received
$0
Pure mark-to-market entry under . Not a cent flowed from Anthropic to Alphabet.
Almost half of Alphabet's quarterly profit ($28.7B of $62.6B) is a circular deal: Alphabet invests in Anthropic → Anthropic's valuation rises through new rounds → Alphabet's own stake gets marked up. The more Alphabet puts in, the larger the paper gain.
Amazon · Q1 2026 book profit
$16.8B[08]
Pre-tax mark-to-market on the Anthropic stake. $8B in, now worth over $70B — an 8.75x paper return in two years.
Free cash flow Q1 2026
$1.2B
−95% year over year. With $30.3B net income and $44B CapEx, the real money goes into datacenters, not the till.
Anthropic locks itself into AWS Trainium with a $100B through 2036 — Amazon invests in Anthropic, Anthropic buys AWS compute back. Book profit up, down.
Nvidia · Q1 FY2027 book profit
$15.9B[02]
Unrealized equity gains. $13.4B on the CoreWeave stake (publicly listed), $2.6B on private stakes (mainly OpenAI).
Cash from the stakes
$0
OpenAI buys Nvidia GPUs — the cash flow comes back as chip revenue through cloud providers, not as a dividend. Nvidia also guarantees a $6.3B backstop to CoreWeave.
Nvidia sells chips to OpenAI, invests $30B back into OpenAI, and lets the resulting valuation gains show up in its own books. A closed liquidity loop.
Microsoft · 9M FY2026 book profit
$5.9B[06]
from the OpenAI recapitalization in October 2025. Equity method (HLBV), not mark-to-market. Microsoft's OpenAI stake: roughly 27%, paper value around $135B.
Revenue share per quarter
$1.5B
20% of OpenAI revenue through 2030 (Wedbush estimate). Separate licensing flow, not return on equity. Exclusivity expires April 2026.
Microsoft feeds OpenAI's valuation through Azure commitments, books the resulting dilution gain — and collects the revenue share as a second stream in parallel. Two sources, one loop.
Oracle · Q3 FY2026 backlog
$553B[09]
RPO +325% year over year. More than half is the $300B OpenAI contract (Stargate initiative).
Cash from the OpenAI deal
$0
Contract starts 2027. Oracle's Q3 cloud revenue of $8.9B comes entirely from other customers.
The largest cloud contract of all time sits in the books, the money flows starting 2027 at the earliest — and the customer (OpenAI) isn't profitable. Backlog without a credit rating.
CoreWeave · Q1 2026 backlog
$99.4B[11]
48x the current quarterly revenue. Top customers: OpenAI ($22.4B), Meta ($35.2B incl. April 2026 expansion), Microsoft.
Q1 2026 revenue
$2.1B
+112% year over year, but $740M net loss — 36% loss margin. 2026 CapEx guidance: $31B to $35B.
Nvidia holds 11% of CoreWeave, is at the same time its main chip supplier, and guarantees $6.3B in unused capacity. Investor, supplier, and backstop guarantor in one.
Q1 2026 aggregate: roughly $67B in book profits sit against $1.5B in actual cash. Alphabet, Amazon, Nvidia and Microsoft alone booked $67.3B in mark-to-market and dilution gains into their quarterly P&L. Real cash flow from the stakes: effectively zero. On top of that, $652B in backlog at Oracle and CoreWeave, of which barely $2B actually moved in Q1. All legal, all readable in the filings — but every piece a link in the same chain.
Sources: filings Q1 2026 / Q3 FY2026 at SEC EDGAR, IR earnings releases. Book profits = US-GAAP profit from "Other Income (Expense)" via mark-to-market adjustments on equity stakes (ASC 321) or dilution gains under the equity method. Full source list at the end of the article.

Act 04 — Historical comparisonsEnron or Cisco?

Looking at $67B in book profits without cash, the reflex is Enron. drew the Cisco parallel in a Substack post on May 22, 2026[05]. Three days later, after Nvidia hit back with a seven-page sell-side memo, he sharpened the point[03]. First what he rules out, then what he means.

Chart 03 · Enron 1991–2001
What Burry does NOT mean: when became fraud
Enron Corporation logo (Paul Rand, 1997)
Fiscal year 2000 · BOOK REVENUE ≠ CASH
Reported revenue
$100.8B
mark-to-market from energy trading
Operating cash flow
$4.8B
actually realized
Ratio 21 : 1 · For every dollar of cash, $21 of book revenue.
From 1991 SEC approval to December 2001 bankruptcy
The Enron pattern
Three steps, the same mechanism. Book profit from future estimates, cash never arrives
① Step 1
MARK-TO-MARKET
(Enron COO) gets SEC approval in 1991 for mark-to-market in energy trading. First time outside of banks.
② Step 2
PHANTOM PROFITS
A 20-year gas contract with estimated full value $1B is booked as profit immediately. Cash arrives, if at all, over twenty years.
③ Step 3
SPEs AS COVER-UP
When the cash didn't come, around 3,000 (Raptor, LJM, Chewco) absorbed the losses, collateralized with Enron stock.
As long as the stock was high, the gaps stayed off-balance-sheet and invisible.
What happened next: in October 2001 the stock fell from $90 to under $1. The SPEs went insolvent, their losses landed back on Enron's balance sheet. Bankruptcy in six weeks, $67B in market value gone. Skilling, Lay, and Fastow went to prison. Mark-to-market approval for energy traders was never granted again.
Sources: McLean / Elkind "The Smartest Guys in the Room" (2003); Powers Report, February 2002 (Enron special committee); SEC litigation releases. Revenue / cash flow 2000 from Enron 10-K (April 2001, later partly restated).
"I am not claiming Nvidia is Enron. It is clearly Cisco."

— Michael Burry, May 2026[03]

What he means: no accounting fraud, no hidden losses. Instead, the growth-through-your-own-credit pattern that Cisco already played out 25 years earlier. Three steps, the same loop pattern as today.

Chart 04 · The Cisco loop
How Cisco financed its own sales in 1999
1999/2000 · BOOKING ≠ CASH
Vendor financing
$2.4B
on Cisco Capital's balance sheet as receivables
Actual cash returning
$0
after 2001 — all written off as
Booked as a receivable, never seen as actual cash.
Fiscal year 1999/2000
The Cisco loop
Three steps, the same pattern as today (shown schematically)
① Step 1
CISCO
lends telecom customers like WorldCom, Global Crossing, and Qwest a total of $2.4B through its in-house Cisco Capital.
② Step 2
CUSTOMERS
use the borrowed money to buy Cisco routers — at roughly the same scale as the loan.
③ Step 3
CISCO
books the sale as revenue — the sits on the balance sheet as a receivable.
Cisco gets the money back as a receivable — not as cash.
What happened next: the 2001 telecom crash. WorldCom, Global Crossing, and Qwest went bankrupt or wrestled with it.[21] Cisco had to write off around $900M in bad debt, the stock fell from around $80 to under $10 — minus 88%. Twenty-six years later, in May 2026, the from March 2000 still hasn't been recovered.
Sources: TechnoStatecraft Cisco/Nvidia comparison 2025[15]; American Affairs Journal "Who Lost Lucent" 2020[21].

Cisco then vs. Nvidia now — bigger and harder to see

The reflex on this comparison: "Cisco was much smaller than Nvidia today, so it's not really comparable." Absolutely true, but relatively misleading. In March 2000, Cisco was the most valuable company in the world — roughly 4% of US GDP, about 4% of the S&P 500. Nvidia today is absolutely larger, but also relatively more heavily weighted in the index: around 6% of the S&P 500, about 12% of US GDP.[15] If something breaks this time, more market cap is affected at once.

Table 01 · Then and now compared
What's visible, what's hidden
Metric Cisco 1999/2000 Nvidia 2026 (reported) Nvidia 2026 (ex. loop bookings)
P/E ratio20130~45–53[△]
Market cap (absolute)$555B$3.4T$3.4T
Share of S&P 500~4%~6%
Share of US GDP~4%~12%
Loop volume$2.4B vendor loans (open on balance sheet)$40B+ equity stakes (in Other Income)hidden in the loop
Q1 book profit from loop$15.9B (ASC 321)stripped from P&L

The Cisco risk in 1999 was obvious: 201, anyone could see that's too expensive. Nvidia's reported P/E of 30 looks cheap by comparison in 2026. Strip the ASC 321 book profits out of the P&L and the real ratio sits closer to 45. And that's conservative: on the buyer side, OpenAI, Anthropic, and CoreWeave hold cash that comes in large part from the hyperscalers' own equity stakes and funding rounds. Strip out that pre-financed buying power too (Goldman Sachs estimates under 15% of Nvidia's revenue as circular through 2027[13]) and the price-to-earnings ratio lands closer to up to 53[△]. This time the risk isn't just bigger, it's also harder to see.

"Money that isn't really there finances buyers who couldn't really pay, from a seller who books a profit on it."
Sources: TechnoStatecraft Cisco/Nvidia comparison 2025[15]; own calculation Q1 FY2027 Nvidia 10-Q ($15.9B unrealized equity gain stripped out)[02]; S&P 500 / US GDP share estimates as of May 2026.

Act 05 — ClosingTwo possible endings

Back to the working hypothesis from Act 01

"Burry was right."

Outcome A · Bullish

"Everything works out."

AI demand is real. OpenAI closes its $122B funding round, hyperscaler investments pay off, Oracle's RPO turns into cash, Microsoft and Amazon stakes keep growing with the valuations. The circular book profits then become prepayment on real growth. The loop legally resolves itself across the contract terms. Nobody loses.

Outcome B · Bearish

"Burry was right."

Just one link in the chain wobbles. OpenAI misses its funding target, Anthropic's valuation is cut in half, Oracle's RPO gets partly revised. The $67B in book profits has to be written down, stakes get repriced, backlog shrinks. A balance-sheet cascade like Cisco 2001. Stocks fall, token prices rise, tools go down. has been wrong more often than right since 2008[12]. But the pattern he describes has happened before: Cisco 1999.

Outcome C · the middle ground (most likely). A mix of both. The really big crash doesn't happen, because the hyperscalers generate operating cash flow and can throttle CapEx if needed. But two or three players in the chain get hit: one AI lab has to raise funding on worse terms, one RPO backlog gets reduced, one hyperscaler stock loses 30 to 40%. Token prices stabilize or rise slightly, a handful of tools disappear. Nobody calls it a "crash," everybody calls it a "repricing."

The Saturday morning with the errors and the headline is hypothetical. But not impossible. Cisco investors saw it on March 27, 2001. Back then it only cost them their portfolio. In 2026 it would also take away the tools they'd otherwise use to rethink their strategy — Claude, ChatGPT, the AI agents that sit in the workflow. The double risk: portfolio drawdown plus the simultaneous failure of the tool you wanted to react with.*

* This site is operated by Max Fraunhofer as a private journalist and is not licensed by BaFin under §32 KWG. The article is journalistic analysis, not investment advice within the meaning of § 1 Abs. 1a KWG or § 2 Abs. 8 WpHG. All figures come from public SEC filings, earnings call transcripts, analyst reports, and media coverage (see source list). Past performance is not an indicator of future results. Anyone making investment decisions based on this article does so at their own risk and should consult a licensed financial advisor or tax advisor.

Making-of — how this article was built
Making-of · last revised June 1, 2026

How this article was built

Six phases over two days. Research pipeline with parallel agents, fact-checking against primary sources, full rebuild after reader feedback. Methodology documented openly.

Pipeline
PHASE 01
Plan-mode + topic brief
Hypothesis clarified: money moves in a loop between the hyperscalers. Source policy: SEC filings < 6 months old plus Burry's original sources.
PHASE 02
4 parallel research agents
Tile data from hyperscaler 10-Q filings · Burry Substack tracking · Cisco 1999 retrospective · AI circular-deals pattern. Four Sonnet runs in parallel.
PHASE 03
Fact-check research
Numbers from 10-Q filings validated against SEC EDGAR. Burry quotes cross-checked against Substack + Stocktwits. Four critical findings corrected (including quote attribution).
PHASE 04
Synthesis
Three acts: mechanics · books · lesson. Book profit vs. cash as the thread. Cisco comparison as the historical frame. Enron block as the negative boundary before the Burry pivot.
PHASE 05
HTML build (v1)
First version on May 27 with a network diagram in the classic loop layout. Loop · books · Cisco loop.
PHASE 06
Reader feedback + rebuild
After feedback ("the chart doesn't get the point across"), full story rebuild on May 28. Network diagram out, charts 01 / 03 / 04 as three-part cards. New hook and ending — from the Burry-quote opener to the Saturday-morning scene.
Methodology for plausibility estimates [△]: The "real P/E ~45 to 53" in Act 04 emerges in two steps. First: from Nvidia's reported P/E (~30), the ASC 321 book profit from Q1 FY2027 ($15.9B unrealized equity gain) is stripped out of the P&L — the range lands around 45. Second: Goldman Sachs' estimate of "under 15% of Nvidia's revenue circular through 2027" applied as the upper bound — yields around 53 as the upper estimate. Both are range estimates, not point forecasts. The $67B aggregate in Act 03 is the sum of the four tile values from the respective 10-Q filings ([06], [07], [08], [02]) — not an own estimate.
Reviewer pass: four parallel reviewers (fact-check, UI/SEO, legal, copy editor) plus voice pass. All P0 and P1 findings integrated before publish.

SourcesFilings, citations, track record

Convention in this article: hard numbers, quotes, and external model outputs are marked in the body with a superscript source number. Click jumps straight to the entry. Three confidence levels are visually distinguished:

[N] hard primary source (SEC filing, earnings call, official press release) [N°] analyst estimate or secondary report [] own plausibility estimate — methodology in the Making-of
  1. 01CNBC · OpenAI closes $122B round at $852B valuation (Mar 31, 2026) · Nvidia with $30B equity stake.
  2. 02CNBC · Nvidia tops $40B in AI equity investments (May 9, 2026) · Aggregate of all Nvidia stakes in 2026.
  3. 03Stocktwits · "Michael Burry vs. Nvidia heats up: Big Short investor stands by his analysis" (May 2026) · Burry's reaction to Nvidia's 7-page sell-side memo, incl. "I am not claiming Nvidia is Enron. It is clearly Cisco."
  4. 05Michael Burry · "The Heretic's Guide to AI's Stars Part III: Tracepalooza & the Bezzle" (May 22, 2026) · Cisco reference ("There is a Cisco at the center of it all"), bullwhip argument.
  5. 06Microsoft 10-Q Q3 FY2026 (SEC EDGAR, April 2026) · RPO $627B, +99% YoY. OpenAI share derived from earnings call comments.
  6. 07Alphabet 10-Q Q1 2026 (SEC EDGAR, April 2026) · Google Cloud backlog $462.3B.
  7. 08Amazon 10-Q Q1 2026 (SEC EDGAR, April 2026) · AWS performance obligations $364B, weighted remaining term 5.5 years.
  8. 09Oracle Q3 FY2026 IR press release (Mar 10, 2026) · RPO $553B, +325% YoY. OpenAI $300B contract.
  9. 10CoreWeave 8-K SEC (Sep 9, 2025) · Backstop agreement with Nvidia, $6.3B through 2032. Primary source for CoreWeave mechanics.
  10. 11CoreWeave 10-Q Q1 2026 (SEC EDGAR, May 2026) · Backlog $99.4B, OpenAI share $22.4B.
  11. 12Elite Currensea · Burry track-record analysis (2024) · 2023 sell tweet, SPY/QQQ puts with 40 to 50% loss.
  12. 13Benzinga · Goldman Sachs flags "Circular Revenue" risk (October 2025) · Estimate: under 15% of Nvidia's revenue through 2027.
  13. 15TechnoStatecraft · Cisco vs. Nvidia — dotcom comparison (2025) · Cisco P/E 201, $555B market cap, 88% fall, ATH March 2000.
  14. 21American Affairs Journal · "Who Lost Lucent?" (August 2020) · Vendor-financing volume 1999–2001, SEC restatement detail.
  15. 22Howard Marks / Oaktree Capital · "Is It a Bubble?" memo (December 2025) · Top-4 hyperscalers $451B operating cash flow 2024, "Some AI revenue is currently circular in nature."