How DeepStack settles a question · Part 03

Is AI debt 1999, 2008 or neither?

The debt behind the machines: 1999 capex, 2008 credit, or neither?

From Part 03: The Machines Will Work. The Debt Might Not.. 6 voices on 2 sides, attributed and dated; the DeepStack Call with its history; 5 dated tests that settle it; the results as they come in.

The mechanism

The machines will earn; the question is who carries the paper behind them. Nebius's own terms put compute in backwardation: buyers pay roughly twice as much for six months or less as for one to three years, $40 to 50 million against $20 to 25 million, a signal about duration that a lender cannot ignore. The 95 percent H100 re-let and a modeled 21 percent return keep the bull case credible, yet neither settles who bears depreciation when financing costs rise, and only 6 percent of the AI paper issued is investment grade. So the credit stack, not the demand, is where the buildout breaks first: if the short-term premium persists and the queued $50 to 60 billion of issuance needs wider concessions to clear, lenders are repricing the risk before the machines stop earning.

Approved by Marco Cauduro on October 10, 2026.

The DeepStack callSkeptical

The machines will earn. The credit stack is where this buildout breaks first.

Bearish on the credit · Conviction High (4/5)

Our Call takes the side “1999 capex plus 2008 credit”.

Read the full call

Call history2 versions · last changed Oct 9, 2026

v2Oct 9, 2026Bearish on the creditConviction High (4/5)Side: 1999 capex plus 2008 creditSide named by the owner.
v1Aug 15, 2026Bearish on the creditConviction High (4/5)First call, as published with the story. The conviction, the actions and the flips were set down for the site on October 7, 2026.

3 voices

1999 capex plus 2008 credit

Michael BurryAug 12, 2026
Nebius extended its server depreciation schedule from four years to five this year even though its own deal pricing implies roughly 50 percent annual decay; customers that would not quit depreciating are what extended the bust twenty-five years ago.
Source
Michael BurryAug 12, 2026
Nebius's own disclosed terms show compute in backwardation: 20 to 25 million dollars per megawatt on one-to-three-year contracts against 40 to 50 million per megawatt for capacity of up to six months; buyers paying twice the rate for near-term use…
Source
Vitaliy KatsenelsonAug 6, 2026
AI profitability is elusive while open-source and Chinese models deliver comparable results at a fraction of the cost, so data-center pricing faces a race to the bottom exactly as fiber optics did after 1999; then the bankruptcies reveal who was…
Source

4 voices

The borrowing is rational

Market SentimentAug 8, 2026
Capacity is sold before it is built: AWS backlog reached 496 billion dollars, Microsoft closed June 2026 with 678 billion of commercial RPO up 84 percent year over year, Google Cloud's backlog hit 514 billion; CoreWeave's 2020-era A100s are fully…
Source
Market SentimentAug 8, 2026
Modeled bottom-up, 100 billion dollars of annual data-center capex sustains roughly 160 billion of steady-state annual revenue, a 1.6x revenue-to-capex multiple and about a 21 percent IRR, even assuming a five-year chip life and a one-third annual…
Source
Oguz ErkanAug 6, 2026
Rental prices for six-year-old A100 GPUs are up more than 20 percent year-to-date, a direct read on AI model providers' willingness to pay and, by extension, end-customer demand.
Source
Variant Perception
Still risk-on, semis capital-scarce, July was a tradeable low.

3 voices

Voices that cross sides

JunkBondInvestor
Three endings fit today's numbers: squeeze, triumph, or spreads just cheapen. Unsortable yet.
Oguz Erkan
His own tripwire: hyperscaler free cash flow may go negative in 2027; leverage on speculative returns repels.
Source
Market Sentiment
His own bar: the spend needs about $960B of cloud revenue; the run rate is about $380B, growing 40% or more.
Source

Where they cross

Who answers whom.

The specific points where one voice meets another: the same evidence read two ways, or the same mechanism with a different sign.

Market Sentiment contradicts Michael Burry, on the evidence
Market Sentiment contradicts Vitaliy Katsenelson, on premise
Michael Burry qualifies Oguz Erkan, on the evidence

Alternative explanations

Three other readings.

What would have to be true for the Call to be wrong, as the other side argues it.

The hardware is holding its price: rents on six-year-old A100s are up more than 20 percent this year, a direct read on willingness to pay that the backwardation signal does not capture (the case Oguz Erkan makes).
Capacity is sold before it is built, so the borrowing is rational: $496 billion of backlog at AWS, $678 billion at Microsoft and $514 billion at Google Cloud (the case Market Sentiment makes).
Open and Chinese models deliver comparable results at a fraction of the cost, so data-center pricing faces the race to the bottom that fiber ran after 1999 (the case Vitaliy Katsenelson makes).

What settles it

The dated tests.

The tests the story set, each with its specification where the owner has written one and the piece of the argument it verifies; on the Docket, resolved on the Results page.

NOV102026
Short-term premiums collapsing toward long-term rates would weaken the backwardation warning; persistence would make it harder to dismiss.
Part 03: The Machines Will Work. The Debt Might Not. Pendingp03-t01Pre-specifiedChecks the backwardation claim
Question
Did Nebius's September-quarter results, in the release, the shareholder letter or on the call, still price compute capacity of six months or less at a premium to its one-to-three-year contracts?
Source
Nebius Group, quarterly results furnished to SEC EDGAR (Form 6-K)
Rule
Read the price per megawatt (or per GPU-hour) Nebius gives for terms of six months or less and for one-to-three-year terms. The answer is yes if the short-term price is at or above 1.5 times the long-term price (the premium persists), and no if the ratio is below 1.5 (the premium collapsed toward long-term rates).
If the source is late, revised or silent
If the results carry no term pricing, or Nebius reports after November 30, 2026, the test is inconclusive, and the Result says what was read and what was missing.
The Call, before the dateExpects yes. Gives the yes answer 60%.
0 · no100 · yes
Next H100 batches
Re-lets near 95 percent support durable value; faster decay would expose a gap in the economics.
Part 03: The Machines Will Work. The Debt Might Not. Pendingp03-t02Open questionChecks the durable-value claim
Specification pending
JAN12027
Resilient hyperscaler free cash flow would ease leverage fears; negative cash flow would validate Erkan’s tripwire.
Part 03: The Machines Will Work. The Debt Might Not. Pendingp03-t03Open questionChecks the leverage claim
Specification pending
SEP82026
The next $50B-$60B issuance will show whether new-issue concessions keep widening or start to normalize.
Part 03: The Machines Will Work. The Debt Might Not. Call heldAnswer: nop03-t04Checks the concession claim
Question
Did the investment-grade corporate bond spread (ICE BofA US Corporate Index option-adjusted spread, FRED series BAMLC0A0CM) read tighter on October 2, 2026, the end of the four weeks after Labor Day, than its August 2026 average?
Source
FRED, ICE BofA US Corporate Index option-adjusted spread (BAMLC0A0CM)
Rule
The reading on October 2, 2026 (the last observation on or before that day) against the mean of the daily observations from August 1 to August 31, 2026: confirmed if it is 3 basis points or more below the August mean, refuted if it is 3 basis points or more above it. The split between AI-related issuers and the rest of the market, as the press reported it on September 22, 2026, is read into the Result as evidence and does not decide the test.
If the source is late, revised or silent
A series that is late, revised after the reading or silent for the window is inconclusive, reviewed by the owner, and the Result says what was read, what was missing and the next step. A difference inside the 3 basis point band is inconclusive and closes the test.
When the streak ends
The 182-day run ending would test whether unusual market calm is breaking; repetition would extend the signal.
Part 03: The Machines Will Work. The Debt Might Not. Pendingp03-t05Open questionChecks the market-calm claim
Specification pending

The test came in

Call held: the result is the one the Call expected.

Resolved October 10, 2026.

Every result is listed on the Results page.

Call heldAnswer: no

Oct 10, 2026 · Part 03: The Machines Will Work. The Debt Might Not. · Test p03-t04

Confirmed: the investment-grade spread ended the window five basis points wider than August's average, as the Call expected.

The test. The next $50B-$60B issuance will show whether new-issue concessions keep widening or start to normalize.

The ICE BofA US Corporate Index option-adjusted spread read 0.85% on October 2, 2026, the end of the four weeks after Labor Day, against an average of 0.80% over August's 21 daily readings (FRED series BAMLC0A0CM). The difference is +0.05%, five basis points, outside the 3 basis point band the rule names, so the answer to the question is no: the market did not price the window tighter than August. That is the outcome the Call expected of the credit stack.

What the series does not show, the press did. On September 22 Reuters reported spreads on AI-related issuers near 115 basis points against 78 for the broader investment-grade market, Alphabet paying a large concession on its August sale, and Aon's $13.5 billion financing drawing $65 billion of orders with its 30-year tranche tightening 35 basis points. Concessions normalized away from the AI complex and stayed wide inside it. That split is evidence here, not the rule.

The Call of Part 03, bearish on the credit of the buildout, expected the spread to widen. It did, by five basis points.

ICE BofA US Corporate Index option-adjusted spread, average 2026-08-01 to 2026-08-31: 0.80%; FRED BAMLC0A0CM · Aug 31, 2026
ICE BofA US Corporate Index option-adjusted spread, 2026-10-02: 0.85%; FRED BAMLC0A0CM · Oct 2, 2026
ICE BofA US Corporate Index option-adjusted spread, reading less average: +0.05%; FRED BAMLC0A0CM · Oct 2, 2026
Spreads on AI-related issuers against the broader investment-grade market, as Reuters reported on September 22, 2026: about 115 basis points against 78; Reuters, Corporate bond buyers get picky with flood of AI debt · Sep 22, 2026

The DeepStack call. Conviction unchanged: a wider spread at the end of the window is what the Call expected of the credit stack.

The result in full: what we said, the rule, what happened, what changes

What changed

The record of Part 03.

The Part’s publication, every change to the Call, every result that came in and the next dated test, oldest first.

Aug 15, 2026PublishedPart 03 published; the Call: Bearish on the credit.
Oct 9, 2026Call updatedThe Call: Bearish on the credit. The machines will earn. The credit stack is where this buildout breaks first. Our Call takes the side “1999 capex plus 2008 credit”. by Marco Cauduro
Oct 10, 2026ResultCall held: the result is the one the Call expected. Call heldAnswer: no by Marco Cauduro
Nov 10, 2026Next testNext: Nov 10, 2026, Short-term premiums collapsing toward long-term rates would weaken the backwardation warning; persistence would make it harder to dismiss. Pre-specified

Sources (3)

Market data are as of the dates cited. Voices are paraphrased from public writing and attributed by name.

Analysis and opinion, not investment advice. Voices are paraphrased from public writing and attributed by name; DeepStack shows no score, ranking or accuracy for any person. Every question, one page each.